Quantitative MRI of the at 7 tesla

COLOFON ISBN: 978‐908‐8915‐18‐5 This thesis was written and typeset in Microsoft Word 2003 using the Palatino Linotype, Baskerville Old Face, and Gaudi Old Style fonts. The cover image was modeled using Google Sketchup and rendered using the Kerkythea renderer. This thesis was printed by Proefschriftmaken.nl || Uitgeverij BOXPress, using 90 gram/m² G‐Print paper. Copyright of this works belongs to the author with exception of chapters, which have been published before elsewhere. The copyright thereto is held by their respective publishers. The publication of this thesis was financially supported by Philips Healthcare Nederland.

© 2012

3

Quantitative MRI of the human brain at 7 tesla

Kwantitatieve magnetische resonantie beeldvorming van het menselijk brein op 7 tesla

(met een samenvatting in het Nederlands)

Proefschrift

ter verkrijging van de graad van doctor aan de Universiteit Utrecht op gezag van de rector magnificus, prof.dr. G.J. van der Zwaan, ingevolge het besluit van het college voor promoties in het openbaar te verdedigen op dinsdag 11 december 2012 des middags te 12.45 uur

door

Daniël Louis Polders

geboren op 28 november 1981 te Haarlem

Promotor: Prof. Dr. P.R. Luijten

Co‐promotor: Dr. H. Hoogduin

Table of Contents

List of abbreviations 1

1. Introduction 5

2. Uncertainty estimations for quantitative in‐vivo MRI T1 mapping 39

3. In vivo three‐dimensional whole‐brain pulsed steady‐state chemical exchange saturation transfer at 7 T 63

4. SNR and uncertainty in DTI at 1.5, 3.0 and 7.0 tesla 91

5. Magnetization exchange effects and quantitative T1 mapping in tumor patients at 7 T 113

6. Multimodal tract‐based analysis in ALS patients at 7 T 145

7. Conclusions 169

8. Conclusies 177

Curriculum Vitae 187

Dankwoord 191

List of abbreviations

3D Three‐dimensional ALS Amyotrophic lateral sclerosis AP Anterior‐posterior APT(R) Amide proton transfer (rate)

B0/1 The vector or magnitude of the main magnetic field / the magnetic field of an RF pulse BW Bandwidth CC Corpus callosum CEST Chemical exchange saturation transfer CNG Cingulum bundles CSF Cerebro‐spinal fluid CST Corticospinal tracts Dlong Longitudinal diffusivity DS Direct saturation (of the bulk water pool) DTI tensor imaging Dtrans Transverse diffusivity EPI Echo planar imaging f0 Resonance frequency of water FA Fractional FH Feet‐head FOV Field of view GRASE Gradient and IR Inversion Recovery LDA Lorentzian difference analysis MD Mean diffusivity MRI Magnetic resonance imaging MS MT(C) Magnetization transfer (contrast)

MTR(asym) Magnetization transfer ratio (asymmetry) NMR Nuclear magnetic resonance NOE Nuclear Overhauser effect

1

ppm Parts per million (fraction of the main magnetic field, 1 ppm = 298 Hz at 7 T) qT1 Quantitative T1 (mapping) RF Radio frequency RL Right‐left RMV Repeated measures variance ROI(s) Region(s) of interest SAR Specific adsorption rate SD Standard deviation SNR Signal to noise ratio SPIR Spectral inversion recovery (for suppression) T1 longitudinal time T2 Transverse relaxation time T2* Transverse relaxation time, including local effects TE Echo time TI Inversion time TR Repetition time WM

2

3

Chapter 1. Introduction

It was eerie. I saw myself in that machine. I never thought my work would come to this. Isidor Isaac Rabi, physicist, 1898 – 1988, discoverer of nuclear magnetic resonance.

5 Chapter 1

6 Introduction

The magnetic resonance phenomenon

In the early 30s of the twentieth century, it was hypothesized and confirmed that an ensemble of paramagnetic spins1, when placed in a sufficiently strong magnetic field, would align either along or against this magnetic field. Furthermore, it was shown that these spins could be perturbed by irradiating spins with electro magnetic pulses at specific radio frequencies (RF), due to a resonance effect which was later named “nuclear magnetic resonance” (NMR). After perturbation the spins would return some of the energy, again in the form of RF radiation (2). This could be detected, although the emitted RF power was many times smaller than the input energy. It was also discovered that the characteristics of this signal (such as amplitude, duration, frequency and phase) reflected the immediate surroundings of the spins.

A method was thus discovered for characterizing the spins’ environment without actually touching them, but by sending RF pulses and “listening” to their echoes. The technique was readily adopted by chemists, who used it to elucidate the bonds, structures, and conformations of molecules of ever increasing complexity. It was not until the seventies that the magnetic resonance phenomenon was used to probe the spatial distribution of spins, and thereby create images of this signal. Also in the early seventies, it was shown in human tissue that differences in the spins’ environment affected the characteristics of the signal they emitted (3). As a result, it became possible to visualize differences in tissue types (i.e. healthy and diseased tissue), leading to

1 Not all isotopes have a magnetic moment, or are sufficiently naturally abundant to generate a detectable signal in tissue. In NMR, the nuclei of 1H atoms (protons), 13C carbon atoms, 23Na, 17O, 19F, and 31P can be used relatively easily. In clinical NMR however, the proton is by far the most observed nucleus. This thesis focuses only on the use of protons for MR measurements. Also, while formally incorrect, the terms 1H nuclei, protons, and spins will be used interchangeably. Note that the terms are used mostly in plural, denoting ensembles of spins. At the ensemble scale, we observe large collections of spins and can pass over much of the quantum mechanics theory (1).

7 Chapter 1 the advent of magnetic resonance imaging (MRI) as a diagnostic tool in medicine.

Nowadays, the knowledge of physiological and molecular processes that underlie disease has increased tremendously and there is a need for specific and quantitative biomarkers that can help to unravel biochemical pathways of disease origin and progressions. To date, MRI is by far the most versatile technology to produce biological pathway information non‐invasively. As the sensitivity of the water proton signal is orders of magnitude larger than all other MR visible signals in tissue, this thesis focuses on methods to obtain biological information by the magnetic behavior of these water proton spins caused by different physiological processes.

The remainder of this chapter will introduce three of the ways in which the MRI signal can be sensitized to reflect certain aspects of the spins’ surrounding. These concern interactions on nanometer and micrometer scales that affect MRI signals and are expanded on in the later chapters of this thesis. The goal for many MRI methods or scan sequences is not to elucidate a specific type of spin interaction, but rather to obtain a certain contrast in the acquired image, to elucidate disease. As a result, conventional MRI images reflect a mixture of several types of interactions, optimized to visualize that what was of interest for a specific application. This thesis focuses on methods that quantify specific interactions between spins to improve the understanding of the physiological processes that influence these interactions in healthy tissue and pathology.

The aim of thiss thesi is to explore the possibilities of performing quantitative MR in human brain using 7 tesla MRI. Quantitative MR methods cover those methods where the resulting images are parameter maps rather than the ‘raw’ MR signal. These parameter maps correspond to a physical characteristic of the measured tissue, e.g. an diffusion constant or a temperature map. Quantitative MR allows researchers to perform longitudinal studies and compare the values found at different time points. With the application of quantitative methods to discern the influence of pathology on a specific parameter, there is also a direct need for knowledge of the errors made in these methods. The last part of this chapter focuses on how to quantify the uncertainty of the result obtained when

8 Introduction fitting acquired data with a model. In other words: how well do we know what we know?

To maximize the sensitivity of the measurements, this thesis explores the application of quantitative methods by using a relatively novel platform for human MR measurements: the 7.0 tesla MRI system. This field strength is a significant increase from the 1.5 T and 3.0 T machines that are now found throughout the world. In past years, there has been a tendency to perform MRI experiments at ever increasing field strengths (4). As the signal of the MR scanner is linearly dependent on the field strength, approximately two times the SNR is expected when going from 3.0 to 7.0 T. This increase in signal can then be used to reduce the necessary scan times or increase the image resolution. Applications that previously had prohibitively large SNR demands might now become possible. Recent developments in DTIe (se below) illustrate how SNR is often the limiting factor to obtain the data required for more advanced models.

There are many confounding factors that can impede the full gain in SNR at 7 T.

For instance, the major relaxation constant T1 is dependent on the resonance frequency and therefore on the field strength (see below). Therefore, simply applying scan sequences that perform well on other fields strengths will result in images that perform sub‐optimal when no care is taken to adjust the timing to correspond to the relaxation rates at 7 T. More fundamental concepts of MRI also change at 7 T. On the microscopic tissue level, the effects of tissue on the magnetic field become more apparent. Boundaries between tissue types with differing magnetic susceptibility cause vessels and blood‐cloths to show as holes many times their actual size in T2* weighted images when this effect is not taken into account.

Technically, 7 T also brings many challenges. While the RF field can be considered relative homogeneous at 1.5 T and sufficient for many applications in the brain at 3 T, this is no longer the case at 7 T. Increasing the resonance frequency reduces the wavelength to ±12 cm in human tissue, causing strong standing wave effects in the human body. This results in signal peaks in the middle and signal voids on the sides of the brain (5). Add to this the fact that the amount of heat generated by RF pulses also increases with the square of the

9 Chapter 1 resonance frequency and the issue to solve the problems in RF inhomogeneity while using pulses that apply less power becomes clear.

Of course, many of these problems are really opportunities in disguise. For example, recent studies have shown that the capability to visualize micro bleeds and small intracranial blood vessels using 7 T is vastly superior in SNR and resolution to lower field strengths, mainly by utilizing these effects (6). It remains to be seen however, if all MR methods really gain much from being applied at 7 T and how the issues described above influence the quantitative nature of these methods. This is part of the exploration described in this thesis.

For the ordering of this chapter: the different methods described in this thesis are presented in order of the scale of the interactions they sensitize for. The introduction starts with nanometer interaction scale sizes with longitudinal relaxation, then to magnetization transfer and chemical exchange saturation transfer, and ending with diffusion tensor imaging on micrometer scales. Note that these are merely three of the myriad of ways an MRI physicist can influence the acquired signal. This does not aim to be a complete overview of all the physical, chemical and biological effects that MRI is capable to detect.

T1 mapping

Longitudinal relaxation and its molecular mechanisms When an ensemble of spins is placed in a strong external magnetic field, the spins align parallel or anti‐parallel to the main magnetic field, B0. At equilibrium, there is a small tendency for this ensemble to align parallel to the field, causing a net magnetization along the direction of B0. This net magnetization is the manifestation of many spins rotating at an angle with respect to the direction of B0, in a motion called precession. The characteristic Larmor precession frequency depends directly on the magnitude of the magnetic field B0 and the gyromagnetic ratio of the nuclei γ. At 7 T this 6   B 267.5 10 7 298 106 Hz amounts to: 220 . When the spins are perturbed with an RF pulse at this resonance frequency, the net

10 Introduction magnetization is tipped away from the direction of B0 and exposes a vector in the transverse plane, which is detected by receiver coils.

Longitudinal relaxation denotes the relaxation of the spin system from this perturbed state back to the initial equilibrium. It is the fundamental measurement characteristic of any nuclear magnetic resonance experiment, as it determines many of the parameters that make up a magnetic resonance experiment (e.g.: repetition time, optimal flip angle). The central concept here is that energy is stored in an excited spin system, but only until it finds a way to release that energy and relax back to its ground state. The longitudinal relaxation time is the amount of time it takes for the spins to relax a factor of

1e  63% . From this perspective, T1 is an empirical quantity resulting from a specific set of experimental parameters, such as field strength, sequence, and sample. In contrast to for instance cell size, a specific T1 value has no direct biological interpretation as such. It is the result of many different interactions and there are probably many ways of combining these interactions to arrive at the same T1 value.

In essence, longitudinal relaxation consists of all the possible mechanisms that spin systems undergo to dissipate the energy delivered by the RF excitation pulse. In the following, the most important molecular processes of longitudinal relaxation will be discussed (7).

Relaxation in liquids: Protons are relaxed by other protons. The magnetic moment of one proton can cause precession of neighboring protons around it. This is known as dipole‐dipole relaxation and is an interaction between motional energy (Brownian motion) of the system and the magnetic (Zeeman) energy. These energies must be compatible for efficient energy transfer, in other words the frequencies must overlap. This compatibility can be expressed by the spectral density function J(ω) or the rotational correlation time τc, which describes the distribution of the motional frequencies present in the pool and the characteristic resonance frequency in a magnetic field ω0 (see Figure 1). For instance, in free water there is a relatively large variation of frequencies present – from very slow to very rapid rotations – and the spectral density function is flat. The amount of overlap of spins with the Larmor frequency is therefore

11 Chapter 1 relatively small and relaxation is inefficient. For spins in a more structured environment the motional freedom is restricted and the spectral density function is limited. As a result, more of the spins have a frequency close to the resonance frequency and relaxation becomes more efficient (T1 becomes shorter).

Figure 1: Spectral density functions for three types of protons: protons bound to the surface of macromolecules display a long correlation time τc and steep spectral density function J(ω). Protons in a structured environment (i.e. hydration layers) display a medium τc and gradual spectral density function J(ω). Protons in a relatively unrestricted environment display a short τc and flat spectral density function J(ω). Reproduced from: “From picture to proton”, D. W. McRobbie, E. A. Moore M. J. Graves and M. R. Prince, Cambridge University Press, Copyright (2006). Spin‐lattice relaxation in liquids is thus depended on the temperature and the motional freedom experienced by the spins. The field strength also influences this interaction, as a higher field strength means a higher resonance frequency and a shift along the spectral density function. An increase in T1 is indeed observed when going from low to high field MRI systems because of the reduction of this interaction.

Relaxation in solids: Brownian motion plays a minimal role in the relaxation in solids and so the amount of dipole‐dipole relaxation is negligible. Still, the

12 Introduction relaxation of magnetization in solids is important, because there are several ways for magnetization to transfer from bulk water to this pool of spins. One of these ways is by means of magnetization transfer, either as a Nuclear Overhauser Effect (NOE), dipole‐dipole interactions, or by means of chemical exchange of protons from labile moieties that are in constant exchange with the bulk pool. See also the next section on chemical exchange saturation transfer.

Regardless of the transfer mechanism, once the magnetization is located at the surface of the solid, it is highly mobile. By the process of ‘spin diffusion’, the magnetization energy is transferred from one spin to the next. In this process, the spins are stationary and no dipole‐dipole relaxation takes place. Relaxation is mediated by spin diffusion by transferring spin‐states rather then particles to nearby relaxation sites that act as magnetization “sinks”.

Much of the relaxation in biological systems takes place at relaxation sites. These are locations where magnetization energy is easily and quickly dissipated. A prime example of a relaxation site is a metal nucleus such as (Gd). Often used as a contrast‐agent in MRI, it allows for significant shortening of the T1 by supplying a relaxation site to many spins. Other metals such as (in ) and zinc (present in many ) similarly act as naturally occurring relaxation sites.

Biological basis of T1 contrast

From a biological perspective, the longitudinal relaxation time T1 is related to three main factors, water content, cellular environment, and the water binding properties of this environment. Another name for T1 relaxation is spin‐lattice relaxation, which describes the main processes involved in this relaxation phenomenon. The cellular environment experienced by the protons (the “lattice”) consists of proteins, lipids, and macro‐molecules such as cell‐ membranes and cyto‐structural complexes. These environmental factors occur in varying combinations, depending on the tissue type. For instance, the observed relaxation in case of cerebro‐spinal fluid (CSF) is relatively slow while much faster relaxation takes place in myelinated white matter. In CSF, the spectral density function is relatively flat, resulting in inefficient dipole‐dipole relaxation and there are few macromolecules to serve as a sink for

13 Chapter 1 magnetization transfer and relaxation in solids. In WM, the environment of the water is dominated by tightly wound sheets of lipid bi‐layers, causing much of the water to be in hydration layers that display much reduced rotational freedom and strongly increased dipole‐dipole relaxation of the spins. In addition, the lipid bi‐layers and proteins within them allow for efficient relaxation once the magnetization is transferred to this solid pool.

Measuring T1 The longitudinal relaxation time is measured by sensitizing the MRI signal for this relaxation behavior, while minimizing the effect of other relaxation effects.

One way to quantify T1 is by sampling the inversion curve. This curve reflects the recovery of MR signal after inversion of the spin system. This curve can then be fitted to an expression including the T1 parameter.

The golden standard for measuring T1 is the single slice inversion recovery (IR) sequence, although due to the long scan times other and faster methods are often used. In the IR method, an inversion pulse is played out and after a varying time period (inversion time, TI), a slice is excited and read out with a short echo time. The inversion curve is sampled by repeating the experiment with another inversion time. Because repeated measurements are necessary for each point on the curve and repetitions are to be acquired at roughly 5 × the estimated T1, this measurement is often prohibitively lengthy for clinical use.

The clinical importance of measuring T1 While the process of longitudinal relaxation was known since the invention of NMR, it was not until the early seventies that is was shown that different tissue types show different relaxation rates (3). The observation that the longitudinal relaxation time was a key characteristic of tissue types, was reproducible over several subjects and showed changes when healthy tissue became diseased, was one of the findings that spurred forward the development of NMR as a diagnostic imaging tool.

The MRI signal is influenced by many different interactions that spins have during an MRI experiment and much of the optimization of MRI sequences focus on maximizing the desired contrast, while minimizing unwanted effects.

14 Introduction

To optimize a method, an MRI physicist needs to know what the relaxation times of the tissues in question are.

When sufficient knowledge of the T1 values displayed by healthy tissue in several subjects is obtained, T1 relaxation time can be interpreted as a tissue parameter and a marker of pathology. By determining the actual T1 value, as opposed to the effect of T1 on the contrast in an image, T1 values can be directly compared between patients and healthy controls. Global changes in T1 can be spotted using histogram analysis and region based analysis can aid in quantitative comparison and heterogeneity assessment.

T1 mapping at 7 T

As stated above, T1 values depend directly on the resonance frequency and it is therefore expected that T1 values increase when measuring at higher field strengths. Some critical reviews have indicated that this increase in T1 could cause a de‐facto decrease in T1 contrast, because the T1 values of different tissue types might start to converge with increasing field strengths (4). Therefore it is of interest to determine the T1 time constants for different tissue types both in healthy and diseased tissue at 7 T. The effective change in T1 contrast when performing MRI at 7 T can be assessed by analyzing the distributions of T1 values found in different tissue types.

T1 mapping in this thesis

In this thesis, the T1 values in different tissue types are investigated in Chapter

2. This way, the effective contrast in T1 values at 7 T was determined. To do so, a fast T1 mapping method that acquires a reasonable volume in just over four minutes was implemented. This allowed the sequence to be included in patient studies, where available scan time is often limited. The T1 values in different tissue types in tumor patients and in ALS patients were investigated in Chapters 5 and 6, respectively.

15 Chapter 1

Magnetization transfer and chemical exchange saturation transfer

Magnetization transfer One important observation in MRI is that while the majority of the signal comes from protons in liquid form, i.e. bulk water or adipose tissue, there are many more protons present in more rigid forms. Many protons are located in lipid bi‐ layers, structural and larger globular proteins, etc. These protons are covalently bound to stationary structures and have very limited motional freedom, resulting in very short T2 relaxation times2. This makes them generally invisible in MRI experiments3. Also, not all bulk water can be considered truly liquid because much of it is located in hydration layers around these macromolecules. The bound water fraction is of specific interest because it too can be visualized, albeit indirectly. Water hydration layers form around the electrical charges present on surfaces and cause water molecules to realign with respect to the surface charge. This alignment is by no means a rigid fixation; much movement is still possible and these protons are in continuous exchange with the bulk water pool. Further away from the surface, this alignment effect is much reduced and water is more free to move according to regular thermal (Brownian) motion. In MRI measurements, the bound water fraction can be observed because it has a much reduced T2 compared to the free bulk water pool.

When looking at spins in the spectral domain, a short T2 is equivalent to a broad resonance peak. In bulk water, the motional freedom of the protons causes

2 T2 relaxation: also known as spin‐spin relaxation, describes the loss of MR signal due to loss of phase coherence of the signal. This results from spins sensing the magnetic effect of neighboring spins, and therefore experiencing a (slightly) different magnetic field. This results in a slight change in resonance frequency which is observed as a net loss of phase coherence of the ensemble. In solids T2 is generally very short; in the order of a few microseconds. 3 With the exception of MRI methods that realize ultra short echo times to allow visualization of covalently bound protons in for instance cartilage and .

16 Introduction motional narrowing of the resonance peak, the resulting width being mostly determined by variations in the B0 field. This is in the order of one hundred hertz or less in the human brain at 7 T. In the case of the bound water pool however, T2 effects causes the resonance spectrum to become as wide as tens of thousands of hertz. It has been shown as early as 1989 that applying RF pulses off‐resonance at a few thousand hertz from the water resonance has a suppressing effect on the acquired signal from the water pool (8). Thus, the magnetization applied off‐resonance is somehow transferred to the bulk water pool (which in itself is insensitive to off‐resonance excitation). Several mechanisms for this magnetization transfer effect have been suggested, such as the exchange of water from the bound water pool to the bulk water pool (a relatively slow process), spin‐diffusion from bound water protons to bulk water protons (much faster), and chemical exchange from exchanging moieties on molecules with labile (non‐permanently bound) protons.

Magnetization transfer by chemical exchange Probing the mechanism of magnetization transfer by chemical exchange has sparked much interest in the last twenty years, starting with its translation from NMR to MRI experiments in 1990 by Wolff and Balaban (9). It was hypothesized that specific exchanging moieties at endogenous (naturally occurring) molecules display a specific resonance frequency, just as in MR spectroscopy. In conventional NMR spectroscopy, these exchanging groups are rarely observed precisely because of the continuous exchange of these groups with the water bulk pool. In this application, chemical exchange saturation transfer offers a way to indirectly visualize exchange of chemically specific protons.

In a chemical exchange saturation transfer (CEST) experiment, the MRI signal is sensitized for chemical exchange using saturation pulses at specific radiation power levels and off‐resonance frequencies tuned to saturate specific moieties of labile protons at known molecules. This is different from the previously described magnetization transfer techniques, where the saturation pulse was optimized to maximize all magnetization transfer effects. For instance, proteins and peptides contain amide groups, which in turn have a labile proton (the

17 Chapter 1 amide proton). From in vitro experiments the resonance frequency and exchange rate have been determined to be +3.5 ppm (from the water resonance) and about 28 Hz at pH = 7 respectively (10).

From exchange theory it is known that the exchange rate of labile protons is governed by the environment of the molecule, especially the acidity (pH) and temperature. By sensitizing the MRI experiment to the exchange of a known type of labile protons, CEST experiments promises to offer a window into these fundamental physical properties and the relative amount of these labile protons. In the case of large proteins in their natural confirmation (i.e. folding of the amino acid chain) not all of the amide protons are at the outside of the protein structure, so not all of them are expected to contribute to the exchange process. If by disease however a protein becomes partly broken down or incorrectly folded, more amide protons are expected to become exposed and thus ‘visible’ in a CEST experiment.

Measuring CEST Designing an MRI‐CEST measurement begins with the realization that the classic saturation based magnetization transfer method has a strong frequency dependency. At any given off‐resonance saturation frequency the signal is suppressed by at least three different mechanisms. These mechanisms (i.e. direct water saturation, magnetization transfer, and chemically exchanging protons)l will al contribute according to some underlying frequency dependency. By repeating the saturation experiment at many different saturation frequencies, a so‐called Z‐spectrum is obtained. This is the summed effect of all magnetization type effects present in the sample, weighed by their respective sensitivity for the saturation pulse used. eSee Figur 2 for a schematic example in the case of a single type of exchanging proton (the solute); amide protons exchanging at +3.5 ppm. There are many ways to design a saturation experiment and many of the current methods are reviewed in the review by Van Zijl and Yadav (11).

18 Introduction

Figure 2: Schematic overview of chemical exchange saturation transfer. a) Solute with labile protons are in continuous exchange with the bulk water pool. b) When applying RF irradiation on the solute’s specific resonance frequency, saturation will reduce the bulk water signal. c) The frequency dependency of the RF saturation pulse is expressed in a Z‐spectrum. d) Asymmetry analysis is applied to discriminate the effect of chemical exchange from that of conventional MT and direct water saturation (the peak at the water frequency). Reproduced from: Magnetic Resonance in Medicine vol. 65, no. 4, p. 927–948, “Chemical Exchange Saturation Transfer (CEST): What is in a Name and What Isn’t?”, Peter C. M. van Zijl and Nirbhay N. Yadav, Copyright (2011), with permission from Wiley and sons. In many clinical settings it is not feasible to collect a densely sampled Z‐ spectrum over a broad frequency range due to scan time limitations and one might compromise with a selection of relevant off‐resonance frequencies of interest.

19 Chapter 1

CEST applications The factors described above make CEST focused at amide proton transfer imaging (APT) a measure that is expected to be sensitive to a wide range of physical and physiological processes. It was shown very early that urea could be visualized in the kidneys by observing the saturation effects on amide protons (12). Other applications of APT imaging have shown pH alteration in rats during early ischemia (10), increase of APT in tumor tissue (13–15), and discrimination between radiation necrosis and tumor recurrence (16).

As the interest in the chemical exchange mechanism grew, many new applications were found for CEST measurements that focused on other endogenous labile protons. Lesions in the patella in the human knee were visualized by focusing on the exchanging hydroxyl protons (‐OH) in glycosaminoglycans (17). Insights into energy metabolism were obtained by observing the exchange of hydroxyl protons in glycogen and glucose (18). Neurological inhibition was investigated by measuring glutamate (19).

Another category of CEST contrasts are the exogenous CEST agents. As they are designed to maximize chemical exchange while minimizing overlap with normal tissue resonances, these particles and aggregates can display CEST enhancements many orders of magnitude larger than the endogenous CEST agents. Their application is currently pre‐clinical and beyond the scope of this thesis. The reader is kindly referred to the review by Zhou and Van Zijl that covers the exciting new possibilities that these compounds bring (20).

CEST at 7 T There are advantages to performing CEST measurements at 7 T and they are analogous to those found in spectroscopy. The spectral dispersion of resonances in Hz increases with the field strength, while the bandwidth of RF pulses remains dependent mainly on the duration and pulse shapes used (irrespective of resonance frequency). This means that it becomes easier to selectively saturate a narrow bandwidth centered on the group of interest. Currently, not much is known of the clinical relevance of the different exchanging protons near the water resonance, but a higher specificity of the saturation process will simplify the interpretation of CEST effects.

20 Introduction

An additional advantage is the increase in T1 relaxation time. As CEST experiments depend on the build‐up of saturation, which is often not achieved in a single saturation pulse, it is of importance to consider that the saturation effect decays with T1. Therefore, longer T1 values allow for the build‐up of more saturation, effectively increasing the sensitivity of the measurement.

Of course the technical limitations of 7 T, especially with regard to the inhomogeneity of the RF field and SAR constraints, make the implementation of an efficient CEST measurement a challenge on its own. Initial results of CEST measurements at 7 T by others (21, 22) have observed larger differences between grey and white matter than earlier observed at lower field strengths and have shown that the method could be well‐suited for the characterization of white matter (WM) pathology.

CEST in this thesis CEST measurements hold great promise as the next step in magnetization transfer imaging that might allow for in‐vivo quantification of many clinically relevant parameters, including pH, temperature, and amide concentration. Therefore, it is a very interesting method to add to the MR toolbox. The aim of Chapter 3 was to develop a method that enables the acquisition of CEST data in a clinical setting. In addition to the limitations described above, this puts constraints on the minimal resolution and maximum scan time.

Previous work on lower field strengths has shown large increases in APT‐CEST effects in tumor patients, credited to elevated levels of free amide protons in and necrosis (15, 16, and 23). Therefore, this patient group forms a good test to validate this new method. In Chapter 5, the method developed in Chapter 3 is applied in a small group of tumor patients.

In addition to tumor characterization, the feasibility of this method to characterize WM pathology is explored. As very little is known of the etiology of amyotrophic lateral sclerosis (ALS), the CEST method was included in a range of MR methods to elucidate this disease in a small cohort of patients. The results of this study are described in Chapter 6.

21 Chapter 1

DTI

Diffusion MRI Brownian motion is the process whereby particles are agitated by the movement of other particles. It is the continuous collisions of particles that cause the seemingly random movement when observing a single particle. The energy needed for this phenomenon comes from heat, hence the term thermal motion. The amount of agition is directly dependent on the temperature of the system. The phenomenon of Brownian motion was discovered in 1828. This was followed by work by Fick who explained the mixture of fluids and gasses with an initial concentration gradient and called this process diffusion (24). It was not until 1905 that Einstein (25) and Smoluchowski (26) independently combined these two processes and showed that diffusion was governed by Brownian motion. They developed the expression for the mean distance traveled by a collection of particles in a given medium. For spherical particles in a low viscosity medium this can be expressed by the Stokes‐Einstein relation:

2 kT xDtD6 B [1] 6r

The mean distance x is directly dependent on the diffusion constant D and time t. The diffusion constant itself is a function of Boltzman’s constant, kB, the temperature, T, the viscosity of the medium, η and the radius of the particles, r. In MRI experiments, the particles and the medium are roughly the same: protons and the water molecules.

The diffusion distances that are probed during a typical MRI experiment are in the order of micrometers. Therefore, diffusion MRI acts as a very sensitive probe for the cellular micro‐structural environment of the protons.

The expression above is only true in the case of free diffusion, where there are no structural boundaries that interact with the particles. This is never fully the case in experiments in tissue. Rather the diffusion constant that is observed is the result of many interactions of water molecules with the cellular environment, i.e. structural proteins, cell walls, etc. As a result, the measured

22 Introduction diffusion coefficient is that of hindered diffusion and is expressed as the apparent diffusion coefficient or ADC.

Additionaly, the expression given above assumes that diffusion is the same in all directions, as is the case in free diffusion. In the case of hindered diffusion in tissue however, this is not always the case. In highly structured tissue such as white matter fibers or muscle fibers, the apparent diffusion coefficient along the fiber structure is markedly larger than perpendicular to the fiber direction (27). In free diffusion or diffusion in unstructured tissue, diffusion is considered isotropic, while in structured media with a certain preferential diffusion direction, diffusion is anisotropic. To account for this effect, the diffusion coefficient can be expressed not as a scalar but as a 3×3 tensor4. Several parameters can be calculated from this tensor, the most important two being: The fractional anisotropy (FA), which is the measure of anisotropy and the primary eigenvector of the tensor, i.e.: the direction in which diffusion is least hindered. These two parameters are used in fiber tracking and to create color coded FA maps. The direction of the primary eigenvector, λ1, is used to determine the color of each voxel in these maps, while the voxel intensity is given by the FA. This 2D representation of WM fiber architecture shows much of the underlying anatomy, even without 3D fiber trajectory reconstruction (Figure 3).

4 This is only a first order approximation, and known to be incorrect. The environment and hindrance of diffusion is not expected to be the same for all protons in the voxel and as such the measured signal is not expected to display the behavior of a single Gaussian distribution.

23 Chapter 1

Figure 3: From Tensor estimation to color coded FA maps. Based on diffusion measurements along multiple axes (A), the shape and the orientation of a ‘diffusion ellipsoid’ is estimated (B). Anisotropy is calculated based on the ratios of lambda 1, 2, and 3, and is visualized from 0 (no anisotropy / isotropic diffusion) to 1 (absolute anisotropy, lambda 2 and 3 are both zero). Also from the estimated ellipsoid (B), the orientation of the longest axis is determined (C) and this primary direction of diffusion is translated to a color (F). Combining (D) and (F), the color coded FA maps is generated (E). Reproduced from: , vol. 51, no. 5, S. Mori and J. Zhang, “Principles of Diffusion Tensor Imaging and Its Applications to Basic Neuroscience Research,” pp. 527–539, Copyright (2006), with permission from Elsevier. Water diffuses easiest along white matter bundles, rather then perpendicular to them. This effect has been exploited by using the primary eigenvector and FA values resulting from the diffusion tensor to reconstruct the fiber bundles in human brain (28). In essence, the approach to reconstruct the fiber pathways is one of brute force. Many possible trajectories are reconstructed from a starting region of interest or even from all voxels in the brain volume. Each voxel is ‘seeded’ with many starting points, evenly spaced over the voxel. Typically 3×3×3 or 5×5×5 seed points are used. Figure 4 illustrates the general approach taken in streamline and the need for multiple seed points per voxel.

24 Introduction

Figure 4: Fiber tracking procedure illustrated in 2 dimensions. Starting at three locations within the same voxel, different fiber trajectories are reconstructed. One of these trajectories is eventually terminated by failing one of the tracking criteria, in this case the maximum angle from one segment to the next.

Measuring DTI To sensitize the MRI scan for diffusion, relatively large gradient fields are applied after excitation (29). This effectively labels the protons with respect to their location in the gradient’s field. After some diffusion time (tens of milliseconds later) the same diffusion gradient strength is used to remove this labeling and realign the phases of the protons. This however only works if the protons were completely stationary during and between these gradient pulses. As diffusion cause the protons to migrate in the field, signal is not fully re‐ phased and the acquired signal is reduced. The amount of signal reduction is directly proportional to the effective diffusion constant of the protons under consideration:

S bD b  e [2] S0

Where Sb/S0 is the relative signal in the presence of an effective gradient field b. In the case of trapezoidal gradient pulses, the amount of diffusion weighting, expressed in the b‐value, is given by the Stejskal‐Tanner relation :

25 Chapter 1

bG222   3 [3]

Where γ is the gyromagnetic ratio, G is the amplitude of the gradient pulse, δ is the duration of a single gradient pulse, and Δ is the duration between two gradient pulses.

Figure 5: Schematic overview of the dephasing effect due to diffusion in the presence of motion probing gradients. Reproduced from: Neuron, vol. 51, no. 5, S. Mori and J. Zhang, “Principles of Diffusion Tensor Imaging and Its Applications to Basic Neuroscience Research,” pp. 527–539, Copyright (2006), with permission from Elsevier. As diffusion is rarely truly isotropic, the amount of dephasing will depend on the direction of the gradient field. To fit the acquired data to the tensor model, This effect is incorporated by assuming the gradient to be a vector in a given coordinate frame and then repeating the diffusion measurement with a collection of different gradient directions. It was shown that a collection of maximally non‐colinear gradient directions, as is obtained by distributing ±30 gradient vectors homogeneously on a unit sphere, is optimal to sample the directional dependency of the diffusion tensor (30). Because of the noisy nature of DTI data and the log transform as a result of eq [2], data quality monitoring is important and outlier rejection can improve the tensor fitting procedure (31).

26 Introduction

Applications of DTI It is important to remember that DTI is a neuro‐anatomical tool first and foremost. It is the best non‐invasive technique to obtain structural information of the brains’ pathways. White matter seems a homogenous substance when observed using a T1 weighted sequence, but post‐mortem studies clearly show that it is heavily structured. Using DTI the orientation of the larger fiber bundles present in the white matter can be elucidated in vivo. Using tractography methods, different cortical areas can be connected, allowing for better understanding of the structural networks that are present in the human brain (32). One obvious application of this anatomical knowledge, apart from basic neuroscience interests, is that of pre‐surgical planning. Full invasive brain surgery is one of the three major options in the treatment of intra‐cranial brain tumors, next to radiation therapy and chemo‐therapy. Knowledge of eloquent cortical areas and the fiber bundles connecting them, holds obvious advantages to patient outcome (33, 34).

In addition to anatomical information, several diffusion parameters convey micro‐structural information and therefore reflect pathology. Of all diffusion parameters, the fractional anisotropy is most often reported to be a sensitive marker, reflecting many different processes, including neuronal degradation, , ischemia, and (35). For instance, it has been shown in multiple sclerosis that FA drops in the case of new/active ring‐enhancing lesions. This indicates that inflammation is somehow correlated to microstructural changes such as cell swelling or changes in the extracellular matrix. Another exciting application of DTI is combining the structural and functional information to investigate pathology where connectivity is expected to be altered, as may be the case in and amyothrophic lateral sclerosis (ALS) (36, 37).

DTI at 7 T Increased SNR is an important reason to perform DTI measurements at higher field strength for the following two reasons:

First, as standard streamline tractography is based on the single tensor model, it is important to realize that there are severe limitations to this model. Most

27 Chapter 1 importantly, this model does not allow for multiple fiber populations with differing principal directions. That is, each voxel is assumed to contain a single direction of neuronal fiber. There are however several regions in the brain where neuronal connections cross, i.e. corpus callosum fibers connecting the left and right hemisphere and cortico‐spinal tracts connecting the cortex to the brainstem. At these crossings, the average FA drops and an average principle eigenvector is measured, which leads to errors in fiber tracking results. One solution for this problem is to simply increase the resolution of the DTI acquisition. This reduces the effect of larger crossing fibers tremendously. However, increasing the resolution is exceptionally SNR demanding. With current methods, it is expected that full brain DTI will not achieve resolutions higher than 1×1×1 mm³.

Second, even at sub‐millimeter resolutions, it is expected that crossing fibers can significantly bias fiber tracking results. Using more sophisticated models that allow for multiple fiber directions is the only way forward and has shown to increase the number, precision and accuracy of reconstructed fibers significantly (38, 39). The acquisition of diffusion data for such higher‐order models often means the acquisition including many more gradient directions and at higher b‐values. This also requires greatly improved SNR.

In this respect, DTI at 7 tesla holds great promise. However, there are a number of drawbacks that are expected to influence the quality of most DTI measurements at high fields. For instance, the basis of the diffusion measurement is a combination of a spin‐echo with an echo planar imaging (EPI) read‐out. This is a combination of refocusing the MR signal using RF pulses (the spin echo) and gradients (EPI). The spin echo is sensitive for B1 inhomogeneity, especially in the case of non‐adiabatic RF pulses as is often the case in imaging methods. The gradient echo technique is sensitive for B0 field inhomogeneities, such as occur at tissue‐air boundaries (i.e. the nasal cavities and ear canals). These two factors directly influence the image quality that can be obtained and are expected to be more problematic at 7 T than at lower field strengths.

28 Introduction

DTI in this thesis Given the promises and anticipated drawbacks described above, it remains an open question whether or not DTI at 7 T is beneficial at all. This issue is addressed in Chapter 4, where DTI performance is compared at three field strengths and the effect of SNR on the resulting DTI parameters is analyzed in detail.

In addition, DTI data obtained at 3 T was used as an anatomical tool to find the cortico‐spinal tracts (CST) of patients and controls in a study that aims to elucidate the pathology in this specific tract. Chapter 6 describes the method to use fiber‐tracking dto fin the voxels that are part of the CST subject‐by‐subject. Also, changes between ALS patients and healthy controls along the CST on a group level are reported, while allowing for variation of each subject’s personal fiber trajectory.

Measuring the uncertainty of fitted values

Uncertainty and bias The concept of measurements and their accompanying errors is well grounded in all the quantitative sciences. Whether scientists discriminate between ‘measurement error’ and ‘measurement uncertainty’, ‘type A’ errors and ‘type B’ errors, ‘accuracy’ and ‘precision’, ‘random errors’ and ‘bias’, the gist remains the same: There are roughly two reasons why one might measure something different than the ‘true’ value.

The first reason has to do with chance, the amount of noise, short term fluctuations in the measured data. This is often called the uncertainty, precision or measurement imprecision. The second type of error includes deviations due to a systematic bias, which are present over a long period in time and are essentially fixed. This type of error is often associated with terms like bias, accuracy, measurement error, etc. Discrimination between these two causes is useful in many cases, for instance in quality assurance of MRI methods where it often proves more relevant to be able to spot a machine drift causing an increasing bias in a measurement, while the noise level remains the same.

29 Chapter 1

Conversely, when comparing a new method with an older one, an essential requirement is that the accuracy remains the same, while improvements are likely to be made in the level of uncertainty.

For MRI, these two types of errors have different sources, with bias arising from hardware properties, operator training, and even the version of the software used in the measurement and subsequent data‐processing. Uncertainty errors arise from variation in physiology, positioning of subjects in the scanner/coil, patient movement (including cardiac pulsation and breathing). To be able to assess the uncertainty and bias in a measurement is an important tool when assessing variation in the data. The measured variation (for instance in a collection of equivalent voxels) is expected to be of the same order as the estimated uncertainty. If however the variation in the measured values is much larger than the uncertainty of the measurement, this hints at some sort of heterogeneity that might arise from some additional factor, (e.g. the voxels are not equivalent, or the SNR is not homogeneous over the selection).

In quantitative MR methods, the measured MR signals are fitted to some kind of model, before the parameter of interest is determined. Especially in in‐vivo data the physiological noise plays an important role, because the amount and nature of this noise is highly variable. In some cases, the noise has a non‐linear effect on the estimated parameter, making it impossible to guess the effect of a change in raw SNR to a change in the variability of the estimated parameter. Therefore, it is of great importance for quantitative MR methods to have a good idea of both the uncertainty and bias present in the method.

Assessing uncertainty To measure the variation present in the estimated parameter from measurement to the next, one can repeat the measurement until a reliable estimate of the variation can be made. However, in most clinical studies this is not feasible for each subject due to time constraints. Therefore, many methods to estimate the uncertainty without repeated measurements have been developed.

One way to assess the measurement uncertainty is by analytical approximations of all errors in the method and error propagation to arrive at the final

30 Introduction uncertainty level that can be expected. For these approaches, an estimation of the SNR is needed. However, in many MRI experiments, especially at high field strengths, SNR varies strongly over the imaging volume and also between scans. This means that the SNR cannot simply be estimated globally and must be measured locally to be reliable. This is not trivial, as there are many confounding factors in high‐field MRI that influence the local SNR, such as receiver coil sensitivities and inhomogeneities in the RF transmit field.

When data is fitted to a model before interpretation, another approach is possible. It is possible to assess the variance in the fit, by feeding gold standard data with known amounts (and distributions) of noise into the algorithm. This way, one can create a look‐up table of the variability of the fit for a given noise level. By measuring the noise level during an experiment, this can then be related to an expression for the expected variance of the fit (i.e. standard deviations, confidence intervals).

Another approach is to use only some of the available data to perform the fit to our model. Many permutations of possible combinations of a subset of the data are fitted and the effect on the estimated fit parameters is observed. This approach makes few assumptions on the level or distribution of the noise in the acquired data. This way, one can observe the robustness of the fit, given the (unknown) noise present in the data, without outside information. Pulling oneself up by ones bootstraps, so to say. Hence the term for this type of uncertainty estimation: bootstrapping (40).

A variation on this theme is not to use a selection of the available data, but rather to use the residuals from a single measurement. Many synthetic datasets are created based on the fit and random resamplings of these residuals. This approach is called wild bootstrapping and can be applied to assess the variance of a fit without the need for repeated measurements (41).

Uncertainty estimation in this thesis Chapter 2 shows for the first time that it is possible to apply the wild bootstrap method to quantitative T1 mapping. It was shown that wild bootstrapping allows researchers to apply a quantitative method with knowledge of the

31 Chapter 1 uncertainty, without the need for repeated measurements or additional SNR determination.

The wild bootstrap approach will be applied in Chapter 4 in a comparison of diffusion tensor imaging at different field strengths. Because the bootstrap approach estimates the uncertainty in all fitted parameters, the effect of SNR on the FA value and the first eigenvector was estimated. As these two parameters are most often used in DTI experiments, for parametric comparison as well as fiber tracking, the amount of uncertainty directly reflects the quality of the DTI measurement.

32 Introduction

Outline of this Thesis

To recapitulate, a short overview of the remaining chapters in this thesis is presented.

In Chapter 2, a rapid slice‐based T1 mapping procedure is implemented. By applying the statistical wild‐bootstrap method to this method, the uncertainty of the fitted T1 values is assessed, without the need for repeated measurements. The performance of the bootstrap method for uncertainty estimation is validated by comparison with simulations and repeated in‐vivo measurements.

In Chapter 3, a new sequence is introduced to rapidly acquire full brain CEST data, in such a way that CEST acquisitions become clinically feasible. This means achieving sufficient saturation effect, while remaining within the SAR and scan time limits.

Chapter 4 investigates the feasibility of diffusion tensor imaging at high magnetic field (i.e. 7.0 tesla), by comparing the performance of these measurements with the comparable sequences performed at 1.5 T and0 3. T. Not only the SNR is investigated, but also the resulting uncertainties in fitted DTI parameters.

In Chapter 5, CEST measurements and T1 mapping are applied to a small group of tumor patients. In this study, conventional contrast enhanced T1‐weighted imaging (at 3 T) and fluid attenuated inversion recovery T2‐weighted imaging is compared with novel measures of CEST, MT, and qT1.

In Chapter 6, the added value of performing quantitative T1 mapping, MT, and CEST measurements at 7 T is explored in a small cohort of ALS patients. The performance of these novel methods is compared with DTI parameters in an effort to further elucidate the etiology of this devastating disease. Because ALS is a neurodegenerative disease which mostly affects the motor , DTI data was used as an anatomical selection tool to compare the different MR measures along individually selected corticospinal tracts of patients and controls.

33 Chapter 1

In Chapter 7, the main results from Chapters 2 to 6 are presented, the feasibility of quantitative methods at 7 T is discussed and possible directions for future research are outlined.

34 Introduction

References

1. Hanson LG. Is quantum mechanics necessary for understanding magnetic resonance? Concepts Magn Reson Part A. 2008 Sep 1;32A(5):329–40. 2. Rabi II, Zacharias JR, Millman S, Kusch P. A New Method of Measuring Nuclear Magnetic Moment. Phys Rev. 1938 Feb 15;53(4):318. 3. Damadian R. Tumor Detection by Nuclear Magnetic Resonance. Science. 1971 Mar 19;171(3976):1151–3. 4. Rooney WD, Johnson G, Li X, Cohen ER, Kim S‐G, Ugurbil K, et al. Magnetic field and tissue dependencies of human brain longitudinal 1H2O relaxation in vivo. Magn Reson Med. 2007 Feb;57(2):308–18. 5. Vaughan JT, Garwood M, Collins CM, Liu W, DelaBarre L, Adriany G, et. al 7T vs. 4T: RF power, homogeneity, and signal‐to‐noise comparison in head images. Magn Reson Med. 2001;46(1):24–30. 6. Conijn MMA. Imaging cerebral small vessel disease at 7 Tesla MRI [Phd Thesis]: Utrecht University; 2011 [cited 2012 Jul 22]. Available from: http://igitur‐ archive.library.uu.nl/dissertations/2011‐0411‐200351/UUindex.html 7. Koenig SH. Molecular basis of magnetic relaxation of water protons of tissue. Acad Radiol. 1996;3(7):597–606. 8. Wolff SD, Balaban RS. Magnetization transfer contrast (MTC) and tissue water proton relaxation in vivo. Magn Reson Med. 1989;10(1):135–44. 9. Wolff SD, Balaban RS. NMR imaging of labile proton exchange. J Magn Res. 1990 Jan;86(1):164–9. 10. Zhou J, Payen J‐F, Wilson DA, Traystman RJ, van Zijl PCM. Using the amide proton signals of intracellular proteins and peptides to detect pH effects in MRI. Nat Med. 2003;9(8):1085–90. 11. van Zijl PCM, Yadav NN. Chemical exchange saturation transfer (CEST): What is in a name and what isn’t? Magn Reson. Med 2011;65(4):927–48. 12. Dagher AP, Aletras A, Choyke P, Balaban RS. Imaging of urea using chemical exchange‐dependent saturation transfer at 1.5T. J Magn Reson Imaging. 2000 Nov 1;12(5):745–8. 13. Zhou J, Lal B, Wilson DA, Laterra J, Zijl PCM van. Amide proton transfer (APT) contrast for imaging of brain tumors. Magn Reson Med. 2003;50(6):1120–6. 14. Zhou J, Blakeley JO, Hua J, Kim M, Laterra J, Pomper MG, et al. Practical data acquisition method for human brain tumor amide proton transfer (APT) imaging. Magn Reson Med. 2008;60(4):842–9. 15. Wen Z, Hu S, Huang F, Wang X, Guo L, Quan X, et al. MR imaging of high‐grade brain tumors using endogenous protein and peptide‐based contrast. NeuroImage. 2010 Jun;51(2):616–22. 16. Zhou J, Tryggestad E, Wen Z, Lal B, Zhou T, Grossman R, et al. Differentiation between glioma and radiation necrosis using molecular magnetic resonance imaging of endogenous proteins and peptides. Nat Med. 2011 Jan;17(1):130–4. 17. Assessment of Glycosaminoglycan Concentration in Vivo by Chemical Exchange‐ Dependent Saturation Transfer (gagCEST). PNAS. 2008 Feb 19;105(7):2266–70.

35 Chapter 1

18. MRI Detection of Glycogen in Vivo by Using Chemical Exchange Saturation Transfer Imaging (glycoCEST). PNAS. 2007 Mar 13;104(11):4359–64. 19. Cai K, Haris M, Singh A, Kogan F, Greenberg JH, Hariharan H, et al. Magnetic resonance imaging of glutamate. Nat Med. 2012 Feb;18(2):302–6. 20. Zhou J, Zijl PCM van. Chemical exchange saturation transfer imaging and spectroscopy. Prog Nucl Mag Res Spec. 2006 May 30;48(2‐3):109–36. 21. Mougin OE, Coxon RC, Pitiot A, Gowland PA. Magnetization transfer phenomenon in the human brain at 7 T. NeuroImage. 2010 Jan 1;49(1):272–81. 22. Dula AN, Asche EM, Landman BA, Welch EB, Pawate S, Sriram S, et. al Development of chemical exchange saturation transfer at 7T. Magn. Reson. Med. 2011;66(3):831–8. 23. Zhao X, Wen Z, Huang F, Lu S, Wang X, Hu S, et al. Saturation power dependence of amide proton transfer image contrasts in human brain tumors and at 3 T. Magnetic Resonance in Medicine. 2011 Oct 1;66(4):1033–41. 24. Fick A. Ueber Diffusion. Annalen der Physik. 1855;170(1):59–86. 25. Einstein A. Über die von der molekularkinetischen Theorie der Wärme geforderte Bewegung von in ruhenden Flüssigkeiten suspendierten Teilchen. Annalen der Physik. 1905;322(8):549–60. 26. von Smoluchowski M. Zur kinetischen Theorie der Brownschen Molekularbewegung und der Suspensionen. Annalen der Physik. 1906;326(14):756– 80. 27. Beaulieu C. The basis of anisotropic water diffusion in the nervous system ‐ a technical review. NMR in Biomedicine. 2002;15(7‐8):435–55. 28. Mori S, Barbara J. Crain, V. P. Chacko, Peter C. M. Van Zijl. Three‐dimensional tracking of axonal projections in the brain by magnetic resonance imaging. Ann Neurol. 1999;45(2):265–9. 29. Basser P, Mattiello J, Le Bihan D. Estimation of the Effective Self‐Diffusion Tensor from the NMR Spin‐Echo. J Magn Res Ser B. 1994 Mar;103(3):247–54. 30. Jones DK. The effect of gradient sampling schemes on measures derived from diffusion tensor MRI: A Monte Carlo study. Magn Reson Med. 2004 Apr 1;51(4):807–15. 31. Chang L, Jones DK, Pierpaoli C. RESTORE: Robust estimation of tensors by outlier rejection. Magn Reson Med. 2005 May 1;53(5):1088–95. 32. Staempfli P, Reischauer C, Jaermann T, Valavanis A, Kollias S, Boesiger P. Combining fMRI and DTI: A framework for exploring the limits of fMRI ‐guide DTI fiber tracking and for verifying DTI‐based fiber tractography results. NeuroImage. 2008 Jan 1;39(1):119–26. 33. Sunaert S. Presurgical planning for tumor resectioning. J Magn Reson Imaging. 2006;23(6):887–905. 34. Pillai JJ, Zaca D, Choudhri A. Clinical impact of integrated physiologic brain tumor imaging. Technol. Cancer Res. Treat. 2010 Aug;9(4):359–80. 35. Horsfield MA, Jones DK. Applications of diffusion‐weighted and diffusion tensor MRI to white matter diseases ‐ a review. NMR Biomed. 2002;15(7‐8):570–7. 36. Mandl RCW, Rais M, van Baal GCM, van Haren NEM, Cahn W, Kahn RS, et al. Altered white matter connectivity in never‐medicated patients with schizophrenia.

36 Introduction

Hum Brain Mapp [Internet]. [cited 2012 Apr 6]; Available from: http://onlinelibrary.wiley.com/doi/10.1002/hbm.22075/abstract 37. Verstraete E, van den Heuvel MP, Veldink JH, Blanken N, Mandl RC, Hulshoff Pol HE, et al. Motor Network Degeneration in Amyotrophic Lateral Sclerosis: A Structural and Functional Connectivity Study. Zhan W, editor. PLoS ONE. 2010 Oct 27;5:e13664. 38. Yamada K, Sakai K, Hoogenraad FGC, Holthuizen R, Akazawa K, Ito H, et al. Multitensor Tractography Enables Better Depiction of Motor Pathways: Initial Clinical Experience Using Diffusion‐Weighted MR Imaging with Standard b‐Value. Am J Neuroradiol. 2007 Oktober;28(9):1668–73. 39. Tournier J, Mori S, Leemans A. Diffusion tensor imaging and beyond. Magnetic Resonance in Medicine. 2011;65(6):1532–56. 40. Grunkemeier GL, Wu Y. Bootstrap resampling methods: something for nothing? Ann Thorac Surg. 2004 Apr;77(4):1142–4. 41. Liu RY. Bootstrap Procedures under some Non‐I.I.D. Models. Ann Stat. 1988 Dec 1;16(4):1696–708.

37

Chapter 2. Uncertainty estimations for quantitative in-vivo MRI T1 mapping

Daniel L Poldersa, Alexander Leemans b, Peter R. Luijten a, Hans Hoogduin a, c a: Department of , University Medical Center Utrecht, Utrecht, the Netherlands. b: Image Sciences Institute, University Medical Center Utrecht, Utrecht, the Netherlands c: Rudolf Magnus Institute of Neuroscience, Department of Neurology and , University Medical Center Utrecht, Utrecht, the Netherlands. This chapter has been published in Journal of Magnetic Resonance vol 224 (2012), 53‐60.

Knowledge is an unending adventure at the edge of uncertainty. Jacob Bronowski, scientist, 1908 – 1974

39 Chapter 2

Abstract

Mapping the longitudinal relaxation time (T1) of brain tissue is of great interest for both clinical research and MRI sequence development. For an unambiguous interpretation of in vivo variations in T1 images, it is important to understand the degree of variability that is associated with the quantitative T1 parameter. This paper presents a general framework for estimating the uncertainty in quantitative T1 mapping by combining a slice‐shifted multi‐slice inversion recovery EPI technique with the statistical wild‐bootstrap approach. Both simulations and experimental analyses were performed to validate this novel approach and to evaluate the estimated T1 uncertainty in several brain regions across four healthy volunteers. By estimating the T1 uncertainty, it is shown that the variation in T1 within anatomic regions for similar tissue types is larger than the uncertainty in the measurement. This indicates that heterogeneity of the inspected tissue and/or partial volume effects can be the main determinants for the observed variability in the estimated T1 values. The proposed approach to estimate T1 and its uncertainty without the need for repeated measurements may also prove to be useful for calculating effect sizes that are deemed significant when comparing group differences.

40 Uncertainty estimations for quantitative in‐vivo MRI T1 mapping

Introduction

The longitudinal relaxation time constant, denoted by T1, is a fundamental parameter in understanding MRI contrast [1]. Water proton T1 is sensitive to many aspects of the local nuclear spin surroundings such as macromolecule concentration, water binding and water content. It is therefore expected to strongly reflect changes that can identify pathology. To allow careful comparisons of T1 values found in healthy and diseased tissue and perform longitudinal or inter‐subject comparisons, T1 measurements should be of high precision, i.e. the uncertainty in T1 should be small. In this work, we focus on determining this uncertainty of the estimated T1 parameter which describes the variance between measurements. By contrast, the accuracy of the T1 mapping MR sequence applied is not the focus of this work. Thus confounds that introduce a systematic bias from the true T1 value are not presented here.

In human studies, there are several subject‐dependent sources that can lead to increased uncertainty in the estimated T1. These sources include subject motion and physiological factors such as breathing, cardiac pulsation, head size and geometry. This issue is further complicated at higher field strengths by the subject‐dependent inhomogeneous B1 fields. All these factors affect the (temporal) stability of the MRI signal and, subsequently, increase the variation in the estimated T1 value. To interpret T1 results unambiguously it is important to gain insight into the uncertainty of the T1 measurement on a subject‐by‐ subject basis.

The most straightforward way to determine the variance between measurements is by repeating them a number of times. This approach, however, is often too time‐consuming, especially in a clinical context. Monte‐ Carlo simulations form another approach to determine the fitting uncertainty. With this method, the uncertainty of a parameter is derived by assessing the behavior of the fit by adding a known noise contribution to ground‐truth data and repeating this procedure many times. Although no additional measurements are needed with these simulations, the characteristics of the real data and noise are typically not fully known, in particular when pathology is

41 Chapter 2 involved. In addition, this approach implicitly assumes spatial and temporal homogeneity of the signal to noise ratio (SNR), which is generally not the case.

The “wild bootstrap” method bridges the gap between the aforementioned approaches by using the residuals from the model fit with respect to the actual measurements to create many synthetic datasets [2]. The variation found in fits of these datasets is a measure of the uncertainty of the fit.

Wild bootstrapping has already been applied successfully in the assessment of diffusion tensor imaging (DTI) parameters [3–8] and the determination of functional MRI resting state network nodes [9]. T1 fitting and conventional bootstrapping (i.e.: random permutations from repeated measurements) were used in spectroscopic imaging to assess the standard error of T1 fitting [10]. To our knowledge, wild bootstrapping has never been applied to T1 inversion recovery (IR) imaging data before.

In this study, we combined the wild bootstrap method with a multi‐slice T1 mapping method to estimate the associated T1 uncertainty map. Inversion recovery data with 23 time points was acquired using a slice‐shifted method proposed before [11]. We validated this approach in‐silico by comparing the estimated uncertainty values with variations found in simulated data with varying noise levels and in‐vivo by comparing the estimated uncertainty with measurements in healthy subjects. First, we compared the estimated uncertainty values from a single subject with the values found in simulations for corresponding noise levels. Second, we analyzed 10 consecutive measurements in the same subject to assess the bias between the estimated uncertainty and the sampled variance in repeated measurements for each voxel separately. Third, we assessed the reproducibility of the method by comparing pairs of scans in four subjects. The differences between the average T1 values and standard deviations (SD) in T1 values over selected regions of interest (ROIs) were compared. Finally, we present examples of how we can improve the interpretation of the fitted T1 maps with knowledge of the accompanying estimated uncertainty.

42 Uncertainty estimations for quantitative in‐vivo MRI T1 mapping

Theory

T1 fitting In this paper we assume that the MR signal follows mono‐exponential recovery of the longitudinal magnetization. This is strictly speaking incorrect, as studies have shown that even without obvious partial volume effects the longitudinal relaxation displays multi‐exponential behavior [12]. However, to be able to reliably estimate the shorter of the T1 components, the inversion curve should be sampled at sufficiently short inversion times and at close intervals. Neither of these two requirements is easily met in clinical inversion recovery sequences. In the data we present here the inversion curve is sampled relatively sparse and it is expected that short T1 components are simply not observed reliably. We model the inversion curve with the following relation:

tT//11 TRT  It I0 12 e  e  (1)

Where I(t) is the measured signal intensity at times t, T1 is the longitudinal relaxation time constant, and TR is the repetition time employed in the sequence.

Wild Bootstrap Method

The uncertainty of the T1 fits was estimated by applying a wild bootstrap method. This method involves many repeats of the fit procedure using   bootstrapped data, Ibootstrap , generated from the initial fit, I fit , and the model residuals,  . The central assumption here is that in the case of a successful fit with a valid model, the residuals represent equivalent samples from the same unknown noise distribution around zero. As a result, neither the location along the inversion curve nor the sign of the residual should greatly influence thet fi result. The variation in the fit result is assumed to be dominated by the overall scaling of the residual distribution (which is equivalent to the SNR of the measurement). A schematic overview of the wild bootstrapping method is shown in Figure 1.

43 Chapter 2

Data

1: Fit T1, I0

Est. Parameters

2: Create Bootstrap sample Repeat Bootstrap Data Nbootstrap times

3: Fit T1, I0

Bootstrap Fits

4: Calculate statistics

SD Mean (uncertainty)

Figure 1: Schematic overview of the wild bootstrapmethod.1: initial fit of the data. 2: Based on the estimated parameters and residuals, many synthetic bootstrap samples are created. 3: These bootstrap samples are fitted and 4: The standard deviation over the bootstrap fits is calculated. Bootstrap data were generated by taking the fitted curve and adding or subtracting randomly chosen residuals in the following manner. The initial collection of residuals of size N, where N is the number of inversion times minus 1 (i.e. 22), is denoted by  (eq. 2). This was combined with its negative  equivalent to give the collection  (size 2×N, eq.3).

     Imeasured I fit (2)

   ,  11 ,..., NN ,   ,...,   (3)

The values bootstrap were then randomly selected from this collection with equal probability, P, for all values, as per eq.4:

1 bootstrap, j 1,..., N i Pbootstrap, j i  (4) 2N

44 Uncertainty estimations for quantitative in‐vivo MRI T1 mapping

 Finally, the bootstrapped time‐curve intensities, Ibootstrap , were calculated by  taking the sum of the initial fitted intensities, I fit , and the resampled residual  values, bootstrap :

   IIbootstrap fit bootstrap (5)

In contrast to wild bootstrapping applications in DTI, we did not use a weighted linear least squares fit in our fitting. Therefore, there was no need to apply weights in a heteroskedasticity consistent covariance matrix estimator (HCCME) and this parameter was omitted.

This way, many different inversion curves are generated. These curves (1000 in our implementation) are all fitted. The standard deviation of the fitted parameters is calculated and we label this the estimated uncertainty.

We compared four kinds of variance. The first is the variance estimated using the bootstrap method, which we will call ‘uncertainty’. We compare this to the variance that we observe by repeated measurements and T1 fitting. We will label this the ‘repeated measures variance’ or RMV. This type of variance is the gold standard that we would like to estimate using the bootstrap method. Thirdly, we also compare the estimated uncertainties to the variance observed between different voxels that are not necessarily the same, but lare al within a single visually homogeneous ROI. Part of this variance is due to the variance of the second type (RMV) and part is due to tissue heterogeneity, the amount of which can be assessed by taking the root of the differences between the squared variances of the second and third type. We will label this type of variance ‘heterogeneity’. Lastly, we compare the uncertainty with the difference between the averages of T1 value within an ROI in two repeated scans. In this case we do not compare differences on a voxel‐by‐voxel level, but rather the differences of ROI averages. This describes another important aspect of the method, i.e. the practical reproducibility of the sequence. This measure of variance will be labeled ‘reproducibility’.

45 Chapter 2

All measures of variance are normalized by the observed mean of the data to give the relative standard deviations or coefficients of variation and are expressed in percentages of the mean.

Methods

Simulations To assess the performance of the wild bootstrap method to estimate the uncertainty in T1 fitting, we compare the estimated uncertainty from a single simulated dataset to the variation found in repeated simulations. The estimated uncertainty and RMV should be in close agreement for the uncertainty to be a meaningful parameter.

T1 relaxation data were simulated by creating a dataset with a single T1 relaxation time constant and multiple noise levels. The inversion recovery curve given in eq. 1 was used to simulate signal recovery after inversion. For the simulations, I0 was set to unity, T1 to 1500 ms, TR to 10000 ms, and the inversion times (t) were chosen similar to those in the MRI sequence from 20 to 5000 ms in 23 linear steps. In this way, the real part of the MR signal was simulated, while the imaginary part of the signal was considered null. Gaussian noise was added to both the real and imaginary channel, with noise amplitudes increasing linearly in 10 steps from 1% to 10% noise relative to I0. This corresponds to SNR levels of 100, 50, 33, 25, 20, 17, 14, 13, 11, and 10, which is a reasonable range for most MRI experiments. The magnitude of the complex signal was calculated, resulting in a simulation of the inversion recovery MR signal with Rician noise characteristics.

For each SNR level, 10000 repetitions were simulated. For each repetition, T1 and bootstrap uncertainty was estimated. The relative standard deviation with respect to the mean of the T1 value was then calculated over all repetitions, giving the RMV. This forms our gold standard of variance between measurements. The bias of the bootstrap uncertainty with respect to the RMV was calculated for each SNR level and expressed in % of the T1 value.

46 Uncertainty estimations for quantitative in‐vivo MRI T1 mapping

MRI measurements After signing an informed consent in agreement with the procedures set by the institution’s ethical board, five healthy volunteers were scanned. All measurements were performed on a 7 tesla Philips Achieva MRI system (Philips Medical Systems, Cleveland USA). A quadrature transmit‐receive head coil was used for transmission of the RF signal. Either a 16 or a 32 channel receive only Nova Medical head coil was used to receive the signals. The same scan parameters were applied for both receive coils.

The scan protocol included a 3.5 mm isotropic B0 mapping sequence. Shim fields up to the third order were fitted to this field map to obtain the optimum shim settings. These were then applied to the following scans.

A multi‐slice inversion recovery T1 mapping sequence based on the sequence presented by Ordridge and later by other groups [11, 13–16] was implemented. This sequence consists of a global non‐selective adiabatic inversion pulse, followed by sequential acquisition of the slices using slice‐selective 90° excitations and EPI read‐outs. During repeated acquisitions of the volume, the slice ordering is shifted, so all slices are acquired at a different inversion time. We applied a “left‐shift” of 2 slices, so when the slice order was initially: [1, 2, 3, 4, …, 45, 46], the following repetition applied the slice order: [3, 4, …,45, 46, 1, 2]. In this way, in 23 repetitions, all 46 slices were sampled at 23 different time points after inversion. This allows for robust fitting of the longitudinal relaxation time constant. The total scan duration for this sequence was 4 minutes and 10 seconds. Based on these parameters, the acquired inversion curves for the voxels in the odd slices differ from those for even slices. The resulting volumes consisted of slices that were sampled at 228 ms intervals with the odd slices starting at 20 ms and the even slices at 134 ms. The latest time points sampled for the odd and even slices were 5036 ms and 5150 ms, respectively.

The single‐shot EPI sequence that forms the basis of the T1 mapping method has the following parameters: FOV: 224×224×91.5 mm3 (RL, AP, FH), acquired (= reconstructed) pixel size 1.0×1.0 mm2 with a 1.5 mm slice thickness and a slice

47 Chapter 2 gap of 0.5 mm. The phase encoding direction was set to be anterior‐posterior, with the fat‐shift direction towards the anterior.

Each slice was acquired in a single shot employing a SENSE acceleration factor of 3.6 in the phase encoding direction and a half‐scan factor of 0.609, resulting in an EPI factor of 65. Fat suppression was accomplished by spectral inversion of the fat signal (SPIR). Repetition time and echo times were 10 s and 8.3 ms, respectively. The bandwidth in the read out and phase encoding directions were 1676 Hz and 19 Hz, respectively. Inversion was achieved by a non‐ selective adiabatic inversion; a 15 ms hyperbolic secant shaped pulse with an effective inversion bandwidth of 440 Hz [17].

For one subject the local SNR was calculated using pure noise data, obtained by acquiring one additional volume without any RF or gradient pulses. The resulting noise image incorporates the variation in noise level over the image due to SENSE reconstruction and non‐uniform filtering performed on the scanner. In another subject, the RMV of the T1 mapping sequence was compared with the estimated uncertainty by repeating the sequence ten times in one scan session. To assess the reproducibility of the method, another three subjects were scanned twice, repeating not only the T1 mapping sequence, but also the calibration scans such as B0 mapping and shimming, and B1 power optimization.

Data processing The acquired data were reordered and the polarity of the magnitude inversion curves was restored using the method proposed in [18]. As it is unknown from the magnitude data whether the smallest value in the curve lies just before or after the zero‐crossing, this value was removed from the fit. This results in 22 points along the inversion curve to fit. Parametric maps of the estimated values for T1, I0, sum of squares error and uncertainty values for T1 and I0 were calculated by fitting the polarity restored inversion curves to eq. 1. The parameters T1 and I0 were fitted by minimizing the sum of squares of the residual values of the fit and the data using a Levenberg‐Marquardt optimization. To reduce the number of required iterations and to stabilize the fit, the parameters were limited to the following parameter space (parameter:

48 Uncertainty estimations for quantitative in‐vivo MRI T1 mapping initial value, [range]): T1: 1200, [0 ‐ 10000] ms, I0: Ihighest,[0 – 10 × Ihighest], where I‐ highest is the highest signal intensity of the sampled curve. The uncertainty values were estimated using the bootstrap method described above, with 1000 bootstrap repetitions.

For one subject, an SNR map was obtained by calculating the local standard deviation of the noise image in a 9×9×9 voxel neighborhood. The SNR of the first image after inversion is calculated following the method described in [19]. Nine ROIs were placed in the following regions: Anterior / posterior , anterior / posterior white matter, caudate, putamen, thalamus, corpus callosum white matter and (see Figure 5). These regions display varying SNR and estimated uncertainty values. These value pairs were compared with the values found in simulated data at varying SNR levels.

Because the inversion time varies by slice in the acquired data, it is not possible to co‐register the raw data, as this would result in rotations through the slices. To enable pixel ‐wise comparison between ten consecutive scans, the fitted parametric volumes were co‐registered using rigid body registration (6 degrees of freedom), applying a cross‐correlation based cost function and tri‐linear interpolation. Then, the same nine ROIs were drawn on the first T1 map and propagated across these volumes. For each ROI, the RMV across the ten scans was compared with the estimated uncertainty derived from the first scan.

ROIs were drawn on each of the subject’s T1 maps separately to assess the reproducibility of the T1 mapping method and the uncertainty estimation in four subjects. Here, we compared the T1 values averaged over each ROI in two scans and the average uncertainty found in each ROI.

All image processing was performed using MIPAV [20] and JIST [21] combined with in‐house developed modules for slice reordering and quantitative T1 fitting. The wild bootstrap algorithm was also implemented in this versatile pipelining framework. All these algorithms are open source and are freely available via the JIST website: http://www.nitrc.org/projects/jist/. Generation of the simulated datasets and calculation of descriptive statistics and plotting was done using Matlab R2009b, the MathWorks Inc.

49 Chapter 2

Results

The bootstrap approach to estimate the T1 mapping uncertainty was first validated by simulations. The relative repeated measures variance (RMV) over 10000 simulated inversion curves was compared to the average uncertainty estimated for these curves by the bootstrap method using 1000 bootstrap samples. Figure 2 shows the RMV of the simulations and mean bootstrap uncertainty with respect nto the mea T1 for different noise levels. This figure also illustrates the difference between these two curves, showing an increasing underestimation of the RMV by the bootstrap results. For data with relative noise contributions less than 5%, the bootstrap method displays an underestimation between 6 and 8% of the RMV. At higher noise levels (lower SNR) this underestimation becomes larger, with 18% underestimation at 10% noise. Additionally, the correspondence of these simulations with in‐vivo data is shown by the stars in Figure 2. The mean uncertainties for the nine selected ROIs follow the expected SNR dependency closely.

Simulated variance and bootstrap uncertainty 8 A) RMV over 10000 simulations 7 B) Bootstrap Uncertainty (±sd) 6 difference (B-A) In-Vivo 5

4

3

2

1

Relative Variance (%) 0

-1

-2

-3 0 2 4 6 8 10 Relative noise contribution (%) Figure 2: Comparing the performance at varying noise levels of the bootstrap method with repeated measures variance (RMV) observed in simulations and measurements in vivo in a single subject. The mean in estimated uncertainty of 10000 simulated inversion curves is plotted with

50 Uncertainty estimations for quantitative in‐vivo MRI T1 mapping the corresponding standard deviation (open circles and error bar). For the RMV (open squares) and in vivo measurements (stars), no standard deviation is available. The difference between the uncertainty and the RMV (closed circles) illustrates the increasing underestimation of the RMV by the bootstrap uncertainty with increasing noise levels. Figure 3 illustrates the data quality obtained in the MRI measurements and shows slices from 16 of the 23 acquired volumes. Three inversion curves from the ROIs containing anterior white matter, posterior grey matter, and cerebro‐ spinal fluid are shown. To avoid showing the best or worst curve from the ROIs, the selected voxels scored the median value of the sum of squares errors for those ROIs. Note that the inversion data are well described by the fit (solid lines). Also, the distribution of residuals for all voxels combined closely follows a Gaussian curve centered on zero.

Ant. WM Pos. GM 1 1

T1 = 1132 ms Uncertainty = 1.5 % 0 20 ms 248 ms 476 ms 704 ms 0 T = 1820 ms Data Fit Residuals 1 Normalized signal Normalized signal Uncertainty = 1.4 % -1 -1

0 1000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000 Inversion Time (ms) Inversion Time (ms) 932 ms 1160 ms 1388 ms 1616 ms b) c) CSF Histogram of all residuals 1

1844 ms 2072 ms 2300 ms 2528 ms 0 Counts

T1 = 4417 ms

Normalized signal Uncertainty = 2.2 % -1

0 1000 2000 3000 4000 5000 -0.2 -0.1 0 0.1 0.2 Inversion Time (ms) Normalized residual intensity a) 2984 ms 3668 ms 4352 ms 5036 ms d) e) Figure 3: Raw data and inversion curves. (a) A single slice from 16 of the 23 volumes after reordering, so that each slice corresponds to one time point along the inversion curve. (b‐d) The acquired inversion curves for a single voxel located in anterior white matter (ant. WM), posterior grey matter (pos. GM) and cerebro‐spinal fluid (CSF), based on selected ROIs. Each voxel was selected by taking the median of the sum of squares error for that ROI. (e) The histogram of all residuals from all voxels in the ROIs, normalized to the fitted intensity I0. To further illustrate the results from this method in in‐vivo data, Figure 4 shows an example of different parameters resulting from the voxel‐wise T1 fit for a single axial slice. Maps for T1, I0, the sum of the squares of the residuals, and the corresponding relative uncertainty maps for T1 and I0 are presented. The enlarged areas depicted in f) and g) show abrupt boundaries between tissue and CSF. This location, where motion artifacts and partial volume effects are

51 Chapter 2 most likely to distort the T1 fit, shows up as a bright rim in the uncertainty image.

4500 a) b) c)

0 10% d) e) f)

g)

0%

Figure 4: Overview of T1 parameters. (a) Quantitative T1 map of one slice. (b) Fitted I0 values. (c) Residual sum of squares error. The second row displays the uncertainty in (d) T1 values and (e)I0. (f) and (g) show an enlargement of the boxes indicated in (a) and (d). Image intensities in (a) and (f) correspond to T1 values as per color‐bar top left. Scaling of (b) and (c) was set to maximize image clarity (arbitrary units). The uncertainties in (d), (e) and (g) are scaled between 0 and 10%, colorbar bottom left. Note that the effects of transmit and receiver coil sensitivities show up in the fitted values for I0, but not in the T1 maps. The agreement between repeated measurements and the calculated uncertainty values in‐vivo was assessed by comparing nine ROIs in ten repeated measurements in a single subject. Table 1 shows the ROI‐averages for the average T1 value (±SD) over repeated scans, heterogeneity over the ROI, averaged over ten scans, RMV over scans, and average estimated uncertainty per voxel. The heterogeneity over any ROI is larger than the mean RMV measured per voxel, which in turn is larger than the mean bootstrap uncertainty. The largest ROI heterogeneity is observed in the posterior cortical grey matter regions, while the smallest heterogeneity is seen in the deeper grey matter regions such as the caudate and putamen.

52 Uncertainty estimations for quantitative in‐vivo MRI T1 mapping

Nine regions of interest were selected in a single slice of the quantitative T1 maps to assess scan‐rescan reproducibility and to facilitate regional comparisons. These regions are indicated in Figure 5a. Figure 5b shows the

Bland Altman plot for the difference in T1 values for two repeated scans, for these nine ROIs over four volunteers. This illustrates the good reproducibility of the T1 mapping method (95% confidence interval ranges from ‐3.2 to 3.3 % of the measured value). Figure 5c compares the reproducibility of the average bootstrap uncertainties over these regions for two repeating scans. We summarize the T1 values and uncertainties found in this study in Table 2. This table also illustrates that the reproducibility of the T1 mapping method applied in this study is high; the between‐scan differences are well within the estimated uncertainty of these scans, indicating that no appreciable difference was found.

5 40

+2sd: 3.3% +2sd: 19.9%

mean: -0.4% 0 0 mean: 0.1%

-2sd: -20.8% scan-rescan difference(%) scan-rescan difference(%) -2sd: -3.2%

-5 -40 1000 2000 3000 4000 5000 6000 0 1 2 3 4 5

Mean qT1 (ms) Mean Unc (%) Frontal GM Putamen Post WM Frontal WM Thalamus Post GM Subjects Caudatus CC WM CSF a) b) c) Figure 5: Schematic overview of location of ROIs and Bland Altman plots of the test‐retest results illustrating the reproducibility of the method. (a) Anatomical left portion of selected ROIs in a single slice. (b) Mean difference of T1 values for each region versus the mean of the two repetitions of the T1 mapping procedure. (c) Mean difference of uncertainty in T1 values for each region versus the mean of two repetitions. Different colors indicate different regions and different markers indicate different subjects. Discussion and Conclusions

In this study, we applied the wild bootstrap method to estimate the uncertainty of T1 fits based on multi‐point inversion recovery MRI data. We compared the performance of this method with the RMV observed in repeated simulations and repeated measurements in healthy subjects.

53 Chapter 2

Validity of the model There are several assumptions that underlie the simple model for inversion recovery we have used in this study. First of all, the model assumes mono‐ exponential behavior. Looking at the residuals of the separate inversion curves, such as displayed in Figure 3, no consistent deviation from the model is observed. Additional autocorrelation analysis of the residual points in time for different lag values (data not shown) shows only a small effect in areas with especially pronounced partial volume effects with CSF, i.e. the boundaries with cortical GM and ventricles. Because magnetization transfer effects can also influence the observed longitudinal relaxation [22], we investigated whether the

SPIR fat suppression pulse had a significant effect on the estimated T1 value and the estimated uncertainty. Comparison in ROIs in a single subject showed that this was not the case (data not shown).

This simple model for inversion recovery also assumes full inversion of the longitudinal magnetization, while this might not always be the case. To account for this, it is possible to incorporate the degree of inversion into the model [23]. We performed the fit with the degree of inversion as a third parameter and found that this parameter is on average 1.01 (with a standard deviation of 0.1) for the selected regions. The mean R² values (adjusted for the number of coefficients in the fit) are 0.9924 and 0.9929 for the two and three component fit, respectively. Therefore, we conclude that fitting the degree of inversion does not add much to the performance of the T1 fitting performance.

Uncertainty and standard deviation in simulations The results shown in Figure 2 illustrate that with increasing noise levels, the proposed method displays an increasing bias in uncertainty with respect to the gold standard (repeated measures variance). This bias constitutes an underestimation of the true uncertainty, which is a known limitation of bootstrap based methods and was also noticed in the implementation in DTI experiments [3–5]. It is in the nature of the bootstrap method to underestimate the variance resulting from random noise in the data because of the method by which all fitting procedures work i.e.: minimization of the residuals. The assumption is that a minimal magnitude of residuals coincides with the fit of

54 Uncertainty estimations for quantitative in‐vivo MRI T1 mapping parameter values which is closest to their true value, but this is only valid for an infinite number of data points or infinite SNR. During the fitting iterations, the fitted parameter values can deviate from the ‘true’ and underlying parameter value, as long as the residuals become smaller. As a result, the error in the fitted parameters accounts for some of the magnitude of the residuals, making the residuals smaller than the true random error at the cost of a bias in the estimated parameter. Because these residuals are then used in the bootstrap method, this leads to an underestimation of the estimated variance. This effect increases with the noise level. Therefore it is important to keep in mind that the bootstrap method cannot give the exact uncertainty of a measurement, but rather estimates the lower bound value of uncertainty.

T1 values and uncertainty in-vivo The uncertainty values observed in a single subject were compared with the simulated uncertainty estimations at corresponding noise levels. We observe in Figure 2 that the values found in in‐vivo data correspond very well with what we expect from simulations.

To further investigate this method in‐vivo, the RMV over ten scans was compared with the estimated uncertainty of a single scan. As can be observed from Table 1, all ROIs show a larger RMV than uncertainty. Averaged over all ROIs, the absolute difference between RMV and estimated uncertainty is ‐1.3%. This constitutes an underestimation of the RMV by the bootstrap method of roughly 40%. For the cortical grey matter regions this is especially pronounced. In addition to the expected underestimation explained above, an additional explanation for the underestimation is that those regions are most susceptible to movement artifacts. Although the scans were co‐registered before comparison, this might still cause an undue increase in RMV. Furthermore, 10 repeated measurements is a relatively small number to accurately estimate the RMV and it is possible that this introduces additional upward bias to the observed RMV.

To assess the reproducibility of the method, we compared the performance in a scan‐rescan setting inr fou healthy volunteers. The results for these measurements show that this method allows for the determination of T1 values

55 Chapter 2 with a reproducibility of about 1%, regardless of ROI (Table 2). This is well below the estimated bootstrap uncertainty.

Interpreting T1 uncertainty maps and applications in research The primary function of the uncertainty maps presented in this work is to assess the quality of the fit, i.e. to highlight those areas where the data are not very well fitted by the model (due to partial volume effects) or otherwise corrupted (due to movement artifacts). Examples of areas with high uncertainty in the T1 fit are found at boundaries between different types of tissue, most prominently at the ventricular wall, where partial volume effects with CSF and cardiac pulsation cause the largest increase in uncertainty. A bright rim around the ventricles can be observed in Figure 4d), e) and g). Interestingly, areas in white matter generally show increased uncertainty compared to grey matter structures (Fig. 4 d and g). In addition to pulsation artifacts close to the ventricles, this can be explained by the relatively wide spacing of inversion times in our experiments (i.e. 228 ms). As white matter displays a T1 value around 1200 ms at 7 tesla, there are only 4‐5 points measured during the first T1 interval (where ~63% of the relaxation takes place). Therefore, fitting of short T1 constants is expected to be less robust than the fitting of voxels displaying longer T1 time constants.

The information in uncertainty maps is similar to that found in the map of residuals or sum of squares errors, which also reflects the success of the fit. However, the wild bootstrap method allows for estimation of the uncertainty of each fitted parameter separately. In this work, we have only discussed uncertainty values for the T1 parameters, but likewise uncertainty values are available for I0. Therefore, we can now assess the extent of how the quality of a fit works through as a minimum uncertainty in each fitted variable. Figure 4d) and g) show that the uncertainty can be as much as 10% in areas where partial volume effects or movement artifacts are particularly pronounced.

The second application of the uncertainty measure introduced in this work is to aid in the comparison of multiple T1 measurements. This is done by estimating the necessary sample sizes (either in subjects or number of voxels in an ROI) to be able to determine significant difference for a given effect size. In longitudinal

56 Uncertainty estimations for quantitative in‐vivo MRI T1 mapping studies, where one would like to follow changes in a single subject, it is often not feasible to conduct enough repeated measurements to get a good estimate of the uncertainty. Furthermore, multiple testing assumes constant uncertainty across the measurements, which might not be the case, especially in patient studies. Because no prior knowledge on the noise level or other sources of artifacts is required for the method presented here, it is particularly suitable in cases where the SNR is expected to vary from one measurement to the next.

Knowledge of the uncertainty also aids in separating the contribution of random measurement errors from true tissue heterogeneity over an ROI. Considering the thalamic region observed in this study, we have determined that in our group of subjects, this region displays a heterogeneity of 9.7%. The mean uncertainty of the fitted T1 values for this ROI, however, is only 2.4%. Therefore, only part of the measured variation can be explained by the estimated uncertainty of the fit.

Another application of uncertainty maps lies in the quantitative analysis of data and data quality monitoring. The uncertainty map can be used as a mask, only allowing voxels with an uncertainty lower than a certain threshold value in subsequent processing steps. This way, outliers are less likely to influence the results. Contrary to using the sum of squares error, we can now determine a threshold that has a physical rationalization (a percentage relative to the T1 value in stead of a value in arbitrary units). More sophisticated approaches such as using a weighted mean rather than a regular mean may now also be feasible for quantitative analyses. Additionally, because this method does not require additional scans, the uncertainty levels can be used routinely to identify datasets that suffer from increased uncertainties due to experimental errors such as subject motion or scanner instability.

Comparison with other studies

T1 mapping forms an immense body of work in MR literature and many studies have been published that focus on the optimization of T1 mapping techniques and the determination of measurement error and bias. See for an extensive review [24]. In many studies however, accuracy and uncertainty are determined only as a way to find optimal sequence parameters for the method [25] or to

57 Chapter 2 compare performance with respect to other methods. Many of the studies used repeated measurements to assess the variation over subjects or over repeated scans [26, 27], which is feasible in the development of a clinical study but not to be performed routinely. On the other hand, researchers have used computer generated datasets to assess the effect of SNR on the error of estimated parameters since the late seventies [28, 29]. The drawback of this is that SNR cannot always be considered constant over the imaging volume, or between measurements.

We have combined repeated measurements and the simulation of datasets to use the noise present in a single measurement to estimate the uncertainty of fitted parameters. We thus provide a method for researchers to assess the quality of each dataset without the need for extra scans. Another study that has approached the fit uncertainty as a practical measure of data quality was performed by Reddick et al. [30]. They used the chi‐squared (sum of all residuals) and coefficient of variance to determine the uncertainty of fitted T1 values. A cut‐off value on each measure of variance was used to classify voxels that showed unacceptable high uncertainty levels. In this approach however, it is difficult to translate the cut‐off level to a percentage of variation in T1, something that is easily accomplished using the method we have proposed.

Conclusions To conclude, we have introduced the wild bootstrap method to quantitatively assess the uncertainties of fitted T1 parameters, per measurement, for each individual voxel. This information contributes to a better assessment of the observed T1 values and its relation to tissue properties one would like to measure. The method takes into account to which extent partial volume effects, low SNR, image distortions, or subject motion will confound the fitted tissue T1 values. Knowledge of the estimated uncertainty in T1 for each specific region in the brain makes it possible to calculate sample sizes necessary to significantly measure differences in (large) group comparisons. This approach also facilitates the investigation of heterogeneity in T1 values separately from the estimated uncertainty of the method. We believe that information on the uncertainty of quantitative T1 per subject, without the need for repeated measurements, will

58 Uncertainty estimations for quantitative in‐vivo MRI T1 mapping contribute to more reliable and reproducible quantitative clinical MRI applications.

Acknowledgements

The authors would like to thank the anonymous reviewers for their constructive criticism, which helped us greatly to improve this paper.

59 Chapter 2

References

1. S.C.L. Deoni, Quantitative Relaxometry of the Brain, Top. Magn. Reson. Imag. 21 (2010) 101–113. 2. R.Y. Liu, Bootstrap Procedures under some Non‐I.I.D. Models, Ann. Stat. 16 (1988) 1696–1708. 3. B. Whitcher, David S. Tuch, Jonathan J. Wisco, A. Gregory Sorensen, Liqun Wang, Using the wild bootstrap to quantify uncertainty nin diffusio tensor imaging, Hum. Brain Mapp. 29 (2008) 346–362. 4. S. Chung, Y. Lu, R.G. Henry, Comparison of bootstrap approaches for estimation of uncertainties of DTI parameters, NeuroImage. 33 (2006) 531–541. 5. T. Zhu, X. Liu, P.R. Connelly, J. Zhong, An optimized wild bootstrap method for evaluation of measurement uncertainties of DTI‐derived parameters in human brain, NeuroImage. 40 (2008) 1144–1156. 6. D.K. Jones, Determining and visualizing uncertainty in estimates of fiber orientation from diffusion tensor MRI, Magn. Reson. Med. 49 (2003) 7–12. 7. D. Jones, Tractography gone wild: Probabilistic fibre tracking using the wild bootstrap with diffusion tensor MRI, IEEE. T. Med. Imaging. 27 (2008) 1268–1274. 8. D.L. Polders, A. Leemans, J. Hendrikse, M.J. Donahue, P.R. Luijten, J.M. Hoogduin, Signal to noise ratio and uncertainty in diffusion tensor imaging at 1.5, 3.0, and 7.0 Tesla, J. Magn. Reson. Im. 33 (2011) 1456–1463. 9. W.F. Auffermann, S.‐C. Ngan, X. Hu, Cluster Significance Testing Using the Bootstrap, NeuroImage. 17 (2002) 583–591. 10. H. Ratiney, S. Noworolski, M. Sdika, R. Srinivasan, R. Henry, S. Nelson, et al., Estimation of metabolite T‐1 relaxation times using tissue specific analysis, signal averaging and bootstrapping from magnetic resonance spectroscopic imaging data, Magn. Reson. Mater. Phy. 20 (2007) 143–155. 11. R.J. Ordidge, P. Gibbs, B. Chapman, M.K. Stehling, P. Mansfield, High‐speed multislice T1 mapping using inversion‐recovery echo‐planar imaging, Magn Reson Med. 16 (1990) 238–245. 12. J.R. Anderson, Q. Ye, J.J. Neil, J.J.H. Ackerman, J.R. Garbow, Diffusion effects on longitudinal relaxation in poorly mixed compartments, J. Magn. Reson. 211 (2011) 30–36. 13. S. Clare, P. Jezzard, Rapid T1 mapping using multislice echo planar imaging, Magn Reson Med. 45 (2001) 630–634. 14. D.C. Zhu, R.D. Penn, Full‐brain T1 mapping through inversion recovery fast spin echo imaging with time‐efficient slice ordering, Magn Reson Med. 54 (2005) 725– 731. 15. F. de Smit, H. Hoogduin, Fast whole brain T1 mapping at 3 Tesla, in: Proceedings 14th Scientific Meeting, International Society for Magnetic Resonance in Medicine, Seattle, Seattle, Washington, USA, 2006: p. 2403. 16. J. Grinstead, W. Rooney, Fast T1 Mapping in Human Brain Using Inversion Recovery EPI with GRAPPA at 3T and 7T, in: Proceedings 16th Scientific Meeting, International Society for Magnetic Resonance in Medicine, Toronto, Toronto, Canada, 2008: p. 3084.

60 Uncertainty estimations for quantitative in‐vivo MRI T1 mapping

17. M.S. Silver, R.I. Joseph, D.I. Hoult, Selective spin inversion in nuclear magnetic resonance and coherent optics through an exact solution of the Bloch‐Riccati equation, Phys Rev A. 31 (1985) 2753. 18. C.G. Bakker, C.N. De Graaf, P. Van Dijk, Restoration of signal polarity in a set of inversion recovery NMR images, IEEE T. Med. Imaging. 3 (1984) 197–202. 19. O. Dietrich, J.G. Raya, S.B. Reeder, M.F. Reiser, S.O. Schoenberg, Measurement of signal‐to‐noise ratios in MR images: Influence of multichannel coils, parallel imaging, and reconstruction filters, J. Magn. Reson. Im. 26 (2007) 375–385. 20. M.J. McAuliffe, F.M. Lalonde, D. McGarry, W. Gandler, K. Csaky, B.L. Trus, Medical Image Processing, Analysis & Visualization in Clinical Research, in: Computer‐Based Medical Systems, IEEE Symposium On, Los Alamitos, CA, USA, IEEE Computer Society, 2001: p. 0381. 21. B.C. Lucas, J.A. Bogovic, A. Carass, P.‐L. Bazin, J.L. Prince, D.L. Pham, et al., The Java Image Science Toolkit (JIST) for Rapid Prototyping and Publishing of Neuroimaging Software, Neuroinformatics. 8 (2010) 5–17. 22. A.M. Prantner, G.L. Bretthorst, J.J. Neil, J.R. Garbow, J.J.H. Ackerman, Magnetization transfer induced biexponential longitudinal relaxation, Magn. Reson. Med. 60 (2008) 555–563. 23. P. Wright, O. Mougin, J. Totman, A. Peters, M. Brookes, R. Coxon, et al., Water proton T 1 measurements in brain tissue at 7, 3, and 1.5T using IR‐EPI, IR‐TSE, and MPRAGE: results and optimization, Magn. Reson. Mater. Phy. 21 (2008) 121–130. 24 P.B. Kingsley, Methods of measuring spin‐lattice (T1) relaxation times: An annotated bibliography, Concept. Magnetic Res.. 11 (1999) 243–276. 25. M. Schabel, G. Morrell, Uncertainty in T‐1 mapping using the variable flip angle method with two flip angles, Phys. Med. Biol. 54 (2009) N1–N8. 26. G. Johnson, I.E.C. Ormerod, D. Barnes, P.S. Tofts, D. MacManus, Accuracy and precision in the measurement of relaxation times from nuclear magnetic resonance images, Br. J. Radiol. 60 (1987) 143–153. 27. R.G. Steen, S.A. Gronemeyer, P.B. Kingsley, W.E. Reddick, J.S. Langston, J.S. Taylor, Precise and accurate measurement of proton T1 in human brain in vivo: Validation and preliminary clinical application, J. Magn. Reson. Im. 4 (1994) 681–691. 28. T.P. Pitner, J.F. Whidby, Simulation of nuclear magnetic resonance spin lattice relaxation time measurements for examination of systematic and random error effects, Anal. Chem. 51 (1979) 2203–2206. 29. J.D. Decatur, T.C. Farrar, Experimental design optimization and confidence interval estimation in multiexponential NMR relaxation measurements, J. Phys. Chem. 93 (1989) 8294–8299. 30. W.E. Reddick, R.J. Ogg, R.G. Steen, J.S. Taylor, Statistical error mapping for reliable quantitative T1 imaging, J. Magn. Reson. Im. 6 (1996) 244–249.

61

Chapter 3. In Vivo Three-Dimensional Whole- Brain Pulsed Steady-State Chemical Exchange Saturation Transfer at 7 T

Craig K Jones1,2, Daniel Polders3, Jun Hua1,2, He Zhu1,2, Hans J. Hoogduin4, Jinyuan Zhou1,2, Peter Luijten3,4, Peter CM van Zijl1,2

1Division of MR Research, Russell H. Morgan Department of Radiology and Radiological Science, Johns Hopkins University School of Medicine, Baltimore, Maryland, USA. 2F.M. Kirby Research Center for Functional Brain Imaging, Kennedy Krieger Institute, Baltimore, Maryland, USA 3Department of Radiology, University Medical Center Utrecht, Utrecht, Netherlands 4Brain Division, University Medical Center Utrecht, Utrecht, Netherlands This chapter was published in Magnetic Resonance in Medicine, vol. 67, no. 6, pp. 1579–1589, 2012.

Knowledge is and will be produced in order to be sold, it is and will be consumed in order to be valorized in a new production: in both cases, the goal is exchange. Jean-Francois Lyotard, philosopher, 1924 – 1998

63 Chapter 3

Abstract

Chemical exchange saturation transfer (CEST) is a technique to indirectly detect pools of exchangeable protons through the water signal. To increase its applicability to human studies, it is needed to develop sensitive pulse sequences for rapidly acquiring whole‐organ images while adhering to stringent amplifier duty cycle limitations and specific absorption rate restrictions. In addition, the interfering effects of direct water saturation and conventional magnetization transfer contrast complicate CEST quantification and need to be reduced as much as possible. It is shown that for protons exchanging with rates of less than 50–100 Hz, such as imaged in amide proton transfer experiments, these problems can be addressed by using a three‐ dimensional steady state pulsed acquisition of limited B1 strength (~1 μT). Such an approach exploits the fact that the direct water saturation width, magnetization transfer contrast magnitude, and specific absorption rate increase strongly with B1, while the size of the CEST effect for such protons depends minimally on B1. A short repetition time (65 ms) steady‐state sequence consisting of a brief saturation pulse (25 ms) and a segmented echo‐planar imaging train allowed acquisition of a three‐dimensional whole‐brain volume in approximately 11 s per saturation frequency, while remaining well within specific absorption rate and duty cycle limits. Magnetization transfer contrast was strongly reduced, but substantial saturation effects were found at frequencies upfield from water, which still confound the use of magnetization transfer asymmetry analysis. Fortunately, the limited width of the direct water saturation signal could be exploited to fit it with a Lorentzian function allowing CEST quantification. Amide proton transfer effects ranged between 1.5% and 2.5% in selected white and grey matter regions. This power and time‐efficient 3D pulsed CEST acquisition scheme should aid endogenous CEST quantification at both high and low fields.

64 In Vivo 3D Whole‐Brain Pulsed Steady‐State CEST at 7 T

Introduction

Chemical exchange saturation transfer (CEST) (1–7) is a contrast mechanism based on the saturation of low‐concentration exogenous or endogenous pools of protons that are in constant exchange with the bulk water pool. One type of CEST imaging, amide proton transfer (APT) of mobile cellular peptides and proteins (8, 9), has a unique contrast that increases for brain tumors (8, 10) and reduces when pH is lowered during acute ischemia (9, 11). Currently, applicability of CEST/APT MRI in the clinic is limited by sensitivity (effect of a few percent), amplifier duty cycle, and specific absorption rate (SAR) restrictions that prohibit use of lengthy saturation schemes. In addition, CEST/APT quantification and image appearance may differ between hospitals due to the interference of competing saturation phenomena such as direct water saturation (DS) and conventional magnetization transfer contrast (MTC), the contributions of which depend on local hardware properties (coils, amplifiers) and on the pulse sequence parameters used. As such there is a need to develop sensitive pulse sequences for rapidly acquiring quantitative whole‐organ images while adhering to amplifier and SAR limitations.

The relative magnitude of CEST and conventional MT contrast can be weighted by varying the saturation length, tsat, and strength, B1 (12). It is not always appreciated that this weighting can be optimized for the particular exchange rate range of interest. Rapidly exchanging protons such as in paraCEST agents and hydroxyl‐ or amine‐based diaCEST compounds require high B1 fields to allow sufficient saturation efficiency α (2, 13, 14):

2  B 1   22 [1]  Bk1  sw in which γ = 267.5 × 106 rad/Ts and ksw is the exchange rate from solute to water.

For slower rates, such as those of amide protons in tissue (ksw =28 Hz; Refs.9 and13), only limited B1 strength is needed (Fig. 1). On the other hand, the DS width (15–17), MTC magnitude (18–20), and SAR decrease strongly when reducing B1. This suggests the opportunity to design a low‐B1 pulse sequence

65 Chapter 3 for APT in which maximum CEST saturation efficiency can be retained under conditions of strongly reduced MTC effects and a narrowed DS line shape.

Figure 1. Saturation efficiency as a function of saturation field strength (B1) for protons with a range of exchange rates (calculated using Eq. 1). Notice that maximum efficiency can be reached at low B1 for slow exchange rates. CEST volume acquisitions have been performed using multislice (21–23) and three‐dimensional (3D; Ref.24) approaches, with the latter having the advantage of the saturation being distributed equally over the volume. CEST imaging has been accomplished using long block pulses (10, 25) or a series of short pulses (8, 26–28) such that the saturation efficiency of the pulse train of pulses is similar to that of the long hard pulse. Alternatively, one can use steady‐state approaches (22, 29) with alternating brief saturation and image acquisition. It is expected from experience with MTC studies (18, 30, 31) that power deposition and thus scan times can be reduced using pulsed steady‐state acquisitions. Here, we use the Bloch equations to simulate the relative contributions of CEST, MTC, and DS as a function of saturation pulse length and strength with the goal of finding suitable 3D steady‐state CEST/APT acquisition parameters that have reduced DS and MTC interference and are not limited by amplifier duty cycle and SAR restrictions. In addition, we show that such acquisition removes the need for

66 In Vivo 3D Whole‐Brain Pulsed Steady‐State CEST at 7 T asymmetry analysis of the saturation effects with respect to the water frequency and allows quantification of CEST/APT effects.

Materials and Methods

Theory We focus on one type of endogenous CEST, namely APT of mobile cellular peptides and proteins (8, 9), for which the exchangeable protons of interest resonate around 3.5 ppm downfield from water (8). CEST effects can be predicted accurately by the Bloch equations, which can include CEST, DS, and MTC effects in a three‐compartment model. However, Bloch simulations are often not that transparent. As the amide proton exchange rate is only about 28 Hz (9, 13), it is also possible to use the analytical equations to verify results. The pure APT ratio can then be described by Refs.2, 9, and13:

tsat T1w APTR xssww k T1 1  e [2]  in which T1w is the longitudinal relaxation time of water and xs is the ratio of solute proton and water proton (111.2 M) concentrations. CEST/APT quantification requires the removal of competing effects of DS, which has been approached by measuring saturation effects as a function of offset frequency (so‐called Z‐spectra; Ref.32 or CEST spectra; Ref.25) and processing such data using so‐called MT ratio asymmetry (MTRasym) analysis with respect to the water resonance frequency. For the APT signals at 3.5 ppm from water this is:

Ssat 3.5 ppm Ssat  3.5 ppm MTRasym 3.5 ppm  [3] S0

However, such analysis is complicated by inherent asymmetries in the MTC effect (33–35) and by possible occurrence of so‐called nuclear Overhauser effects (NOEs; Refs. 2 and 36) upfield (i.e., at negative frequency offsets) from water.

Thus, the MTRasym measured in tissue is a convolution of multiple effects:

67 Chapter 3

endo MTRasym 3.5 ppm  APTR PTR3.5 ppm [4] MTC NOE MTRasym 3.5 ppm MTRasym 3.5 ppm in which PTRendo accounts for endogenous proton transfer effects other than due to amides. In general, it is assumed that MTRasym = APTR and images are called APT images, but it would be more correct to use APT‐weighted for the terminology. Each of the contributions in Eq. 4 depends differently on B1 and tsat, and, consequently, MTRasym(3.5 ppm) effects may vary (between −5% and 5%) between laboratories or even between different experiments in the same laboratory. Of course, saturation effects can only be positive, because additional water signal can unfortunately not be created and negative asymmetries are due to larger saturation effects upfield from water. In a practical approach (24, 37), APT images are often acquired using saturation parameters that provide approximately zero MTRasym(3.5ppm) for normal tissue, thereby highlighting lesions. This is not a major problem when using APT to determine relative contrast between affected and normal tissue such as in tumors or ischemia, but it may be an issue for diseases where lesions are less clear. Here, we investigate the convolved effects of MTC, DS, and APT during steady state, which reveal some troubling aspects when applying asymmetry analysis, leading to the need to analyze data without this approach.

Pulse Sequence and Simulations In Fig. 2, a 3D CEST technique is shown in which chemical exchange transfer is built up over multiple saturation pulses. After every short saturation pulse, a segmented echo‐planar imaging (EPI) readout is acquired. This fast readout and short repetition time (TR = 65 ms) result in an efficient whole ‐brain CEST acquisition of less than 11 s per irradiation frequency, allowing a full Z‐ spectrum to be acquired in a clinically reasonable time. The rapid scanning leads to a magnetic steady‐state situation.

68 In Vivo 3D Whole‐Brain Pulsed Steady‐State CEST at 7 T

Figure 2. Pulsed CEST acquisition consisting of a frequency‐selective sinc‐gauss saturation pulse followed by a short partial EPI readout. This TR interval is repeated continuously (166 times for the pulse sequence parameters described in the Materials and Methods section) with the saturation pulse applied at one frequency to fill 3D k‐space, starting with the high frequency region of k‐space during steady state build up. No intervolume delay is used when acquiring the multiple frequencies for a Z‐spectrum. To obtain insight into this, CEST, MTC, and DS effects for white matter (WM) were simulated using the Bloch equations for a three‐pool model (38, 39) consisting of macromolecular, amide, and bulk water protons. For convenience, the MTC effect was assumed to be symmetric so that MTRasym(3.5 ppm) equals the APT effect. These simulations were used to determine (a) the optimum B1 and duration of the saturation pulse for maximum APT effect and (b) the number of saturation pulses needed to obtain steady state. The situation of macromolecular and amide protons exchanging with the bulk water pool was modeled for 7 T based on previously published relaxation and exchange rates for free water and amide pools summarized in Table 1 (31, 38). Note that exchange rates, concentrations, and chemical shifts in ppm are not field dependent. The T2 values were estimated from 3 T literature (13, 18) by assuming a reduction of 40% with respect to 3 T. For convenience, all T1 values were set at 1.7 s. An offset frequency of 3.5 ppm was used for amide protons, while 0 ppm was used for all other protons. The APT and MTC effects were simulated as a function of saturation pulse power (0.1– 5 μT in steps of 0.1 μT)

69 Chapter 3

and duration (5–65 ms in steps of 5 ms) to determine parameters to maximize APT and minimize MTC. A 40 ms delay after each saturation pulse was simulated during which exchange and relaxation occur. This delay was based on the shortest possible TR (65 ms). The simulated signal was quantified at 301 saturation frequencies linearly spaced between −10 and 10 ppm. To determine the number of consecutive pulses required to reach steady state, the signal intensity after each sinc‐gauss pulse (applied on the APT frequency of 3.5 ppm) was plotted as a function of time postsequence onset. For comparison, a Z‐ spectrum was simulated with a continuous saturation pulse of duration 2500 ms and 1 μT at the same saturation frequencies as in the pulsed scheme. This length of pulse is similar to previous CEST work (24, 27) and was intended to be comparable to 100 repetitions of pulsed saturation.

Table 1. Parameters Used in the Three‐Pool Bloch Equation Simulations Proton Exchange Chemical concentration

T1 (s) T2 (ms) rate (s−1) shift (ppm) (mM) Ref. Macromolecule 1.7 0.01 4.6 0 7150 30 Amide 1.7 33a 28 3.5 72 13 Bulk water 1.7 48a – 0 111,200 18 a: Estimated based on the listed references, but adjusted for field strength assuming a reduction of 40% with respect to 3 T. Note that the value of T2 is needed for the Bloch simulation but will not strongly affect the CEST effect.

Data Acquisition The study was approved by the Johns Hopkins Medicine IRB and performed on six normal controls who provided informed consent. MRI data was acquired on a 7 T Philips Achieva system (Philips Healthcare, Best, The Netherlands) using a quadrature transmit head coil and a 32‐channel Nova Medical Inc (Wilmington MA) phased array receive coil. High dielectric pads (40) were placed on either side of the head by the temporal lobes for padding to minimize movement and to flatten the signal intensity across the head (40). Third‐order shims were optimized over the brain. Imaging data were acquired using the 3D multishot gradient‐echo (EPI factor 7) sequence of Fig. 2 with TR/TE/flip angle (FA) = 65 ms / 7.2 ms / 12° across 40 slices at 2 × 2 × 2 mm3 isotropic resolution across a field of view of 220 × 220 mm2. The multishot EPI used a linear readout

70 In Vivo 3D Whole‐Brain Pulsed Steady‐State CEST at 7 T so the zero k‐space profile was acquired half way through the volume acquisition. The parallel imaging SENSE factor was set to 2 × 2 (RL × AP). The saturation pulse was a 1 μT, 25 ms single‐lobe sinc‐gauss pulse. The TR of 65 ms was the shortest duration possible given the saturation pulse duration, readout, and delay imposed by the pulse sequence to allow for radiofrequency and gradient duty cycles. The time for whole‐brain acquisition per irradiation frequency at 2 × 2 × 2 mm3 resolution was 10.9 s. No intervolume delay was used after each volume acquisition. Following two dummy acquisitions, an unsaturated volume was acquired followed by 77 volumes at saturation frequency offsets of 0.0, ±0.2, ±0.4, ±0.6, ±0.8, ±1.0, ±1.2, ±1.4, ±1.6, ±1.8, ±2.0, ±2.2, ±2.4, ±2.6, ±2.9, ±3.1, ±3.2, ±3.3, ±3.4, ±3.5, ±3.6, ±3.7, ±3.9, ±4.1, ±4.4, ±4.6, ±4.8, ±5.0, ±5.5, ±6.0, ±6.5, ±7.0, ±7.5, ±8.0, ±8.5, ±9.0, ±9.5, ±10.0 ppm (relative to the water frequency). Total acquisition time was 14 min 24 s. The unsaturated reference was acquired using the same sequence as the saturated volume (including the TR) except that the radiofrequency saturation pulse was turned off.

Data Processing All data were registered to the first volume (unsaturated) using the rigid body (six degrees of freedom) registration algorithm FLIRT (FSL, FMRIB Centre, University of Oxford) with the normalized mutual information cost function and sinc resampling. All further data processing was done using Python, Scipy, and Numpy (www.phyton.org, www.scipy.org, and numpy.scipy.org, respectively) using code written in‐house.

A confound in selective frequency saturation is the underlying variability in the

B0 inhomogeneity due to local susceptibility‐induced field shifts. These frequency variations can be as high as 1 ppm in regions near the nasal cavity and ear canals, while smaller variations are seen across the whole brain. The advantage of Z‐spectra is that signals are acquired at many saturation offset frequencies, allowing the minimum signal (maximum DS) in each voxel to be found and then shifted to the assigned water frequency of 0 ppm for referencing. No obvious MTC effects could be discerned in the steady‐state Z‐

71 Chapter 3 spectra allowing us to fit the direct saturation contribution using a Lorentzian line shape (15) in each voxel:

2 LW 12 LAbLW, ,12 , fshift  100  A2 b [5] LW4 f f 12 0 shift where A is the amplitude, LW1/2 is the water linewidth, fshift is the frequency shift of the Z‐spectrum due to magnetic field inhomogeneities, and b is a baseline offset. The Lorentzian fit was based on points from the Z‐spectrum around the water frequency (|fshift| < 1 ppm) and points above 6 ppm from water

(fshift ≥ 6 ppm), where limited endogenous CEST and NOE effects are expected. Such Lorentzian fitting has been used previously in paraCEST imaging (41) when occurrence of CEST effects at positive and negative frequencies with respect to water prohibited the use of asymmetry analysis. However, for the B1 needed in paraCEST imaging, this would not be precise in vivo because MTC effects cause the water saturation line shape to be non‐Lorentzian. The

Lorentzian curve was used to shift the acquired data to correct for B0 inhomogeneity and to determine the CEST/APT effects. Two types of analyses were done for measuring APT numbers: (i) standard asymmetry analysis for the range 3.3–3.7 ppm according to Eq. 3 and (ii) Lorentzian difference analysis (LDA), in which the acquired data were subtracted from the Lorentzian curve and the mean signal around from 3.3 to 3.7 ppm was quantified without the need to use the upfield side as a control.

Regions were drawn on a midaxial slice through representative WM and GM regions and CSF. The mean MTRasym(3.5ppm) and APT signals within each region were calculated and displayed. A t‐test was used to test the hypothesis that these quantities differed between regions.

Results

Simulations In Fig. 3a, the three‐pool Bloch simulation of the APT effect is shown as a function of saturation B1 and duration. Maximal APTR (1.7%) was found for a

72 In Vivo 3D Whole‐Brain Pulsed Steady‐State CEST at 7 T

1 μT, 25 ms saturation pulse, and a TR of 65 ms. When simulating only the MT and DS effects for these saturation settings (Fig. 3b), it can be seen that MTC and DS effects are very small at 3.5 ppm. These optimized saturation parameters were subsequently used to simulate the signal intensity as a function of saturation pulse number (Fig. 3c). For the first volume, signals at 3.5 and −3.5 ppm were within 1% of the final steady‐state signal by the 60th and 57th TR interval, respectively, assuming an initial fully relaxed signal. For a TR of s65 ms, thi would correspond to 3.9 and 3.7 s of scanning time, which is approximately 1/3 of the k‐space volume. When acquiring a Z‐spectrum, the signal can be in steady state with far fewer pulses (and therefore much faster) by having no delay between the volume acquisitions at different frequencies. This is shown in Fig. 3c by the dotted and dash‐dotted lines for which the previous saturation point was 0.1 ppm away from the APT and control frequencies. When simulating steady‐state Z‐spectra with intervolume delays of 0 s (Fig. 3d) and 7 s (Fig. 3e), there was negligible difference, giving an APTR of 1.7% in both cases. It can be seen that Z‐spectra simulated using a continuous 1 μT hard pulse of length 2 s show strongly increased contributions from MTC and DS (dashed lines) compared to those simulated in steady state (solid lines). TheR APT quantified from the continuous hard pulse Z‐spectra was 1.3% for both 0 and 7 s intervolume delays. This reduced APT sensitivity compared to the steady‐state approach is likely due to the larger competing contributions from MTC and residual DS at the 3.5 ppm offset. Another Z‐spectrum was simulated (for the 0 s intervolume delay) using a continuous 0.4 μT hard pulse (length 2 s) to match the expected APTR of a 25 ms 1 μT sinc‐gauss pulse (Fig. 3c; dotted line). The saturation‐matched Z‐spectrum and difference compared well to that from the pulsed sequence.

73 Chapter 3

Figure 3. Simulations of the effect of saturation strength and duration on the pulsed steady‐state APT effect for a three‐compartment model of semisolid macromolecular protons, solute amide protons, and bulk water protons (Table 1). a: APTR for TR = 65 ms, showing a maximum at tsat = 25 ms dan B1 = 1 μT. b: MTC + DS effect at 3.5 ppm for TR = 65 ms using zero APTR. A strong reduction with decreasing B1 can be seen, leading to very small MTC for tsat = 25 ms at B1 values of 1 μT or less. c: Normalized water signal intensity (S/S0) as a function of number of TR intervals to determine the number of saturation pulses needed to reach steady state. Simulations for saturation at 3.5 ppm (APT frequency) and −3.5 ppm (control frequency) for an initially fully

74 In Vivo 3D Whole‐Brain Pulsed Steady‐State CEST at 7 T relaxed signal (solid and dashed lines, respectively) and for a previously obtained steady state at a frequency differing by 0.1 ppm (dotted and dash‐dotted lines, respectively). d: Comparison of Z‐ spectra for pulsed steady‐state CEST (solid curve), continuous saturation CEST (dashed line) with a 0 s intervolume delay and continuous saturation with B1 (0.4 μT) designed to match pulsed saturation (dotted line). The inset shows the MTRasym for both simulations. e: As (d) but with a 7 s intervolume delay. f: Mean signal intensity (and standard deviation) of an in vivo region in WM (major forceps) acquired from one acquisition with 0 s intervolume delay (dashed line) and one acquisition with 7 s intervolume delay (solid line). This pair of acquisitions was used to corroborate the simulations that showed little effect of the minimal intervolume delay.

Human Studies To corroborate the negligible dependence of the saturated steady‐state signal on intervolume delay, an in vivo experiment was done. A single volunteer was scanned twice using the pulsed CEST sequence with saturation offsets ranging from −1 to 1 ppm (21 step acquisition), both with a 7 and a 0 s intervolume delay. WM Z‐spectra from the major forceps were quantified (Fig. 3f). A spill‐ over effect from one saturation frequency to the next would skew the shape of the Z‐spectrum acquired with 0 s intervolume delay to the positive side, but this was not the case. This is consistent with the simulated results in Fig. 3d,e. Therefore, all in vivo data were acquired for APT quantification used the 0 s intervolume delay. The added benefit is that the acquisition is much faster and the number of TR intervals to get into steady state significantly reduced (as seen in the simulated data of Fig. 3c).

In Fig. 4a, an example slice stack is shown for the unsaturated volume of the steady‐state acquisition. Images as a function of saturation frequency for one anatomical slice are shown in Fig. 4b, indicating that there was sufficient saturation as reflected in the images near DS. Figure 5 shows a Z‐spectrum acquired data using 77 frequency offsets in the major forceps. The normalized signal intensities above 6 ppm are indistinguishable from 100%, suggesting negligible MTC effects. In addition, the saturated water line shape is very narrow. These features allowed the use of Lorentzian fitting to points between ±1 ppm and above |frequency offset| of 6 ppm (dashed line). This fit can be assumed to be dominated by DS effects and was used in each voxel to correct for the frequency shift of the Z‐spectrum due to B0 field inhomogeneities. The difference between the Lorentzian curve and acquired data should represent signal related to chemical exchange between tissue solute protons and bulk

75 Chapter 3 water. The downfield difference (shown in the inset in Fig. 5) revealed a broad group of resonances between 0 and 5 ppm that may be attributed to exchangeable protons. The range around 3.5 ppm was previously assigned to APT (42–44). Interestingly, there are appreciable saturation effects at frequencies lowern tha water (between 0 and −5 ppm). This is the range where Ling et al. (36) suggested the presence of NOE signals, which can probably be attributed to exchange‐relayed NOEs (2, 45). Such signals complicate the use of MTRasym(3.5 ppm) calculations for the amide protons. Fortunately, the minimal MTC effects in the low‐power pulsed CEST data allow the use of LDA to estimate the APT effects from the difference between the Lorentzian fit and the acquired data (see Materials and Methods section).

Figure 4. a: Slices of one volume (unsaturated) of the pulsed 3D acquisition to show the excellent image quality. b: The saturated volumes of one slice for saturation offset frequencies ranging from −10 to 10 ppm. The images near the middle are dark because of DS.

76 In Vivo 3D Whole‐Brain Pulsed Steady‐State CEST at 7 T

Figure 5. In vivo Z‐spectra for a region in WM (major forceps). Z‐spectrum acquired at frequencies denoted by the small black dots along the bottom. The lack of detectable saturation at frequencies >6 and <−6 ppm confirm the expected negligible MTC for the steady‐state acquisition. The acquired adat for the frequencies indicated by the large red circles were fitted to a Lorentzian function. The difference between the Lorentzian (green dashed line, assumed to be based only on DS) and acquired data indicate that exchange effects occur both at positive and negative frequencies with respect to water. Inset: The mean signal between the Lorentzian fit and data acquired between 3.3 and 3.7 ppm was used to calculate the APTR.

Spatial maps of the B0‐corrected APTR between 3.3 and 3.7 ppm created using this LDA technique are shown in Fig. 6a. For comparison, a traditional APT‐ weighted (asymmetry) map calculated using Eq. 3 is shown in Fig. 6b. The most important difference is that the signal in the new APTR map is positive, whereas it is mostly negative in the traditional asymmetry map, due to interference of the exchange‐relayed NOE signals. The mean APTR from six normal controls was quantified for nine brain regions (Fig. 7a,b and Table 2). There was no statistical difference (P > 0.01) between any tissue region (WM or RGM). APT values in all WM and GM regions were statistically different from the APT in CSF (P < 0.01), except for the frontal grey matter (GM). The lack of a statistical difference between frontal GM and CSF could be due to partial

77 Chapter 3 volume effects with CSF in the cortex. For comparison, the MTRasym at 3.5 ppm was calculated (Fig. 7c and Table 2), which is commonly used to determine APT differences in patients. Again, none of the tissues showed statistical differences (all P > 0.01). The only deviation from the mean was the CSF measure from volunteer 1, which we attribute to partial volume effects due to small ventricles. The other CSF values were close to zero.

78 In Vivo 3D Whole‐Brain Pulsed Steady‐State CEST at 7 T

Figure 6. a: An amide proton transfer ratio (APTR) map quantified from the mean signal between 3.3 and 3.7 ppm in the difference plot (Fig. 5 inset) on a per voxel basis. b: MTRasym(3.5ppm) map based on asymmetry analysis of the same data as in (a). c: Slice 20 of the APTR map in (a). d: Slice 20 of the MTRasym(3.5ppm) map in (b). Note the negative signal intensity in the MTR asymmetry method.

79 Chapter 3

Figure 7. a: Example regions drawn on a single slice of one control where 1 = genu of corpus callosum (gCC), 2 = splenium of CC (sCC), 3 = minor Forceps (minFor), 4 = major Forceps (majFor), 5 = anterior internal capsule (aIC), 6 = caudate nucleus (Caud), 7 = frontal cortical GM (frGM), and 8 = posterior cortical GM (postGM), and 9 = CSF. b: The mean and standard deviation of the APTR (%) for each volunteer organized by region. c: The mean and standard deviation of the MTRasym(3.5ppm) (%) for each volunteer organized by region.

80 In Vivo 3D Whole‐Brain Pulsed Steady‐State CEST at 7 T

Table 2. APT Effects Quantified for Nine Brain Regions (n = 6) in Normal Controls Region APTR (3.3–3.7 ppm) MTRasym (3.5 ppm) gCC 2.4 ± 1.4% −4.0 ± 2.4% sCC 2.9 ± 1.4% −2.8 ± 2.7% majFor 1.9 ± 0.4% −3.3 ± 1.0% minFor 1.8 ± 0.7% −4.1 ± 1.3% aIC 1.4 ± 0.8% −3.3 ± 1.0% caudate 1.4 ± 0.5% −2.7 ± 2.6% frGM 1.5 ± 0.9% −1.5 ± 1.3% posGM 1.8 ± 0.6% −1.5 ± 1.3% CSF 0.1 ± 1.3% −1.1 ± 0.7% Standard deviation is given.

Abbreviations—gCC: genu of corpus callosum; sCC: splenium of CC; majFor: major forceps; minFor: minor forceps; aIC: anterior internal capsule; Caud: caudate nucleus; frGM: frontal cortical grey matter; postGM: posterior cortical GM; CSF: cerebrospinal fluid. Discussion

We developed a fast low‐B1 pulsed CEST steady‐state acquisition able to acquire 3D whole‐brain volumes in less than 11 s per frequency offset at a 2 mm isotropic resolution. Bloch equation simulations were used to optimize the saturation pulse parameters with respect to reducing contributions from MTC and DS at 3.5 ppm while retaining close to maximal APT effects. The resulting parameters allowed acquisition within amplifier duty cycle limitations and are well within SAR guidelines, even at 7 T (SAR was approximately 1.5 W/kg). Interestingly, the resulting Z‐spectra for pulsed CEST showed prominent upfield exchange transfer effects, which confounded analysis of CEST effects using the commonly used approach of asymmetry analysis. The resulting data were therefore processed using Lorentzian fitting of the DS contribution and subtracting this from the experimental Z‐spectrum. This was possible due to the narrow appearance of the DS line shape, which was a consequence of the low B1 power used. To better facilitate such DS fitting, the sampling frequency of the Z‐spectrum around the water frequency was increased, which did not cause a major increase in overall scan time due to the short 3D volume acquisition time. LDA allowed quantification of the APT effect without being affected by potential MTC and exchange‐based asymmetries with respect to the water

81 Chapter 3 frequency. The resulting LDA‐based APTR maps differed strikingly from the

MTRasym(3.5 ppm)‐based images (Fig. 6) in that the range of values was more restricted (2.1 ± 0.6% versus −3.9 ± 1.6%, respectively) and positive. The most likely explanation is that the asymmetry measures vary due to intensity differences in the upfield region of the spectrum where the pulsed CEST acquisition showed significant signal saturation from −1 to −5 ppm.

The magnitude of both CEST and NOE effects is a function of saturation strength and duration and will most likely differ for different coil setups. Thus, absolute quantification of CEST effects will always require specification of B1, tsat, and the pulse sequence timing details, as corroborated by other papers (e.g., Refs.28,36). The finding of a negative MTR asymmetry is probably not restricted to pulsed style acquisition and it is clear that one must be careful in interpreting APT (and CEST) effects based on MTR asymmetry analyses. Importantly, saturation effects cannot be negative because this would correspond to an increase in water signal. Thus the asymmetry analysis data cannot be a correct value for APTR. Using LDA, the effect size measured in selected WM and GM regions reflects well the values found based on Bloch equation simulations (Fig.. 3) Using the values in Table 1 for the exchange rate,

T1w, and the ratio xs of solute protons versus water protons (111.2 M), the maximum effect (infinite saturation time tsat) for a B1 of 1 μT (267.5 rad/s) as based on the analytical solution is 1.4%, in good agreement with the simulation results (1.7%). Experimental APTR values (Table 2) varied from the same range to slightly higher, but such small discrepancies can easily occur based on differences in relaxation and exchange parameters between the simulations and experiments. We, therefore, conclude that, when MTC effects are small and the DS line shape is sufficiently narrow, the LDA approach provides a suitable alternative for quantifying APT. One surprising aspect of the results was that the MTC contribution was lower than simulated. We attribute this to B1 loss in the cables between the amplifier and the coil. Note that this should not affect the saturation efficiency for APT but will affect MTC strongly (Fig. 3a,b). For systems with more efficient power transfer, an even more reduced B1 setting may, therefore, be possible. Another potential factor affecting APTR is B1 inhomogeneity. Again, however, as shown in Fig. 1, this is not a main issue for

82 In Vivo 3D Whole‐Brain Pulsed Steady‐State CEST at 7 T slowly exchanging protons as long as variation is not more than 10–20% from the optimal B1. It is a major issue for paraCEST and amine and OH protons.

It is clear that the low‐power steady‐state pulsedCEST acquisition reveals additional information in the Z‐spectrum, which may provide information complementary to amide proton exchange. Most prominent are saturation effects upfield from water (−1 to −5 ppm), which are in the range where NOEs can be expected (2, 36). These could be due to both direct protein–water (36) or exchange‐relayed (2) NOEs, with the latter more likely based on the known time scale of such transfers for mobile proteins (43, 45, 46). These signals may offer new information for protein/peptide content or pH quantification independent of the amide proton peak at 3.5 ppm. We are currently investigating this in more detail. In addition, there is abundant CEST signal difference between 0 and 3 ppm from water. This is where other amides, amines, and hydroxyl groups resonate. It is important to realize that the relative contributions of these different groups to the signal will vary with the saturation parameters. For instance, high B1 approaches will allow detection of fast‐exchanging protons, while for low‐power approaches, such as used in the pulsed steady‐state CEST here, slowly exchanging protons such as the amides will be more visible (Fig. 1). It should be clear from this discussion that CEST and APT quantification are dependent on the pulse sequence parameters. In addition, changing the saturation power will affect both the direct saturation and MTC contribution in the Z‐spectrum. Until standardization of CEST approaches occurs, one must, therefore, be careful in comparing numbers between different sites.

Technical Details The data acquisition took 14 min and 24 s. This is mostly because of the very high spectral resolution (77 saturation frequencies) and we expect that this number can be reduced based on specific experimental needs. A more important detail is that the acquisition time was 11 s per 40‐slice 3D volume. The first in vivo APT measurements (10, 37) were single slice experiments with 1.5×1.5×5 mm3 resolution, taking approximately 30 s per slice. A recent multislice CEST FISP (28) sequence allowed acquisition of 13 slices in

83 Chapter 3 approximately 10 s. The only other 3D readout (24) used a GRASE sequence and was a 2.2×2.2×4.4 mm3 acquisition with 30 slices per volume and each volume acquired in 20 s. Thus the current sequence is the fastest per volume and has more slices per volume.

A single lobe sinc‐gauss pulse was used for saturation, and we found this simple pulse to provide good saturation and power tradeoffs with minimal side‐lobe contamination. Z‐spectra simulated with short square‐shaped pulses had large alternations in signal due to side lobes because of the sinc‐like profile in frequency space (data not shown). Other saturation pulses with more optimized shape should be possible but were not tested in this work.

The short‐TR pulsed CEST technique requires a sufficient duration to build up the saturation‐transfer steady state for amide protons exchanging with the water peak. For the 3D technique proposed here, we set the readout to begin at the high frequency end of k‐space to allow a sufficient number of saturation pulses to be applied before the phase encoding scheme encoded near the center of k‐space. Based on the simulations and the tissue parameters, approximately 50 TR intervals are required for the steady state to be reached (Fig. 3c). This will be dependent on T1w and TR. When performing lower resolution acquisitions, this may be problematic as the number of phase encodes reduces and therefore, the number of saturation pulses may not be sufficient to reach steady state within the first one or two frequencies of the Z‐spectrum. It should be noted that fewer TR intervals would be required to reach steady state at 3 T, due to the decreased T1w with respect to 7 T. Initially, we included an intervolume delay to reduce the possibility of saturation from one frequency over into the subsequent volume (saturated at a different frequency). For the study here, we found minimal bleed over, even without delay between volumes. The likely reason for this is the placing of the initial 50–60 saturation pulses at the high frequency k‐space lines.

Fat/lipid saturation may cause ‐based artifacts such as banding around the brain (24) or through the brain when using EPI‐type imaging (47). This generally becomes an issue when performing asymmetry analysis, which is not necessary to get APTR here. However, even the MTRasym(3.5ppm) images

84 In Vivo 3D Whole‐Brain Pulsed Steady‐State CEST at 7 T seem to be free of any lipid banding effects, such artifacts seem to be absent. We have no certain explanation for this, but it may be due to the overall low power used being insufficient for the lipid saturation (similar to the low MT effect). Other metabolites in principlen ca give NOEs in the upfield region, the size of which will vary depending on molecular motion and pulse sequence timing. We are currently looking into this.

In the simulations, all T1 values were conveniently set at 1.7 s. This is not expected to affect the results appreciably, because the exchange rates correspond to average life times of about 200 ms for the macromolecular protons and 35 ms for amide protons, both small relative to T1. In addition, T2s were estimated and thus not precise. However, this rate affects mainly the saturation efficiency (Zhou; Ref.13) and is generally negligible (neglected in Eq. 1 too).

The pulsed CEST sequence should also be suitable for lower field strengths such as 3 T, where body coil excitation is used. We confirmed this experimentally (results not shown).

The direct saturation line shape of the Z‐spectrum was fit using a Lorentzian function, but other interpolation options are possible, such as a spline (1) or high‐order polynomial (13). The Lorentzian was chosen because it properly describes the DS curve during steady state (15) and uses physically intuitive parameters. A more exact representation would require fitting of the whole Z‐ spectrum to the Bloch equations (e.g., Refs. 9, 14, and 20) based on three pools (bulk water, macromolecular, and amide). In normal brain tissue, it may be possible to estimate the pool parameters, but this would likely be more difficult in the presence of pathology (e.g., tumor, , lesions).

Conclusions

We showed that a pulsed saturation, steady‐state 3D whole brain CEST experiment is feasible on a high‐field (7 T) imaging system. The pulse amplitude and duration were optimized for maximal APT and strongly reduced MTC and width of the DS line shape. The narrow DS line could be fitted using Lorentzian line shape analysis, allowing us to quantify the APT

85 Chapter 3 effect independent of assumptions about exchangeable protons upfield from water. This technique is expected to be a highly efficient and clinically feasible method to acquire whole brain CEST data for slowly exchanging protons

(ksw < 50–100 Hz)

Acknowledgements

Equipment used in the study is manufactured by Philips. Dr. Craig Jones is partially paid through a grant from Philips Medical Systems to the Kennedy Krieger Institute. Prof. Dr. Van Zijl is a paid lecturer for Philips Medical Systems. Prof. Dr. Van Zijl is the inventor of technology that is licensed to Philips. This arrangement has been approved by Johns Hopkins University in accordance with its conflict of interest policies.

86 In Vivo 3D Whole‐Brain Pulsed Steady‐State CEST at 7 T

References

1. Ward KM, Balaban RS. Determination of pH using water protons and chemical exchange dependent saturation transfer (CEST). Magn Reson Med 2000; 44: 799– 802. 2. van Zijl P, Yadav N. Chemical exchange saturation transfer (CEST): what is in a name and what isnʹt? Magn Reson Med 2011; 65: 927–948. 3. Guivel‐Scharen V, Sinnwell T, Wolff SD, Balaban RS. Detection of proton chemical exchange between metabolites and water in biological tissues. J Magn Reson 1998; 133: 36–45. 4. Zhou J, van Zijl P. Chemical exchange saturation transfer imaging and spectroscopy. Progr Nucl Magn Reson Spectrosc 2006; 48: 109–136. 5. Sherry AD, Woods M. Chemical exchange saturation transfer contrast agents for magnetic resonance imaging. Annu Rev Biomed Eng 2008; 10: 391–411. 6. Terreno E, Castelli DD, Aime S. Encoding the frequency dependence in MRI contrast media: the emerging class of CEST agents. Contrast Media Mol Imaging 2010; 5: 78–98. 7. Aime S, Castelli DD, Crich SG, Gianolio E, Terreno E. Pushing the sensitivity envelope of lanthanide‐based magnetic resonance imaging (MRI) contrast agents for applications. Acc Chem Res 2009; 42: 822–831. 8. Zhou J, Lal B, Wilson DA, Laterra J, van Zijl PCM. Amide proton transfer (APT) contrast for imaging of brain tumors. Magn Reson Med 2003; 50: 1120–1126. 9. Zhou J, Payen J‐F, Wilson DA, Traystman RJ, van Zijl PCM. Using the amide proton signals of intracellular proteins and peptides to detect pH effects in MRI. Nat Med 2003; 9: 1085–1090. 10. Jones CK, Schlosser MJ, van Zijl PCM, Pomper MG, Golay X, Zhou J. Amide proton transfer imaging of human brain tumors at 3T. Magn Reson Med 2006; 56: 585–592. 11. Sun PZ, Zhou J, Sun W, Huang J, van Zijl PCM. Detection of the ischemic using pH‐weighted MRI. J Cereb Blood Flow Metab 2007; 27: .1129–1136 12. Desmond K, Stanisz GJ. Pulsed saturation transfer for quantifying CEST in the presence of MT. In: Proceedings of the 17th Annual International Society for Magnetic Resonance in Medicine, Honolulu, Hawaii, 2009. p 4494. 13. Zhou J, Wilson DA, Sun PZ, Klaus JA, Van Zijl PCM. Quantitative description of proton exchange processes between water and endogenous and exogenous agents for WEX, CEST, and APT experiments. Magn Reson Med 2004; 51: 945–952. 14. McMahon MT, Gilad AA, Zhou J, Sun PZ, Bulte JWM, van Zijl PCM. Quantifying exchange rates in chemical exchange saturation transfer agents using the saturation time and saturation power dependencies of the magnetization transfer effect on the magnetic resonance imaging signal (QUEST and QUESP): Ph calibration for poly‐L‐ lysine and a starburst dendrimer. Magn Reson Med 2006; 55: 836–847. 15. Mulkern RV, Williams ML. The general solution to the Bloch equation with constant RF and relaxation terms: application to saturation and slice selection. Med Phys 1993; 20: 5–13.

87 Chapter 3

16. Kim M, Gillen J, Landman BA, Zhou J, van Zijl PCM. Water saturation shift referencing (WASSR) for chemical exchange saturation transfer (CEST) experiments. Magn Reson Med 2009; 61: 1441–1450. 17. Smith SA, Bulte JWM, van Zijl PCM. Direct saturation MRI: theory and application to imaging brain iron. Magn Reson Med ;2009 62: 384–393. 18. Stanisz GJ, Odrobina EE, Pun J, Escaravage M, Graham SJ, Bronskill MJ, Henkleman RM. T1, T2 relaxation and magnetization transfer in tissue at 3T. Magn Reson Med 2005; 54: 507–512. 19. Henkelman RM, Stanisz GJ, Graham SJ. Magnetization transfer in MRI: a review. NMR Biomed 2001; 14:. 57–64 20. Wolff SD, Balaban RS. Magnetization transfer contrast (MTC) and tissue water proton relaxation in vivo. Magn Reson Med 1989; 10: 135–144. 21. Sun PZ, Murata Y, Lu J, Wang X, Lo EH, Sorensen AG. Relaxation‐compensated fast multislice amide proton transfer (APT) imaging of acute ischemic stroke. Magn Reson Med 2008; 59( 5): 1175–1182. 22. Dixon WT, Hancu I, Ratnakar SJ, Sherry AD, Lenkinski RE, Alsop DC. A multislice gradient echo pulse sequence for CEST imaging. Magn Reson Med 2010; 63: 253– 256. 23. Dula AN, Asche EM, Landman BA, Welch EB, Pawate S, Sriram S, Gore, JC, Smith, SA. Development of chemical exchange saturation transfer at 7T. Magn Reson Med 2011; published online Mar 22. 24. Zhu H, Jones CK, van Zijl PCM, Barker PB, Zhou J. Fast 3D chemical exchange saturation transfer (CEST) imaging of the human brain. Magn Reson Med 2010; 64: 638–644. 25. Ward KM, Aletras AH, Balaban RS. A new class of contrast agents for MRI based on proton chemical exchange dependent saturation transfer (CEST). J Magn Reson 2000; 143: 79–87. 26. Salhotra A, Lal B, Laterra J, Sun PZ, van Zijl PCM, Zhou J. Amide proton transfer imaging of 9L gliosarcoma and human glioblastoma .xenografts NMR Biomed 2008; 21: 489–497. 27. Mougin OE, Coxon RC, Pitiot A, Gowland PA. Magnetization transfer phenomenon in the human brain at 7 T. Neuroimage 2010; 49: 272–281. 28. Shah T, Lu L, Dell KM, Pagel MD, Griswold MA, Flask CA. CEST‐FISP: a novel technique for rapid chemical exchange saturation transfer MRI at 7 T. Magn Reson Med 2011; 65: 432–437. 29. Liu G, Ali MM, Yoo B, Griswold MA, Tkach JA, Pagel MD. PARACEST MRI with improved temporal resolution. Magn Reson Med 2009; 61: 399–408. 30. Sled JG, Pike GB. Quantitative imaging of magnetization transfer exchange and relaxation properties in vivo using MRI. Magn Reson Med 2001; 46: 923–931. 31. Smith SA, Farrell JAD, Jones CK, Reich DS, Calabresi PA, van Zijl PCM. Pulsed magnetization transfer imaging with body coil transmission at 3 Tesla: feasibility and application. Magn Reson Med 2006; 56: 866–875. 32. Bryant RG. The dynamics of water–protein interactions. Annu Rev Biophys Biomol Struct 1996; 25: 29–53.

88 In Vivo 3D Whole‐Brain Pulsed Steady‐State CEST at 7 T

33. Pekar J, Jezzard P, Roberts DA, Leigh JS, Frank JA, McLaughlin AC. Perfusion imaging with compensation for asymmetric magnetization transfer effects. Magn Reson Med 1996; 35: 70–79. 34. Swanson S, Pang Y. MT is symmetric but shifted with respect to water. In: Proceedings of the 11th Annual Meeting of ISMRM, Toronto, Canada, 2003. p 660. 35. Hua J, Jones CK, Blakeley J, Smith SA, van Zijl PCM, Zhou J. Quantitative description of the asymmetry in magnetization transfer effects around the water resonance in the human brain. Magn Reson Med 2007; 58: 786–793. 36. Ling W, Regatte RR, Navon G, Jerschow A. Assessment of glycosaminoglycan concentration in vivo by chemical exchange‐dependent saturation transfer (gagCEST). Proc Natl Acad Sci USA 2008; 105: 2266–2270. 37. Zhou J, Blakeley JO, Hua J, Kim M, Laterra J, Pomper MG, van Zijl PCM. Practical data acquisition method for human brain tumor amide proton transfer (APT) imaging. Magn Reson Med 2008; 60: 842–849. 38. Li AX, Hudson RHE, Barrett JW, Jones CK, Pasternak SH, Bartha R. Four‐pool modeling of proton exchange processes in biological systems in the presence of MRI‐paramagnetic chemical exchange saturation transfer (PARACEST) agents. Magn Reson Med 2008; 60: 1197–1206. 39. Woessner DE, Zhang S, Merritt ME, Sherry AD. Numerical solution of the Bloch equations provides insights into the optimum design of PARACEST agents for MRI. Magn Reson Med 2005; 53: 790–799. 40. Haines K, Smith NB, Webb AG. New high dielectric constant materials for tailoring the B1+ distribution at high magnetic fields. J Magn Reson 2010; 203: 323–327. 41. Liu G, Li Y, Pagel MD. Design and characterization of a new irreversible responsive PARACEST MRI contrast agent that detects nitric oxide. Magn Reson Med 2007; 58: 1249–1256. 42. van Zijl PCM, Zhou J, Mori N, Payen J‐F, Wilson D, Mori S. Mechanism of magnetization transfer during on‐resonance water saturation. A new approach to detect mobile proteins, peptides, and lipids. Magn Reson Med 2003; 49: 440–449. 43. Mori S, Eleff SM, Pilatus U, Mori N, van Zijl PC. Proton NMR spectroscopy of solvent‐saturable resonances: a new approach to study pH effects in .situ Magn Reson Med 1998; 40: 36–42. 44. Chen W, Hu J. Mapping brain metabolites using a double echo‐filter metabolite imaging (DEFMI) technique. J Magn Reson 1999; 140: 363–370. 45. Hwang TL, van Zijl PC, Mori S. Accurate quantitation of water‐amide proton exchange rates using the phase‐modulated CLEAN chemical EXchange (CLEANEX‐PM) approach with a Fast‐HSQC (FHSQC) detection scheme. J Biomol NMR 1998; 11: 221–226. 46. Liepinsh E, Rink H, Otting G, Wüthrich K. Contributions from hydration of carboxylate groups to the spectrum of water–polypeptide proton–proton Overhauser effects in aqueous solution. J Biomol NMR 1993; 3: 253–257. 47. Sun PZ, Zhou J, Sun W, Huang J, van Zijl PCM. Suppression of lipid artifacts in amide proton transfer imaging. Magn Reson Med 2005; 54: 222–225.

89

Chapter 4. SNR and Uncertainty in DTI at 1.5, 3.0 and 7.0 tesla

Daniel L. Polders MSc1, Alexander Leemans PhD2, Jeroen Hendrikse MD1, Manus J. Donahue PhD3, Peter R Luijten PhD1, Johannes M Hoogduin1,4 PhD

1: Department of Radiology, University Medical Center Utrecht, Utrecht, the Netherlands 2: Image Sciences Institute, University Medical Center Utrecht, Utrecht, the Netherlands 3: Department of Clinical Neurology, Oxford University, Oxford, United Kingdom 4: Rudolf Magnus Institute of Neuroscience, Department of Neurology and Neurosurgery, University Medical Center Utrecht, the Netherlands This chapter was published in Journal of Magnetic Resonance Imaging, vol. 33, no. 6, pp. 1456–1463, 2011.

These motions were such as to satisfy me, after frequently repeated observation, that they arose neither from currents in the fluid, nor from its gradual evaporation, but belonged to the particle itself. Robert Brown, botanist, 1773 – 1858, on the random movement of small particles in fluids.

91 Chapter 4

Abstract

Purpose: To compare Diffusion tensor imaging (DTI) measurements at ultra high field strength (7 tesla) in human volunteers with DTI measurements performed at 1.5 and 3 tesla.

Materials and Methods: The signal to noise ratio (SNR) and the uncertainty in fitted DTI parameters fractional anisotropy and primary eigenvector are assessed with tractography based regions of interest, measured in nine volunteers at 1.5 T, 3 T and 7 T with clinically available hardware configurations.

Results: An increase in SNR is observed on the 7 tesla system compared to the 1.5 or 3 T system. The measured increase in SNR at 7 T is larger than expected from field strength alone, indicating the large influence of improved receive coil hardware. Additionally, while the average fractional anisotropy remains relatively constant across field strengths, a decrease in uncertainty in the fitted values for fractional anisotropy and the principal eigenvector of the DTI tensor was found. Increased spatial heterogeneity of signal intensities is observed at 7 T.

Conclusions: Given the current hardware constraints, DTI at ultra‐high field strengths is possible with improved performance in selected regions of interest.

92 SNR and Uncertainty in DTI at 1.5, 3.0 and 7.0 tesla

Introduction

Diffusion tensor imaging (DTI) is an MRI method that uses directionally varying gradient fields to elucidate white matter architecture (1). DTI in combination with fiber tracking enables the visualization of white matter pathways in vivo (2). DTI is based on the application of strong gradient fields which cause an attenuation of the measured MR signal. Combined with long echo times, this makes DTI an inherently signal to noise ratio (SNR) sensitive technique (3,4). The introduction of ultra high field MRI scanners operating at field strengths above 3 T allows for imaging of the human brain at an increased SNR (5, 6). For DTI, this allows reduced scan time, increased spatial resolution, or increased diffusion weighting while retaining the SNR of the acquired images. The first steps towards these experiments at ultra high field strength have been reported recently (7, 8). However, there are several confounding factors at increased field strength that might actually annihilate the advantages of increased SNR.

Increased field strength substantially shortens transverse relaxation rates, with

T2 values in white matter dropping from 94 ms at 1.5 T to 77 ms and 50 ms at 3.0

T and 7.0 T (9). Related to shorter T2 values, previous investigations have questioned the expected gain in SNR of DTI at 7 T, reporting a 15% SNR loss for DTI measurements at 7 T as compared to 3 T (9, 10). Even more so, increased susceptibility effects resulting in much shorter T2* values may additionally affect the acquired diffusion weighted signals. As a result, it remains unclear whether 7 T DTI imaging translates to more sensitive and specific fiber tracking as compared to 1.5 or 3.0T.

DTI‐based parameters reflect the microstructure of tissue and are not expected to depend on the strength of the main magnetic field at which they are determined. The diffusion tensor is calculated using the voxel intensities of a collection of acquisitions, i.e. non‐weighted and diffusion weighted in different spatial directions. It is non‐intuitive how changes in SNR of these source images influence the fitted DTI parameters. Previous studies have shown that the uncertainty in DTI‐based parameters decreases with increasing SNR, while the accuracy remains largely unaffected (3). To assess the quality of the DTI data,

93 Chapter 4 wild bootstrapping has been developed to determine the uncertainty in fitted DTI parameters (11).

The aim of this study is to compare DTI at 1.5 T, 3 T and 7 T in the same volunteers using standard clinically available hardware. We investigated potential differences between these field strengths in terms of 1) the SNR of the b ~ 0 s/mm2 images, 2) the uncertainty in fractional anisotropy (FA) and 3) the uncertainty in the first eigenvector of the diffusion tensor. Global distributions of these parameters were inspected, as well as regional averages from nine volunteers. Three selected regions of interest (ROIs) were compared: sections of the corpus callosum (CC), cortico‐ spinal tract (CST) and cingulum bundles (CNG).

Methods

Nine healthy volunteers, with average age 27.6 years old (range 20‐54), signed an informed consent according to the guidelines of the local ethical committee. Three successive DTI experiments were performed on all volunteers on three Philips MRI scanners operating at field strengths of 1.5 T, 3 T and 7 T.

Data Acquisition For excitation the body coil was used at 1.5 and 3 T, and a quadrature volume T/R coil at 7 T. For signal reception an 8 channel parallel receive head coil was used at 1.5 T and 3 T and a 16 channel parallel receive head coil at 7 T.

Ass the focu of this work was on the performance of a given DTI sequence, rather than to optimize DTI sequences per scanner, a basic DTI sequence as provided by the vendor was applied. A single shot echo planar multi‐slice pulsed gradient spin echo sequence was used. The number of )(2 mm slices was 49, the EPI factor was 47, the field of view 224×224 mm2. Parallel imaging was applied using a SENSE acceleration factor of 2.5 and applying a half fourier factor of 0.68. The acquisition matrix of 112×109 was interpolated to a 128x128 matrix with isotropic voxels of 2 mm.

For each field strength, the experimental protocols were optimized and equalized as much as possible within the given hardware constraints. Gradient

94 SNR and Uncertainty in DTI at 1.5, 3.0 and 7.0 tesla performance is different on all three scanners and as a result, the diffusion, read‐out and echo times differ per scanner. The scanner dependent imaging parameters are listed in table 1. Diffusion measurements consisted of a baseline image at b ≈ 0 s/mm2 and 32 diffusion weighted images using b = 1000 s/mm2 applying the diffusion gradients along different spatial directions homogeneously distributed over a sphere.

A noise image was acquired by adding an additional image acquisition to the sequence without RF or gradient pulses. The measured signal constitutes a single sampling of the noise from the coil elements and incorporates reconstruction effects such as the parallel reconstruction sensitivities.

Table 1: Details of scanning parameters FIELD (tesla) 1.5 3.0 7.0 Repetition time (ms) 9 837 8 384 11 268 Echo time (ms) 65 51 71 Phase bandwidth (Hz) 21.0 29.7 29.6 Read bandwidth (Hz) 2101 2969 2555 Diffusion time (Δ) (ms) 32.0 25.1 35.2 Diffusion gradient dur. (δ) (ms) 17.5 13.3 23.8 Scan duration 6ʹ04ʹʹ 5ʹ10ʹʹ 6ʹ56ʹʹ Receive channels 8 8 16 Receive coil diameter (mm) 240 240 200 Max. gradient (mT/m) 66 80 30 Max. gradient slew rate (T/m/s) 50 100 100

Post-processing Calculation of the DTI parameters was performed with the ExploreDTI toolbox (12). Residual eddy currents and motion artifacts were corrected by aligning (3‐ D affine registration) the diffusion weighted images to the b = 0 s/m² image with reorientation of the B‐matrix (13). The noise image was processed without image registration.

Diffusion tensors were estimated by using a weighted linear least squares optimization with an anisotropic covariance matrix, which takes the log transform into account (14). In doing so, we mitigate the introduction of an

95 Chapter 4 additional bias, which is known to exist for the non‐weighted linear least squares fitting procedures (4).

Region Of Interest Selection Tract‐specific ROIs were selected by 1) performing whole‐brain deterministic streamline fiber tractography (15); 2) selecting a specific segment of well‐known anatomical fiber bundles and 3) selecting the voxels that contained one or more fibers. The aim was to select parts of the fibers that were well defined, compact and approximately straight. By choosing fiber bundles which were roughly perpendicular with respect to each other, potential directional differences in uncertainty in DTI parameters can be evaluated. Parts of the following pathways were segmented and used as the final ROIs: the corpus callosum (CC), cortical spinal tract (CST) and cingulum bundles (CNG).

Whole‐brain tractography was performed with the following settings: minimum FA = 0.2, maximum angle = 20°, step size = 1 mm, minimum fiber length = 50 mm. Seed points were distributed homogeneously over the brain volume, spaced 3x3x3 mm3 apart. Fibers were selected and cut between two masks that define the boundaries of tract segments and voxels that were intersected by these selected fibers were considered for further analysis in that ROI. Figure 1 shows an example of the selected tract segments.

96 SNR and Uncertainty in DTI at 1.5, 3.0 and 7.0 tesla

Fig. 1: Localization of Regions of Interest in a dataset acquired at 7 T. Three orthogonal slices and a 3‐D rendering showing the segments of fibers used in the selection; a segment of the corpus callosum, cingulum bundles and cortico‐spinal tracts. (Axial slice is shown from above.) The masks were drawn on color coded FA maps of cross‐sectional slices through the fibers, where the fibers were easily recognized. The masks were placed as follows: the CC masks were placed in sagittal slices, one at the left CNG and one at the right CNG. This allowed CC fibers to cover approximately 10 slices from left to right. For the CST, transversal masks were placed between the lowest level of the CC and lowest slice in the dataset. This resulted in approximately 15 slices superior‐to‐inferior. The two CNG masks were drawn on coronal slices, one located within the genu and one in the splenium of the CC, covering approximately 20 slices anterior‐to‐posterior.

Signal To Noise Calculation Signal to noise calculations were performed using the method described by Dietrich et al. (16), which is also described in the National Electrical

97 Chapter 4

Manufacturers Association standard ʹMS 1‐2001ʹ (17). This method utilizes the separate acquisition of a noise image, assumes a Rayleigh distribution for the acquired noise image and applies a scaling factor accordingly:

mean S r SNR  rROI [1] ROI 2 stddev N r  rROI 4 

Voxels in S(r) are selected from the tractography based ROIs in the b ~ 0 s/mm2 images. Voxels in N(r) are selected from the same ROI positions in the noise images. As follows from this method, it is possible to determine the local SNR. To do so, a 9x9x9 voxel neighborhood was used to calculate the local standard deviation of the noise per voxel. The rationale there is tha the global noise changes are due to the diminishing coil sensitivities towards the center, which span much larger regions than this neighborhood and that local standard deviations can be assumed constant over the neighborhood. This method allows for the calculation of SNR maps as a supplement of SNR averages over the selected ROIs calculated using equation [1].

Uncertainty Calculation Uncertainty measures were calculated using the wild bootstrap technique described by Whitcher et al (18). This allows for the estimation of the uncertainty in fractional anisotropy and principal eigenvector direction without the need for repeated measurements. Wild bootstrapping is a model‐based resampling technique, applied to the residuals of the diffusiont tensor fi at each voxel and can be specifically designed to work when the model is heteroscedastic, i.e. the variance of the errors is not constant for all observations. In the case of DTI, this corresponds to the assumption of non constant variance for the log‐transformed NMR signal. In this work wild bootstrapping was performed with 1000 repetitions and the resulting standard deviations of FA and cones of uncertainty (the standard deviation of the direction of the first eigenvector of the DTI tensor relative to the mean direction), are reported.

98 SNR and Uncertainty in DTI at 1.5, 3.0 and 7.0 tesla

Results

Tractography Based ROI Selection Streamline tractography was successfully performed on all acquired datasets. For the CC, the ROIs consisted of 2200 (± 330 across all datasets) voxels. The segmented CST fiber bundles consisted of 890 (±360) voxels. The CNG ROI consisted of 360 (± 110) voxels.

SNR Analysis Mean SNR values of b ~ 0 images over all volunteers were compared between 1.5 T, 3 T and 7 T for the ROIs in the CC, CST and CNG (Fig. 2). For the CC the mean (± standard deviation) SNR was 11±1, 24±3 and 41±6 at 1.5, 3 and 7 T respectively. For the CST these SNR values were 9±1, 19±3 and 30±6. For the CNG the SNR values were 10±1, 22±3 and 27±5. For all three field strengths the highest SNR was found in the CC. Averaged over all regions of interest, the relative ratios of SNR were found to be 1 : 2.1 : 3.2 for 1.5 T, 3 T and 7 T respectively.

50 CC CST CNG 45 40

35

30 SNR 25 20 15

10

1.5T 3T 7T 1.5T 3T 7T 1.5T 3T 7T Fig. 2: Box‐and‐whisker plot of SNR found in all subjects as a function of field strengths for different ROIs.

99 Chapter 4

Fig. 3a shows the SNR maps for a b ~ 0 volume at field strengths of 1.5 T, 3 T and 7 T in a single volunteer. At all field strengths the highest SNR is found close to the anterior and posterior surface of the brain, due to the smaller distance to the receive coils. Apart from CSF showing high signal intensity at the b = 0 images, we can also observe B1 inhomogeneity at 7 T. This effect causes low signal intensities especially inferior in the brain and in the lateral sides of the insula. Fig. 3b shows the histograms of the SNR values of the brain volume in the same volunteer. Notice the increased width of the distribution of the SNR values with increased field strength.

1.5 T 3 T 7 T LRLR 110

90

70

50 AP

AP 30

10 a LR a.u.

10 30 50 70 90 110 10 30 50 70 90 110 10 30 50 70 90 110 b Fig. 3: SNR maps of one volunteer over three field strengths. 1.5 (left), 3.0 (middle) and 7.0 tesla results (right). The rows show axial, sagittal and coronal slices, as well as histograms of the voxel values over the complete volume. All maps are scaled according to the adjacent intensity‐bar.

100 SNR and Uncertainty in DTI at 1.5, 3.0 and 7.0 tesla

Uncertainty In FA Fig. 4 shows directionally encoded FA maps for all three field strengths. The mean value of FA for each ROI was calculated for all subjects. The distribution over the group of subjects are shown in Fig. 5a. For the CC the mean FA, averaged over subjects, was relatively constant with 0.66±0.02, 0.65±0.02 and 0.64±0.03 at 1.5, 3.0 and 7 T respectively. For the CST these FA values were 0.64±0.03, 0.62±0.04, 0.62±0.04. For the CNG the FA values were 0.55±0.02, 0.52±0.03 and 0.52±0.03.

Fig. 4:Color‐coded FA maps of an axial slice of the same subject on three field strengths. Red, green and blue indicate regions with diffusivities oriented primarily left‐right, frontal‐posterior, and superior‐inferior respectively.

CC CST CNG CC CST CNG 0.18 0.65 0.16 0.6 0.14 FA 0.55 0.12 Uncertainty in FA

0.5 0.1

0.08 0.45 ab1.5T 3T 7T 1.5T 3T 7T 1.5T 3T 7T 1.5T 3T 7T 1.5T 3T 7T 1.5T 3T 7T

Fig. 5: Fractional anisotropy (a) and uncertainty in fractional anisotropy (b) for different ROIs and field strengths.

101 Chapter 4

The uncertainty in the calculated FA values was also assessed and averaged over the ROIs. The group distributions are shown in Fig. 5b. A clear decrease in the estimated uncertainty of FA can be observed for higher field strengths. To assess the distribution of FA uncertainty in a single volunteer, the FA uncertainty is mapped onto the fiber pathways, shown in Fig. 6.

Fig. 6: Saggital view of one volunteer of FA uncertainty mapped unto tracts, brighter sections indicating regions with higher uncertainty (standard deviation) in the calculated FA values.

Cone Of Uncertainty The effect of field strength on the uncertainty in the angular direction of the first eigenvector of the diffusion tensor can be described by an elliptical cone shape, known as the cone of uncertainty (19, 20). In this study we have used the circular representation of the cone of uncertainty, which is illustrated in Fig. 7. Here the cones of uncertainty in one volunteer on a single slice are shown. It can be observed that with increasing field strength, the angular uncertainty decreases. Again the average uncertainty for each ROI is calculated for all subjects. The group distributions, of the angular uncertainty are shown in Fig. 8. A strong decrease in uncertainty at higher field strengths can be observed. In the CC, the mean cone of uncertainty in all subjects decreases from 33° at 1.5 T to 24° at 3 T and 20° at 7 T. Similar trends are present in thed CST an CNG, with 38°/27°/23° and 29°/17°/15° respectively.

102 SNR and Uncertainty in DTI at 1.5, 3.0 and 7.0 tesla

Fig. 7: Cone of uncertainty map of a single slice of one volunteer, on three field strengths; 1.5 T (top), 3 T (middle) and 7 T (bottom). Images in the left column show the FA map, with the location of the cones of uncertainty map indicated by the box. The cones of uncertainty map in the right column shows a zoomed region, as indicated by the box in the middle column. The width and color of the cones reflect the uncertainty in the primary eigenvector of the diffusion tensor of corresponding voxels; narrow and blue cones indicate a low uncertainty while wided an red cones indicate a high uncertainty.

103 Chapter 4

CC CST CNG

40

35

30

25

20 Cone of Uncertainty (degrees)

15

1.5T 3T 7T 1.5T 3T 7T 1.5T 3T 7T Fig. 8: Cones of uncertainty for three ROIs and field strengths Discussion

The goal of this research was to investigate whether performing DTI at 7 tesla is beneficial when compared to DTI at 1.5 and 3.0 tesla. Two main descriptors of these benefits are SNR and the uncertainty in estimated DTI parameters. As SNR is gained it can be used to increase resolution, diffusion weighting, or simply to acquire images with a lower noise level. The effect is that subsequent fitting of the data to the diffusion tensor model is improved, reducing uncertainty in DTI estimates. The uncertainty in the FA and the principal eigenvector was assessed with wild bootstrapping and compared across the three field strengths. The results show a significant increase in SNR and decrease in the uncertainties of estimated FA and principal eigenvector for the compared ROIs.

SNR The focus of this work was to compare scanner performance, given the hardware configurations. Therefore, the results found in this study are confounded by differences in these configurations. Apart from the magnetic field strength, the most important influences on SNR are the difference in gradient performance and the coils used for signal reception. Differences in gradient performance result in a variation in the time needed to apply a

104 SNR and Uncertainty in DTI at 1.5, 3.0 and 7.0 tesla diffusion gradient pulse equivalent to b = 1000 s2/m, resulting in varying echo times. The difference in the number of receive coil elements, 8 channels at 1.5 and 3 T and 16 channels at 7 T, and a smaller inner diameter of the 7 T coil strongly influences the obtained signal to noise levels. Additionally, variations in gradient performance resulted in varying echo times. The effect on echo time is especially detrimental for the measurements performed at 7 T, due to both weaker gradients and shorter T2 values.

A secondary effect of the differences in gradient performance was expressed in a variation of repetition times for the three scanners. As this scan is not T1 weighted, it is not expected that this has a significant effect on the obtained SNR.

The MR signal from a given unit of tissue depends on (10):

1 TE SNR B0 exp fcoil [2] BW T2

Where B0 is the main magnetic field strength, BW the acquisition bandwidth, echo time TE, and transverse relaxation time T2. fcoil is a function describing the influence of transmit and receive coils on the signal and describes the effect the number of coil elements, their quality factors, geometry, spatial sensitivities, the effective g factor when using parallel imaging, etc. This function is spatially dependent, especially at higher field strengths.

In this study, an relative increase in SNR of 1 : 2.1 : 3.2 was found when averaging the ROI’s and subjects. At this point, it is of interest to discuss the potential origins of these changes in SNR. Is the measured gain in SNR the result of an increased B0 field or due to improvements in receive coil technology? Using the relation given in equation [2] and the acquisition parameters from Table 1, we can estimate the various sources of signal increase. To illustrate this, we distinguish three cases (see Table 2). The first case is the proton density weighted signal, where all signal parameters except for field strength are considered equal (i.e. ignoring both tissue and hardware effects: the theoretical upper limit). The second case represents an acquisition where only the differences in transversal relaxation rates are taken into account. This

105 Chapter 4 case approximates the signal measured at the shortest reasonable echo time of

51 ms. We can appreciate that the effect of T2 on the available signal is significant. This case also illustrates the potential gain in SNR when going to higher field strengths. The third case takes both the differences in T2 and gradient performance (affecting TE and BW) into account, but still neglects the coil effects. The reduced gradient performance at 7 T reduces the potential gain in SNR. When the gradient performance at 7 T is on par with that at 3 T, a greater increase in SNR is to be expected.

Table 2: Estimated effect of field strength, T2 and gradient performance on relative SNR, compared to measured SNR levels. The estimated values are normalized to the 1.5 T value of case 3. Case 2: Case 3: Case 1: Effect of Effect of Field Strength Effect of field Measured transverse gradient strength relaxation rate performance 1.5 T 2.0 1.2 ≡1.0 1.0 3.0 T 4.0 2.1 2.1 2.1 7.0 T 9.3 3.4 2.4 3.2

1 Estimated using equation 2, assuming fcoil to be unitary. 2 TE = 51 ms, T2 values used: 94 ms, 77 ms and 50 ms for 1.5, 3.0 and 7.0T respectively (9) 3 TE = 65 ms, 51 ms and 71 ms, BW = 21.0 Hz, 29.7 Hz, 29.6 Hz for 1.5, 3.0 and 7.0T respectively. The effect of the differences in receive coils were not approximated, but we believe it is reasonable to attribute the difference between case 3 and the actual measured SNRs to this effect. A large amount of the measured increase in signal must therefore be contributed to the increased performance of the receive coils. The receive coil used at 7 T consists of 16 elements, tightly fitting the subjects head. Although the individual elements are smaller, they are much closer to the tissue, resulting in higher signal per channel. Consequently, the recent introduction of 32 channel receive coils on 3 T and 7 T systems will further improve SNR (21).

Comparing SNR maps between field strengths, a loss in homogeneity of the acquired images can be observed at higher fields. The variance of the SNR over the ROIs, expressed as a percentage is approximately 15% for the 1.5 and 3 T

106 SNR and Uncertainty in DTI at 1.5, 3.0 and 7.0 tesla scanners and 20% at 7 T, which is comparable. The SNR maps (Fig. 3) however, show severe inhomogeneities. These inhomogeneities occur mainly at the periphery and inferior part of the brain. The main cause of this lies in the inhomogeneous RF excitation of spins, which is particularly pronounced at 7 T

(5). Further developments on coil hardware and B1 shimming and transmit SENSE at ultra high field are likely to improve these issues (22). With the current state of the art of 7 T DTI, the data acquired at 7 T does not allow for fiber tracking of the whole brain. Especially regions of the temporal lobe lack sufficient signal to perform robust fiber tracking. Given the current inhomogeneities in SNR, the influence on SNR of fiber orientation with respect to B0 remains unclear.

Uncertainties In DTI Parameters DTI measurements performed with increased SNR result in decreased uncertainty (4). With the use of the wild bootstrap method in the uncertainties of calculated DTI measures are shown to decrease with increased field strength. When uncertainties in FA and first eigenvector decrease, the usability of these measures may improve in fiber tracking applications both in single subject and longitudinal studies. Qualitatively, all datasets allow for extensive fiber tracking of the selected fiber pathways. Placing masks to segment fibers yielded similar numbers of voxels, which were used in subsequent analysis. In quantitative terms, the calculated DTI measures are comparable, with FA distributions overlapping but showing a trend towards a small decease in FA. This is in concurrence with previous work of Jones and Basser (4), where it was shown that at SNR <25 the FA is overestimated. Calculating FA at a higher SNR should thus yield a decreased FA, with a reduced overestimation.

In conclusion, given the shorter T2 values and the technical challenges of inherent B1 inhomogeneity using single excitation coils at 7 T, DTI is one of those techniques for which the actual gain by going to higher field strength is not obvious. This study investigated regions of interest chosen in segments of the corpus callosum, cortico spinal tract and singulum. SNR and the uncertainties of DTI based parameters at 7 T show substantial improvement compared with DTI results at 3 T or 1.5 T. It is shown that the observed increase

107 Chapter 4 in SNR can be attributed to increased signal from the stronger magnetic field and the improved receiver coil. The potential gain in SNR at 7 T has not been fully accomplished due to reduced gradient performance. However, the improved sensitivity of the receiver coil at this field strength compensates largely for this detrimental effect.

With further improvements in gradient performance and (multi transmit) B1 optimization strategies to come, DTI at 7 T holds great promise.

108 SNR and Uncertainty in DTI at 1.5, 3.0 and 7.0 tesla

References

1. Basser P, Pierpaoli C. Microstructural and physiological features of tissues elucidated by quantitative‐diffusion‐tensor MRI. J Magn Reson Ser B 1996:111:209‐ 219. 2. Mori S, Zijl PCMV. Fiber tracking: principles and strategies ‐ a technical review. NMR Biomed 2002:15:468‐480. 3. Farrell JA, Landman BA, Jones CK, Smith SA, Prince ,JL Zijl PCV, Mori S. Effects of signal‐to‐noise ratio on the accuracy and reproducibility of diffusion tensor imaging‐derived fractional anisotropy, mean diffusivity, and principal eigenvector measurements at 1.5 T. JMRI 2007;26:756‐767. 4. Jones DK, Basser PJ. Squashing peanuts and smashing pumpkins: How noise distorts diffusion‐weighted MR data. Magn Reson Med 2004;52:979‐993. 5. Vaughan J, Garwood M, Collins C, Liu W, DelaBarre L, Adriany G, Andersen P, Merkle H, Goebel R, Smith M, Ugurbil K. 7T vs. 4T: RF power, homogeneity, and signal‐to‐noise comparison in head images. Magn Reson Med 2001;46:24‐30. 6. Zwanenburg J, Hendrikse J, Visser F, Takahara T, Luijten P. Fluid attenuated inversion recovery (FLAIR) MRI at 7.0 Tesla: comparison with 1.5 and 3.0 Tesla. European Radiology 2010;20:915‐922. 7. Mukherjee P, Hess CP, Xu D, Han ET, Kelley DA, Vigneron DB. Development and initial evaluation of 7‐T q‐ball imaging of the human brain. Magn Reson Imaging 2008:26:171‐180. 8. Mukherjee P, Chung S, Berman J, Hess C, Henry R. Diffusion Tensor MR Imaging and Fiber Tractography: Technical Considerations. AJNR 2008;29:843‐852. 9. Cox E, Gowland P. Measuring T2 and T2ʹ in the brain at 1.5 T, 3 T and 7T using a hybrid gradient echo‐spin echo sequence and EPI. In: Proceedings 16th Scientific Meeting, International Society for Magnetic Resonance in Medicine, Toronto. 2008 p. 1411. 10. Speck O, Zhong K. Diffusion Tensor Imaging at 7T: Expectations vs. Reality Check. In: Proceedings 17th Scientific Meeting, International Society for Magnetic Resonance in Medicine, Honolulu. 2009 p. 1462. 11. Chung S, Lu Y, Henry RG. Comparison of bootstrap approaches for estimation of uncertainties of DTI parameters. NeuroImage 2006;33:531‐541. 12. Leemans A, Jeurissen B, Sijbers J, Jones D. ExploreDTI: a graphical toolbox for processing, analyzing, and visualizing diffusion MR data. In: Proceedings 17th Scientific Meeting, International Society for Magnetic Resonance in Medicine, Honolulu. 2009 p. 3537. 13. Leemans A, Derek K. Jones. The B‐matrix must be rotated when correcting for subject motion in DTI data. Magnetic Resonance in Medicine 2009;61:1336‐1349. 14. Basser P, Mattiello J, Le Bihan D. Estimation of the Effective Self‐Diffusion Tensor from the NMR Spin‐Echo. J Magn Reson Ser B 1994;103:247‐254. 15. Basser PJ, Sinisa Pajevic, Carlo Pierpaoli, Jeffrey Duda, Akram Aldroubi. In vivo fiber tractography using DT‐MRI data. Magnetic Resonance in Medicine 2000;44:625‐632.

109 Chapter 4

16. Dietrich O, Raya JG, Reeder SB, Reiser MF, Schoenberg SO. Measurement of signal‐ to‐noise ratios in MR images: Influence of multichannel coils, parallel imaging, and reconstruction filters. Journal of Magnetic Resonance Imaging 2007;26:375‐385. 17. National Electrical Manufacturers Association (NEMA). Determination of signal‐to‐ noise ratio (SNR) in diagnostic cmagneti resonance imaging. In: NEMA Standards Publication MS 1‐2001. Rosslyn: National Electrical Manufacturers Association; 2001 p. 15. 18. Whitcher B, David S. Tuch, Jonathan J. Wisco, A. Gregory Sorensen, Liqun Wang. Using the wild bootstrap to quantify uncertainty in diffusion tensor imaging. Human Brain Mapping 2008;29:346‐362. 19. Jones DK. Determining and visualizing uncertainty in estimates of fiber orientation from diffusion tensor MRI. Magnetic Resonance in Medicine 2003;49:7‐12. 20. Cheng Guan Koay, Nevo U, Lin‐Ching Chang, Pierpaoli C, Basser P. The Elliptical Cone of Uncertainty and Its Normalized Measures in Diffusion Tensor Imaging. , IEEE Transactions on 2008;27(6):834‐846. 21. G.C. Wiggins, C. Triantafyllou, A. Potthast, A. Reykowski, M. Nittka, L.L. Wald. 32‐ channel 3 Tesla receive‐only phased‐array head coil with soccer‐ball element geometry. Magnetic Resonance in Medicine 2006; 56: 216‐223. 22. Setsompop K, Alagappan V, Gagoski B, Witzel T, Polimeni J, Potthast A, Hebrank F, Fontius U, Schmitt F, Wald LL, Adalsteinsson E. Slice‐selective RF pulses for in vivo B1 inhomogeneity mitigation at 7 tesla using parallel RF excitation with a 16‐ element coil. Magnetic Resonance in Medicine 2008;60(6):1422‐1432.

110 SNR and Uncertainty in DTI at 1.5, 3.0 and 7.0 tesla

111

Chapter 5. Magnetization exchange effects and quantitative T1 mapping in tumor patients at 7 T

Daniel Polders1, Craig Jones2,3, Jeroen Hendrikse1, Piere Robe4, Eduard Voormolen4, Peter Luijten1, Hans Hoogduin1,5

1: Department of Radiology, UMC Utrecht, Utrecht, The Netherlands 2: Division of MR Research, Russell H. Morgan Department of Radiology and Radiological Science, Johns Hopkins University School of Medicine, Baltimore, Maryland, USA. 3: F.M. Kirby Research Center for Functional Brain Imaging, Kennedy Krieger Institute, Baltimore, Maryland, USA. 4: Department of Neurosurgery, UMC Utrecht, Utrecht, The Netherlands 5: Brain division, UMC Utrecht, Utrecht, The Netherlands This chapter is in preparation for submission to Journal of Magnetic Resonance Imaging

Exchange is creation. Muriel Rukeyser, poet, 1913 – 1980

113 Chapter 5

Abstract

This study presents magnetization exchange and quantitative T1 measurements in a small group of intracranial tumor patients at 7 T. The results from a recently introduced method for chemical exchange saturation transfer measurements were compared for several regions of interest, comparing healthy and diseased tissues. To distinguish the different effects contributing to signal saturation, magnetization transfer ratio images resulting from saturation at +3.5 ppm, ‐3.5 ppm, and asymmetry scores were analyzed. These magnetization exchange based measures were quantitatively compared with T1 maps. The magnetization transfer ratio images acquired with saturation at ‐3.5 ppm displayed the largest contrast between diseased tissue types and normal appearing white matter. In addition, a strong correlation between T1 values and magnetization transfer images and asymmetry scores was observed. This illustrates that many factors influence the observed saturation effects and care should be taken when interpreting asymmetry results based on these measurements.

114 Magnetization exchange effects and qT1 mapping in tumor patients at 7 T

Introduction

MRI methods to characterize tumor tissue heterogeneity and predict, monitor and guide tumor treatment are becoming increasingly important. In spite of the widespread use of conventional (spin denstity, T1 and T2 weighted) and contrast enhanced (dynamic susceptibility enhanced) MRI, there is still a need for more sensitive and specific methods for tumor characterization, preferably based on endogenous contrast mechanisms. In this study we have combined the high senstitivity of ultra high field MRI (7 T), quantitative T1 and magnetization exchange effects to compare different contrast mechanisms in a smal group of intracranial tumor patients.

Chemical exchange saturation transfer (CEST) is a class of MRI methods that aims to sensitize the MRI signal to reflect the chemical exchange of labile protons from specific molecules to the bulk water pool (1,2). The approach to measure CEST is similar to magnetization transfer (MT) techniques, i.e. off‐ resonance saturation transfers to the bulk water pool, causing a reduction of the measured on‐resonance signal. In many CEST experiments, the saturation frequency is varied over multiple repeated saturation experiments, probing the frequency dependency of saturation transfer and resulting in a Z‐spectrum (3). CEST measurements that focus on amide protons as the exchangeable proton of interest (known as amide proton transfer, APT), have been shown in vivo to display contrast between healthy tissue and tumor (4,5), ischemic (6), and necrotic (7) tissue. Several underlying mechanisms that could explain the observed APT effects have been suggested. In tumors, the increase in APT signal was explained by an increase of amide protons due to increased protein expression levels as shown by proton MR spectroscopy (8). In ischemia, APT differences were suggested to be a result of a change in acidity that alters the exchange rates (6). Finally, increased APT effects are expected to be observed in the case of cellular degeneration and necrosis, where the (partial) breakdown, dissociation and unfolding of proteins exposes backbone amide groups to the bulk water (9), enabling more amide protons to participate in exchange with the bulk water pool.

115 Chapter 5

In past research, asymmetry analysis has been applied to discriminate the CEST effect from conventional MT effects. By working under the assumption that the MT effect is approximately symmetric around the water resonance, the amount of CEST is calculated by determining the difference between saturation effects downfield from those symmetrically upfield from the water resonance. However, studies have shown very low or even negative CEST values in healthy tissue. In the paper by Ling and coworkers (10), the probable source of this negative asymmetry score was identified as nuclear Overhauser effects (NOEs) from (non‐exchanging) aliphatic protons, which manifest when saturating upfield from the water frequency. In the following we investigate the combined magnetization exchange effects, comprised of conventional MT, APT, and NOE.

There are several potential advantages of applying this promising technique at higher field strengths, i.e. 7 tesla. The differences in APT between normal and pathological tissue are expected to be in the range of only a few percent, thus an increased SNR allows for more robust measurement at reasonable resolutions. The increased chemical dispersion at higher field strengths (11) allows for better separation of the saturation frequencies and less contamination by direct saturation of the bulk water pool. T1 is longer at higher field strengths (12) and thus saturation of the bulk water pool persists longer, allowing for more accumulated saturation and potentially improving the sensitivity of the measurement.

The application of CEST at higher field strengths has been hampered by the stringent SAR constraints. Scan time increases dramatically at higher field strengths, for many CEST methods depend heavily on lengthy and high amplitude saturation pulses. Recently, a new steady state approach to obtaining CEST information was introduced, which applied small saturation pulses interleaved with a segmented read‐out scheme and therefore benefits from reduced SAR depositioning (13). This approach allows for rapid acquisition of the volumes, while achieving sufficient saturation in clinically feasible scan time.

116 Magnetization exchange effects and qT1 mapping in tumor patients at 7 T

There are several topics that are explored in this study. First, the effect of pathology on the measured saturation effect on the separate downfield and upfield images are of interest, as they can elucidate how contrast may change as a function of different relaxation and magnetic transfer mechanisms. The combined saturation effects as a result of APT, NOE and conventional MT effects are investigated in healthy and diseased tissue. Secondly, because the amount of saturation build‐up is dependent on the T1 of the bulk water pool, it is of interest to investigate these effects in pathology where T1 values are known to be altered. The separate effects that contribute to signal saturation are expected to depend on T1 differently. Finally, it is of interest to investigate how

B1 inhomogeneity effects influence the measured magnetization exchange effects.

In this work, the initial results with frequency specific magnetization exchange measurements at 7 T in a small heterogeneous group of brain tumor patients is presented. These results are compared with conventional qualitative measurements, i.e. contrast enhanced T1 weighted imaging (CE‐T1w) and fluid attenuated inversion recovery (FLAIR). The effects of B1 inhomogeneities on the acquired magnetization exchange results are qualitatively discussed. To study the role of T1, the quantitative MTR scores are compared with quantitative T1 data in selected ROIs.

Materials and Methods

Patient selection Six intracranial tumor patients were included after signing an informed consent in agreement with the guidelines set by the local ethical board. They received the 7 T scanning in addition to the regular 3 T pre‐operative protocol. The surgical procedure consisted either of complete resection of the tumor mass, or the collection of biopsy samples for pathological characterization. The PA results for these patients are summarized in Table 1.

117 Chapter 5

Table 2: Patient characteristics Subject Sex Age Gd enhancement? PA grading 1 F 41 Yes meningioma (WHO grade 1) 2 M 48 No oligodendroglioom (WHO grade 2) 3 M 24 No oligo‐astrocytoom (WHO grade 2) 4 M 60 Yes glioblastoom multiforme (WHO grade 4) 5 F 59 Yes glioblastoom multiforme (WHO grade 4) 6 F 62 Yes glioblastoom (WHO grade 4)

MRI sequences: The following MRI sequences were acquired on the patients. Please note that the 7 T scanning was performed before the 3 T protocol, to prevent any interference of the Gd contrast injection.

At 3 T, a T1 weighted post Gd image was acquired to visualize the enhancing region of the tumor. This sequence applied a sagittal 3D fast field echo (FFE), TE / TR = 14 / 475 ms, reconstructed voxel size 1 mm³, field of view (FOV): 240×240×180 mm³ (AP×FH×RL), acquired after injection of 0.1 ml/kg of a gadolinium‐containing contrast agent (Gadobutrol, Gadovist 1.0 mmol/mL, Bayer Schering Pharma, Newbury, UK).

All other scanning was performed on a 7 tesla Philips Achieva MRI system (Philips Medical Systems, Cleveland USA). For excitation, a quadrature transmit‐receive head coil was used. Either a 16 or a 32 channel receive only Nova Medical head coil was used to receive the signals. For both receive coils, the same scan parameters were applied.

The scan protocol included a 3.5 mm isotropic B0 mapping sequence. Shim fields up to the third order were fitted to this field map to obtain the optimum shim settings. These were then applied to the following scans.

A fluid attenuated inversion recovery (FLAIR) sequence with magnetization preparation was used to visualize edema and fluid‐rich tissue regions following the method by Zwanenburg (14). While this sequence was acquired at a resolution of 1 mm³, it was reconstructed to 1.5 mm³ isotropic for ease of subsequent registration steps and ROI delineation.

118 Magnetization exchange effects and qT1 mapping in tumor patients at 7 T

The effective B1+ amplitude was determined per subject using the method described by Sacolick and coworkers (15). The method acquired images at a resolution of 2.8×2.8×5 mm, covering a field of view of 200×250×35 mm (RL‐AP‐ FH). The phase shift was realized by an 8 ms, 500° Fermi pulse, played 1.5 kHz off‐resonance. B1 maps were calculated off‐line by algorithms developed in‐ house using the Matlab programming language (v 7.9.0, The Mathworks, Natick, Massachusetts USA).

CEST measurements The CEST measurement was performed following the method described by Jones in (13, see Chapter 3 of this thesis). This method employs a series of saturation and acquisition elements, reaching steady state before the central k‐ space line is reached. The volumes were acquired in a 3D segmented (multishot) EPI readt‐ou scheme, where each shot was preceded by a saturation pulse. In our implementation, the saturation pulse was optimized for the maximum amount of asymmetry, not to minimize concomitant MT effects. Previous research indicated that for relatively slowly exchanging protons (i.e. amide protons) the observed proton transfer rate varies very rlittle fo saturation powers between 1.0 μT and 2.0 μT (16). As B1 inhomogeneities are considerable at 7 T, there are regions in the brain where no more than 80% of the desired B1 amplitude is achieved. Therefore, it is prudent to use slightly larger pulses to compensate for this. As a result, the saturation pulse was a 1.8 μT 50 ms block pulse, followed by a crusher gradient to remove the net transverse magnetization. The volume was acquired at a resolution of 2 mm³ isotropic, with a field of view of 224×224×100 mm³ (RL×AP×FH). The EPI factor was 15, along the AP direction; this resulted in 190 shots to acquire each volume in 20.3 seconds.

To illustrate this method in a healthy volunteer, a Z‐spectrum spanning ±16.8 ppm was sampled densely at 67 equidistant frequency offsets (step size: 0.5 ppm). The total scan duration, including 2 seconds relaxation delay between the volumes, was 25 min and 17 seconds. This way, the peak saturation effects at ±3.5 ppm due to APT and NOE can be clearly observed in white matter, while being absent in CSF.

119 Chapter 5

For the patient scans, rather than densely sampling the Z‐spectrum, which would be prohibitively lengthy in the patient protocol, we opted to sample the Z‐spectrum only at the frequencies of interest, similar to the approach described by Zhou and coworkers (17). In total, 18 volumes were acquired, including the unsaturated volume. Off‐resonance frequencies were set to ‐16.78, ‐4.03, ‐3.69, ‐ 3.35, ‐3.02, ‐2.68, ‐0.5, ‐0.25, ‐0, 0.25, 0.50, 2.68, 3.02, 3.35, 3.69, 4.03 and 16.78 ppm with respect to the water resonance. The total scan time for this sequence was 6 min 40sec.

The saturation observed at a given saturation frequency is the sum of several mechanisms of magnetization transfer. Four factors we distinguish are: direct saturation (DS) of the bulk water pool by the saturation pulse, ‘conventional’ magnetization transfer (MT) from water bound to the surfaces of large macromolecules to bulk water by dipole‐dipole interactions, chemical exchange of molecules with labile protons, and nuclear Overhauser effects (NOE) from non‐exchanging aliphatic protons to either directly to bulk water, or via fast‐ exchanging protons (such amine or hydroxyl groups). Amide protons form an endogeneous pool of labile protons that are directly coupled to the amount proteins and peptides and display a resonance frequency of +3.5 ppm down field from the water resonance (≡ 0.0 ppm). The NOE effects are thought to arise from interactions of bulk water pool with aliphatic protons that resonate up field from bulk water over a wider range, spanning ‐2 to ‐4 ppm (13).

To distinguish the different saturation effects often an asymmetry analysis is applied, where saturation effects from DS and MT are assumed to be symmetrical around the water resonance pool. The two remaining effects, CEST and NOE, both add to the measured saturation, but on opposite sides from the water resonance. As such, the effects counteract each other when looking at the asymmetry score and simple subtraction of one side from the other is more likely to obscure than to elucidate the effects that may be caused by disease.

Because of the combination of different effects contributing to the signal suppression, we will use the more general term of magnetization transfer ratio (MTR) to denote the relative signal suppression due to saturation at a given saturation frequency. To quantify the MTR signal at a specific off‐resonance

120 Magnetization exchange effects and qT1 mapping in tumor patients at 7 T frequency, the data is first normalized to the unsaturated volume. Then, the data needs to be corrected for residual B0 offsets and the influence of direct water saturation. These two issues are addressed by fitting those points of the sampled Z‐spectra that are dominated by direct saturation to a Lorentzian line shape in a similar approach to (13). In this study we fitted the central five offsets and the outermost offsets, (‐4, ‐0.5, ‐0.3, ‐0, 0.3, 0.5, and 4 ppm) to the following equation:

w 2 LAwf,, 1 A [1] sat 22 wff4 0 sat 

Where A is the amplitude, w is the fitted line width, fsat is the saturation offset frequency with respect to bulk water, f0 is the difference in bulk water resonance frequency due to inhomogeneities in the B0 field.

The MTR at the frequencies of interest were calculated by

Sfsat sat  MTR fsat L fsat [2] S0

Where Ssat(fsat) /S0 is the measured signal intensity at saturation offset frequency fsat normalized to the unsaturated signal, and L(fsat) is the fitted Lorentz curve at that frequency.

The calculated MTR volumes were then registered unto the FLAIR volumes, using a normalized mutual information cost function and tri‐linear interpolation. In the following, we report the MTR intensities at the amide proton transfer frequency, MTR(+3.5ppm), the opposite side, MTR(‐3.5ppm), and the difference between the two, which we will label the asymmetry score at ±3.5 ppm.

T1 measurements

A multi‐slice T1 mapping sequence based on the sequences presented first by

Ordridge and later Clare (18,19) was implemented. In this T1 mapping

121 Chapter 5 approach, an inversion pulse is played out after which all slices in the volume are acquired successively using slice‐selective 90° excitations and EPI read‐outs. During the following repetitions, the slice ordering is shifted, so all slices are acquired at a varying inversion times. A slice‐shift of 2 slices was applied, so that in 23 repetitions, all 46 slices were sampled at 23 different time points after inversion, allowing for precise estimation of the longitudinal relaxation time.

A single‐shot EPI sequence forms the basis of the T1 mapping method. The acquired and reconstructed matrix of 224×224 covered an FOV of 224×224 mm. The 46 slices had a thickness of 1.5 mm with a 0.5 mm slice‐gap, covering 91.5 mm in the feet‐head direction. The phase encoding direction was set to be anterior‐posterior, with the fat‐shift direction towards the anterior. The sequence was accelerated using a SENSE acceleration factor of 3.6 in the phase encoding direction, a half‐scan factor of 0.609, and an EPI factor of 65. Fat suppression was accomplished by spectral inversion of the fat signal (SPIR) before each slice excitation. Repetition time and echo times were 10 s and 8.3 ms, respectively. Inversion was achieved by a non‐selective adiabatic inversion. Inversion times for odd slices ranged from 20ms to 5036 ms, and even slices from 134 to 5150 ms. The total scan duration for this sequence was 4 minutes and .10 seconds

The relaxation times were estimated by fitting the reordered and polarity restored volumes to the following expression:

tT//11 TRT  It I0 12 e  e  [3]

Where I(t) is the measured signal intensity at inversion time t, I0 is the fitted voxel intensity at the fully relaxed state, T1 is the longitudinal relaxation time, and TR is the repetition time of the sequence. The difference between the data and this expression was minimized using a Levenberg‐Marquardt optimization within the following parameter space: T1: [0 ‐ 10000] ms, I0: [0 – 10 × Ihighest], where Ihighest is the highest signal intensity of the sampled curve.

122 Magnetization exchange effects and qT1 mapping in tumor patients at 7 T

Finally, the fitted quantitative T1 (qT1) maps were registered to the FLAIR volumes by rigid body registration using a normalized mutual information cost function and applying tri‐linear interpolation.

ROI selection and statistics To select regions of pathological and normal appearing tissue, regions were drawn over multiple slices in the FLAIR image. Under guidance of an experienced radiologist, the following seven regions were selected; 1) enhancing tumor tissue (as based on the 3 T Gd enhanced T1‐weighted images), 2) non‐enhancing solid, 3) non‐enhancing cystic (both based on the 7 T FLAIR images), 4) edema, 5) normal appearing white matter (NAWM), 6) normal appearing grey matter (NAGM), and cortico‐spinal fluid (CSF).

To compare the effect of pathology, rather than the inter‐subject differences, voxels efrom th same ROIs over the different subjects were pooled before analysis. Because the number of voxels for each ROI is large, between 400 and 10000, comparing the means using a T‐Test or ANOVA gives highly significant differences (P < 1×10‐4). In this case, the magnitude of the differences (i.e. the effect size) might bring more insight. The normalized contrast‐to‐noise ratios were calculated using Hedges’ G measure of effect size. Hedges’ G is a standardized measure for the difference between two groups, while accounting for the variation found within each group. It is calculated by taking the difference between the means of two groups and normalizing that to the pooled variance in the two groups (20). The calculations were performed using the measures of effect size toolbox for Matlab (21). Additionally, correlation between saturation‐based parameters and qT1 values was assessed by Pearsons correlation coefficient and illustrated by scatter plots.

Results

Figure 1a illustrates the acquired MTR data and fit in a healthy volunteer for two voxels; one located in CSF and the other in white matter, as indicated by circles and squares, respectively. Shown are the signal intensities measured at different saturation frequencies (open markers) normalized to the unsaturated signal. The central five and outermost points are used to fit a Lorentzian line

123 Chapter 5 shape, which allows for residual B0 shift correction and removes the direct water saturation effect from the measured saturation ratio. Closed markers indicate the difference between the acquired data and the Lorentz fit. The large grey markers indicate the fitted intensities at ±3.5 ppm, which are used in the asymmetry calculation. Figure 1b‐g show the parameter maps that result from this fitting procedure. The fitted intensities at +3.5 ppm (fig. 1b) and ‐3.5 ppm (fig. 1c) both show some residual contrast between white and grey matter, probably due to MT effects that co‐occur in this measurement. To illustrate the variation in the Lorentz‐curve fit, we show the other parameter maps that result from this approach (fig. 1 e‐g).

124 Magnetization exchange effects and qT1 mapping in tumor patients at 7 T

1

0.8

0.6

Data Fit Diff ±3.5 ppm CSF 0.4 WM Normalized intensity

0.2

0

15 10 5 3.5 0 -3.5 -5 -10 -15 Saturation frequency (ppm) A)

MTR(+3.5 ppm) 0.3 MTR(-3.5 ppm) 0.3 MTR asymmetry 0.1

0.25 0.25

0.2 0.2 0

0.15 0.15

0.1 0.1 -0.1

0.05 0.05

B)0 C)0 D) -0.2 f0 (ppm) 0.2 A w (ppm) 3 1.2 0.15 2.5 1.1 0.1

0.05 1 2

0 0.9 1.5 -0.05 0.8 E)-0.1 F) G) 1 Figure 1: Example of acquired MTR curves and fitting and fitted parameter maps. A) Densely sampled Z‐spectra for a voxel located in CSF (circles) and white matter (squares). The fitted Lorentz peak using the central five and outermost two offsets is indicated by the (dashed) curves. The difference between this Lorentz curve and the measured intensities is indicated by the closed markers. The interpolated intensities at ±3.5 ppm is indicated by the grey marker. Panels B to G show the resulting parameter maps. B) and C) fitted intensities at ±3.5 ppm, D) asymmetry score. E) Fitted residual B0 offets, in ppm, F) fitted peak height A, G) fitted peak width w in ppm.

125 Chapter 5

Figure 2 shows an overview of all patients included in this study and shows a single slice through the main tumor mass from the following volumes:

1) contrast‐enhanced T1‐ weighted image, 2) fluid attenuated image, 3) qT1 map, 4) MTR(+3.5 ppm), 5) MTR(‐3.5 ppm), and 6) the MTR asymmetry. Based on the contrast in the CE‐T1w and the FLAIR images, ROIs were drawn on the FLAIR volumes, selecting healthy and diseased tissue for quantitative comparison.

grade CE-T1w FLAIR qT1 MTR(+3.5ppm) MTR(-3.5ppm) Asymmetry

0123400.10.20.3-0.2 -0.1 0 0.1

I

II

II

IV

IV

IV

Figure 2: Overview of six tumor cases, showing the T1 weighted images post contrast injection, FLAIR, quantitative T1 maps, MTR at +3.5 ppm and ‐3.5 ppm, and the asymmetry score.

126 Magnetization exchange effects and qT1 mapping in tumor patients at 7 T

For each subject, a B1 map was acquired, to probe the inhomogeneity of the RF excitation. This is a well known issue at higher field strengths, especially when only a single channel is used to achieve excitation of the spins (22,23). Figure 3 illustrates these inhomogeneities in approximately the same slices as figure 2. A standing wave effect that causes constructive interference in the middle of the brain, while destroying magnetization at the sides, can be observed in all subjects. Depending on the feet‐head location of the slice, this effect becomes more or less pronounced.

B1 Maps B1 Profiles 050050100 % 1 0.8 0.6 0.4 0.2

0 amplitude (%) + 1 1 0.8 0.6 Nominal B 0.4 0.2 A) B) 0

Figure 3: A) B1 maps of all subjects, at approximately the same slices as shown in figure 2. B) the average B1 profile (left to right) for the area between the dashed lines. Table 2 indicates the mean (± standard deviation) MTR values, asymmetry scores, and qT1 values for these patients. Figure 4 shows this information in an error‐bar plot to highlight the relative differences. The absolute difference in asymmetry score between NAWM and diseased tissue range from 3% for edema to 6% and 8% for non‐enhancing and enhancing tumor tissue, respectively. Cystic tissue MTR asymmetry also shows 8% difference to NAWM, while the MTR values in those regions are much reduced compared to the other tissue types.

127 Chapter 5

Table 2: Average values for different tissue types and standard deviations (SD). MTR while saturating at +3.5 ppm, ‐3.5 ppm , the asymmetry score for those frequencies, and the quantitative T1 values.

Tissue type MTR(+3.5) MTR(‐3.5) Asymmetry qT1 (ms) 0.128 0.146 ‐0.018 2152 Enhancing Tumor (±0.037) (±0.038) (±0.046) (±340) 0.107 0.15 ‐0.043 2185 Non‐enh. Tumor solid (±0.036) (±0.041) (±0.037) (±380) 0.062 0.078 ‐0.016 2046 Non‐enh. Tumor cystic (±0.028) (±0.039) (±0.041) (±449) 0.111 0.176 ‐0.065 1766 Edema (±0.033) (±0.04) (±0.03) (±366) 0.157 0.252 ‐0.095 1110 NAWM (±0.026) (±0.027) (±0.028) (±182) 0.132 0.194 ‐0.062 1665 NAGM (±0.039) (±0.05) (±0.044) (±366) 0.0599 0.057 0.003 4313 CSF (±0.038) (±0.039) (±0.039) (±441)

128 Magnetization exchange effects and qT1 mapping in tumor patients at 7 T

MTR(+3.5 ppm) MTR(-3.5 ppm) 0.3 0.3

0.25 0.25

0.2 0.2

0.15 0.15 MTR MTR 0.1 0.1

0.05 0.05

0 0 1 2 3 4 5 6 7 1 2 3 4 5 6 7 ROI ROI MTR Asymmetry qT 0.1 5000 1

0.05 4000 0 3000

-0.05 (ms) 1 T -0.1 2000 MTR difference MTR -0.15 1000 -0.2 1 2 3 4 5 6 7 1 2 3 4 5 6 7 ROI ROI Figure 4: Tissue specific averages and standard deviations over all subjects. ROIs: 1: enhancing tumor. 2: Solid non‐enhancing tumor. 3: Cystic non‐enhancing tumor. 4: Edema. 5: Normal appearing white matter. 6: Normal appearing grey matter. 7: CSF. Additionally, Table 3 compares mean values over different ROIs to that of normal appearing white matter, expressed in Hedge’s G scores. Of interest here is not to interpret the absolute values of this measure of contrast between the tissues for different scan types, but to compare the performance of the different sequences. Comparing the MTR values with the asymmetry score, we can conclude that the biggest effects are observed at MTR(‐3.5ppm), indicating that differences in NOE effects due to pathology are larger than the differences in

APT.When comparing different tissue types to NAWM, we observe that the qT1 sequence rates best in discriminating one tissue type from NAWM (5 out of 7 tissue types). Exceptions are: non‐enhancing cystic, which can be discriminated from NAWM better using the MTR(‐3.5ppm)scans.

The differences observed in the MTR based methods are largest for MTR(‐ 3.5 ppm): about 10 % (with respect to the unsaturated signal) when comparing NAWM with enhancing tumor and even 17 % for cystic regions compared to

129 Chapter 5

NAWM. The effects observed for the MTR asymmetry scores are smaller, with 4.5 % and 4.9 % for these regions, respectively. The effects of MTR(+3.5 ppm) display differences of 2.8 % and 9 %, respectively.

Table 3: Standardized differences of mean values of each tissue type compared to normal appearing white matter. Hedges’ G was calculated as a measure of effective contrast‐to‐noise.

Tissue Type MTR(+3.5) MTR(‐3.5) Asymmetry qT1 Enhancing Tumor ‐0.83 ‐3.02 1.85 3.38 Non‐enh. Tumor solid ‐1.64 ‐3.05 1.64 3.98 Non‐enh. Tumor cystic ‐3.56 ‐5.78 2.54 3.5 Edema ‐1.56 ‐2.27 1.04 2.38 NAWM 0 0 0 0 NAGM ‐0.72 ‐1.39 0.87 1.78 CSF ‐3.45 ‐6.74 3.35 11.26

When the average MTR based values are plotted against the observed qT1 values, strong correlations become obvious (see figure 5). MTR(+3.5 ppm) and

MTR(‐3.5 ppm) show a strong negative correlation with T1, with Pearsons correlation coefficients of ‐0.79 and ‐0.89 respectively (two‐tailed, p < 1×10‐6).

The asymmetry scores display a significant positive correlation with T1, a Pearsons correlation coefficient of 0.81 (two‐tailed, p < 1×10‐6). In this analysis, CSF was left out the collection of ROIs.

130 Magnetization exchange effects and qT1 mapping in tumor patients at 7 T

MTR(+3.5ppm) MTR(-3.5ppm) 0.3 0.3 Correlation: -0.79 Correlation: -0.89 0.25 0.25

0.2 0.2

0.15 0.15 MTR MTR

0.1 0.1

0.05 0.05

0 0 1000 2000 3000 1000 2000 3000 T (ms) T1 (ms) 1

CEST asymmetry 0.1 Correlation: 0.81

0.05 Tumor Enh. Tumor nonEnh. solid 0 Tumor nonEnh. cystic Edema -0.05 NAWM MTR NAGM -0.1

-0.15

-0.2 1000 2000 3000 T (ms) 1 Figure 5: Scatterplots of MTR values versus T1. Illustrating the correlation between MTR and T1 measures for the selected ROIs, excluding CSF.

131 Chapter 5

Discussion and Conclusions

In this paper the magnetization exchange effect measured by a novel pulsed steady state CEST method was investigated. The MTR values at +3.5 ppm, ‐ 3.5 ppm, and the resulting asymmetry score were reported. These values were compared to quantitative T1 mapping and relative intensities in FLAIR imaging. To do so, ROIs were delineated in six intra‐cranial tumor patients. The measured values in these ROIs and standardized effect sizes with respect to normal appearing white matter were compared. These quantitative comparisons show that the saturated images upfield from the water resonance (‐3.5 ppm) shows a stronger effect of pathology than the images where saturation was applied at the amide proton resonance. In addition, it was shown that MTR values have a strong negative correlation with the observed T1 values in the selected ROIs, while the asymmetry score has a positive correlation to T1.

Magnetization exchange effects in Tumors at 7 T The pulsed steady state method, using short interleaved saturation pulses rather than a long continuous wave presaturation pulse with a much larger effective pulse angel, allows rapid signal acquisition with sufficient saturation while remaining within SAR limits. Also, the smaller saturation pulses reduce the overall MT effect, highlighting the CEST andE NO effects (13). However, the MT effects can still be significant and should be taken into consideration when modeling these measurements. We have shown that under the experimental conditions in this study, strong effects are observed on both sides of the water resonance. The origin of these effects is expected to be conventional MT effects, as well as APT and NOEs from aliphatic protons. As a result, direct interpretation of an asymmetry score is greatly hampered and might obfuscate rather than elucidate the effects that take place in pathology. For example, previous studies have observed a baseline negative asymmetry score in healthy tissue in human subjects (24) and animal studies (7) indicating that NOEs might have played a role in those experiments as well.

132 Magnetization exchange effects and qT1 mapping in tumor patients at 7 T

In our study, the combined effect of conventional MT and NOEs observed in pathological tissue is larger on the MTR(‐3.5 ppm) side than the combined effect of MT and APT on the MTR(+3.5 ppm) side. For conventional MT is was already known that MTR correlates with viscosity, viable cell density and total protein concentration in lesions (25,26). A recent study at even higher field strength showed that in contrast to APT, NOEs are insensitive to pH and may provide a good biomarker for macromolecular content (27). Therefore, we propose that NOEs can be used in addition to conventional MT contrast, effectively increasing the sensitivity of the MTR measurement.

It would be of interest to further investigate the molecular basis of the observed NOE. The current hypothesis that NOEs are observed from aliphatic protons suggests that it these effects might reflect the cellular lipid‐bi‐layer enviroment. By combining high ‐resolution Z‐spectra with conventional magnetic resonance spectroscopy focussing on choline quantification, a direct relation to magnetization exchange effects and specific molecules that are modulated in pathology might be uncovered. In addition, these NOE effects might prove to be an additional tool for characterizing white matter pathology where the axonal or integrity is compromised.

To the best of our knowledge, this presents the first intra‐cranial tumor patient magnetization exchange data acquired at 7 T. Clinical tumor CEST has been shown before at lower field strengths (5,17,24) and qualitatively our MTR asymmetry results show the same trends; increasing asymmetry with increasing tumor malignancy. Differences in pulse sequences, saturation parameters and relaxation constants make quantitative comparison not feasible. MTR and CEST in healthy volunteers at 7 T have been explored before (28) and in that work they report an increase in asymmetry as well as SNR when comparing 3 T with .7 T data

Sensitivity to B0 inhomogeneities While shimming terms up to the third order were fitted to a separately acquired

B0 map to reduce B0 variations as much as possible, there are considerable shifts in locations where B0 changes too quickly to be completely compensated (i.e. nasal cavities and ear canals). To certain extent, variations in B0 are

133 Chapter 5 accommodated for by the Lorentz fit, by which the f0 shift is directly determined by the direct saturation peak (see fig. 1e). The residual B0 offsets are reflected in the f0 map and closely mimic the information contained in a regular

B0 map. The span of the central saturation frequencies used in the Lorentz fit limits the maximum frequency range of this map, to roughly 75 % of this span. In the case for this healthy volunteer, the saturation frequencies were spaced 0.5 ppm apart, thus the middle five frequencies covered 2 ppm, allowing a shift of about 1.5 ppm to be fitted reliably. Offsets that are beyond this range, as are present directly above the nasal‐cavities and the ear canals cannot be fitted using the Lorentz curve approach and this directly influences our ability to fit the MTR at these anatomical locations as well. In the patient scans, the central offsets used in the Lorentz fit were spaced ‐0.5, ‐0.25, 0, 0.25, 0.5 ppm, allowing for a maximum B0 shift of about 0.75 ppm (223.5 Hz). A simple solution for this limitation is of course to include a wider range off offsets to fit the DS effect, but one has to consider the asymmetries that might arise from exchanging protons that resonate closer to the water resonance frequency, such as exchanging –OH groups (1 ‐ 2 ppm) (29).Determination of the f0 offset is only part of the solution to residual B0 inhomogeneities after shimming. As can be observed in fig. 1b, the same pattern as in fig. 1e is present in the fitted MTR values. The main reason for this is that however many points are acquired in this experiment, the step‐size is still quite large: 0.5 ppm. The peak‐widths due to APT or NOE magnetization transfer are in the same order of magnitude and only one or two points lie on this peak. After the f0 shift is determined the correct MTR intensity is found by spline interpolation with the neighboring points. If those points are too far from the actual resonance frequency, i.e. the frequency shift is large, the true MTR at the frequency of interest is increasingly underestimated. For the patient data, the saturation points are spaced 0.34 ppm apart and thus the bias is expected to be smaller than in the data shown for the healthy subject.

Additionally, while the Lorentz‐fit method to correct for B0 inhomogeneities is superficially similar to the one proposed by Zaiß (30), there are considerable differences. First of all, this study did not fully model the underlying principles to different pools of magnetization transfer. Not only these contributions unknown, but also the sparse sampling of the Z‐spectrum in only 17 points

134 Magnetization exchange effects and qT1 mapping in tumor patients at 7 T makes the dataset too small to return reliable results. Secondly, as was indicated in (30), the system is too much dependent on the effective B1 power to be suitable for the variations that we encounter in our patient data. Therefore, the Lorentz fit approach we take in this paper is an attempt to subtract most of the DS effect and possibly also correct for some portion of the conventional MT, but the resulting MTR values should not be interpreted as full modeling of the underlying MT processes.

Sensitivity to B1 inhomogeneities

There are considerable variations in the B1 excitation field and it is important to address the effects on the measured values for the MTR‐based methods used in this paper. Specifically, the relative contribution of the different magnetization transfer effects might change when the effective B1 power changes. We will discriminate conventional effects from DS, conventional MT, NOE and APT.

For DS, the effect of variations in B1 amplitude is linear, i.e. the DS can be considered to be an on‐resonance pulse with an amplitude given by the center frequency and effective bandwidth of the pulse. Conventional magnetization transfer effects also strongly depend on B1, as transfer of magnetization from the bound water pool to free bulk water is practically instantaneous and the amount of saturation scales directly with the B1 power (31). The parameter maps resulting from the Lorentz peak fit, such as the fitted peak amplitude A (fig. 1f) and w (fig. 1g), illustrate the combined effects of these two mechanisms. The fitted peak amplitude A reflects the total amount of saturation at 0 ppm compared to the two rpoints fa off‐resonance, which are assumed to be affected only by conventional MT effects. Similarly, the fitted peak width w reflects the amount conventional MT effect present in these measurements, as increased MT effects directly causes an increase in the measured Z‐spectrum.

With our experimental parameters, the contribution of conventional MT effects is considerable. This makes these data less suitable for quantitative modeling of the CEST effect as a separate contribution. To do so, the effect of conventional MT should be minimized by using a shorter and smaller saturation pulse, as was applied in (13). This approach however, assumes relatively homogeneous

B1 fields, which are not always available at 7 T.

135 Chapter 5

Similarly to MT effects, NOEs area also expected to depend on B1, although the mechanism of magnetization transfer is different. As described in (2), it is expected that much of the NOEs are in fact exchange mediated, i.e. saturation of protons at aliphatic proteins is first transferred to secondary exchanging groups by means of spin diffusion, before contributing to the bulk water saturation. Because the transfer rates of these secondary groups are much higher than amide protons (500‐10000 Hz) and the amount of aliphatic protons is large, we can assume that the limiting factor in saturation is the amount of saturation power.

Previous studies have shown that the observed proton transfer rate (PTR) can be expressed as (2):

tsat T1w PTR xssww k T1 1  e [4] 

Where the maximum observed proton transfer rate is given by the fraction of amide protons xs, the labeling efficiency α, exchange rate ksw, the effective saturation time tsat, and longitudinal relaxation time of the bulk water pool T1w. When neglecting transverse relaxation during the saturation pulses, the labeling efficiency is approximated by (2,32):

2  B1   22 [5]  Bk1  sw

Which incorporates the gyromagnetic ratio γ in rad/Ts, the B1 power and exchange rates ksw. For amide proton exchange, ksw is expected to be around

30 Hz and the effective B1 field we obtained in our experiments lies between 40‐

100 % of the nominal B1 power of 1.8μT (See Figure 3). Therefore, the labeling efficiency lies between 96‐99 % for our experiments. When taking into account the errors due to B0 inhomogeneities and SNR, and the fact that APT is only a 3‐ 5 % with respect to the unsaturated volume, it is not expected that the sensitivity of these measurements is sufficient to pick up these variation resulting from B1 inhomogeneities.

136 Magnetization exchange effects and qT1 mapping in tumor patients at 7 T

qT1 and FLAIR We compare the MTR based measured to two other MRI sequence types; FLAIR and quantitative T1 mapping. The FLAIR sequence is essentially a T2 weighted sequence in which signal from CSF is minimized by inversion. FLAIR is often used in clinical settings, because of the excellent delineation of lesions and edema. At 7 T, conventional FLAIR contrast is more difficult to accomplish due to the shifted T1 of tissue, but this has been ameliorated by means of an magnetization preparation (14). As a result however, the sequence consists of multiple RF pulses and the realized B1 power influences the obtained signal intensity non‐linearly. Inhomogeneities in the B1 field cause shading effects in these FLAIR images, making them less suited for quantitative analysis than for qualitative comparison. Nonetheless, tissue containing tumor or edema are easily recognized on FLAIR images, especially cystic regions which show strongly decreased T2 weighting. In these measurements, increases in FLAIR image intensity correlate well with increases in water content or, inversely, with cellular breakdown.

The calculated T1 maps show no B1 inhomogeneities in the fitted T1 values, as B1 affects the voxels intensities in a constant manner over the different inversion times and thus show up as a variation in the fitted voxel intensity I0 and not in

T1. On the T1 maps, tumors are recognizable as regions with elongated T1 values. It is of interest to note that the regions labelled as non‐enhancing cystic, do not show especially large T1 values. This indicates that while water content is high (as indicated by FLAIR contrast), it displays roughly the same relaxation behavior as other tissue. This might be interpreted as ‘dirty’ fluid filled cavities with increased viscosity or breakdown products containing relaxation sites, reducing the effective T1. This might also explain the large variation observed in CSF ROIs, as not all CSF can be considered ‘clean’ but might also be contaminated with breakdown products.

Correlation to T1

Direct water saturation and NOE is expected to depend directly on T1 (10), while for APT, only small effects are expected (13). The conventional MT effect is not expected to depend on T1, although it might correlate with T1, as both T1

137 Chapter 5 and MT are measures that are sensitive to tissue water content (33). When plotting the average MTR based values versus the longitudinal relaxation times for selected ROIs, a strong correlation between the two was observed. In the measurements where conventional MT plays a significant role, the correlation is negative. This follows from the fact that MT and T1 are sensitive indicators of water content. This was shown nearly twenty years ago (34), although it was indicated that T1 and T2 depend much more on water content than MT does. In the asymmetry score we find a positive correlation, which agrees with the dependency expressed in equation 3. An additional effect is that of increasing saturation due to elongated T1 values. Once saturation has been achieved, the bulk water pool takes longer to return to equilibrium. Thus, it is easier to accumulate saturation when T1 is longer. Recently, is was shown in animal models that the observed APT effect scales with T1 and should be taken into account when applying these measurements to assess changes in pH (35).

It is important to note that in our experiments, we have not measured the dependency of MTR‐based measures on T1, but only the correlation of these two measures. The measures displayed in figure 5 represent distinctly different types of tissue and there are many possible sources of magnetization relaxation or transfer, that are not necessarily the same for all tissue types. For instance, one might be tempted to assign water content as the underlying principle governing both MTR values and qT1. The effects observed in non‐enhancing cystic tumor tissue, which is expected to contain necrotic fluid, do not fully follow this scheme however, there is little difference in qT1 with respect to enhancing tumor, but the MTR values are much reduced. This indicates that structural integrity is an important aspect of MT as well.

To conclude, this study shows the application of a novel CEST method and quantitative T1 mapping in a small group of intra‐cranial tumor patients at 7 T. Using a pulsed steady state method that achieves saturation while remaining within SAR limits, the first intra‐cranial tumor magnetization exchange data at 7 T was presented. We have observed that these methods reflect tumor heterogeneity in varying degrees. In terms of discriminative power, the MTR(‐3.5 ppm) image performs best, because conventional MT as well as NOEs add to the magnetization exchange effect. Also, strong correlations between

138 Magnetization exchange effects and qT1 mapping in tumor patients at 7 T

MTR‐based values and the corresponding T1 relaxation times were observed. As the the applied experimental settings will strongly influence the relative contribution of these magnetization exchange effects, these measurements cannot be interpreted as representative for all types of CEST experiments. However, these confounding factors are expected to be present in some degree and need to be otaken int account when interpreting CEST results in pathology. Further research in clinical neuroscience and MR methodology is needed to fully elucidate the molecular basis of the observed MRI contrast.

139 Chapter 5

References

1. Ward K., Aletras A., Balaban R. A New Class of Contrast Agents for MRI Based on Proton Chemical Exchange Dependent Saturation Transfer (CEST). Journal of Magnetic Resonance 2000;143:79–87. 2. van Zijl PCM, Yadav NN. Chemical Exchange Saturation Transfer (CEST): What Is in a Name and What Isn’t? Magn Reson Med 2011;65:927–948. 3. Grad J, Bryant RG. Nuclear Magnetic Cross‐relaxation Spectroscopy. Journal of Magnetic Resonance (1969) 1990;90:1–8. 4. Zhou J, Lal B, Wilson DA, Laterra J, Zijl PCM van. Amide Proton Transfer (APT) Contrast for Imaging of Brain Tumors. Magn Reson Med 2003;50:1120–1126. 5. Wen Z, Hu S, Huang F, Wang X, Guo L, Quan X, Wang S, Zhou J. MR Imaging of High‐grade Brain Tumors Using Endogenous Protein and Peptide‐based Contrast. NeuroImage 2010;51:616–622. 6. Sun PZ, Zhou J, Huang J, Zijl P van. Simplified Quantitative Description of Amide Proton Transfer (APT) Imaging During Acute Ischemia. Magn Reson Med 2007;57:405–410. 7. Zhou J, Tryggestad E, Wen Z, Lal B, Zhou T, Grossman R, Wang S, Yan K, Fu D‐X, Ford E, Tyler B, Blakeley J, Laterra J, van Zijl PCM. Differentiation Between Glioma and Radiation Necrosis Using Molecular Magnetic Resonance Imaging of Endogenous Proteins and Peptides. Nat Med 2011;17:130–134. 8. Howe FA, Barton SJ, Cudlip SA, Stubbs M, Saunders DE, Murphy M, Wilkins P, Opstad KS, Doyle VL, McLean MA, Bell BA, Griffiths JR. Metabolic Profiles of Human Brain Tumors Using Quantitative in Vivo 1H Magnetic Resonance Spectroscopy. Magnetic Resonance in Medicine 2003;49:223–232. 9. Clarke J, Itzhaki LS. Hydrogen Exchange and Protein Folding. Current Opinion in Structural Biology 1998;8:112–118. 10. Ling W, Regatte RR, Navon G, Jerschow A. Assessment of Glycosaminoglycan Concentration in Vivo by Chemical Exchange‐Dependent Saturation Transfer (gagCEST). PNAS 2008;105:2266–2270. 11. Tkáč I, Öz G, Adriany G, Uğurbil K, Gruetter R. In Vivo 1H NMR Spectroscopy of the Human Brain at High Magnetic Fields: Metabolite Quantification at 4T Vs. 7T. Magnetic Resonance in Medicine 2009;62:868–879. 12. Rooney WD, Johnson G, Li X, Cohen ER, Kim S‐G, Ugurbil K, Springer CS. Magnetic Field and Tissue Dependencies of Human Brain Longitudinal 1H2O Relaxation in Vivo. Magn Reson Med 2007;57:308–318. 13. Jones CK, Polders D, Hua J, Zhu H, Hoogduin HJ, Zhou J, Luijten P, van Zijl PCM. In Vivo Three‐dimensional Whole‐brain Pulsed Steady‐state Chemical Exchange Saturation Transfer at 7 T. Magn Reson Med 2012;67:1579–1589. 14. Visser F, Zwanenburg JJM, Hoogduin JM, Luijten PR. High‐resolution Magnetization‐prepared 3D‐FLAIR Imaging at 7.0 Tesla. Magnetic Resonance in Medicine 2010;64:194–202. 15. Sacolick LI, Wiesinger F, Hancu I, Vogel MW. B1 Mapping by Bloch‐Siegert Shift. Magnetic Resonance in Medicine 2010;63:1315–1322.

140 Magnetization exchange effects and qT1 mapping in tumor patients at 7 T

16. Sun PZ, van Zijl PCM, Zhou J. Optimization of the Irradiation Power in Chemical Exchange Dependent Saturation Transfer Experiments. J Magn Res 2005;175:193– 200. 17. Zhou J, Blakeley JO, Hua J, Kim M, Laterra J, Pomper MG, Zijl PCM van. Practical Data Acquisition Method for Human Brain Tumor Amide Proton Transfer (APT) Imaging. Magn Reson Med 2008;60:842–849. 18. Ordidge RJ, Gibbs P, Chapman B, Stehling MK, Mansfield P. High‐speed Multislice T1 Mapping Using Inversion‐recovery Echo‐planar Imaging. Magn Reson Med 1990;16:238–245. 19. Clare S, Jezzard P. Rapid T1 Mapping Using Multislice Echo Planar Imaging. Magn Reson Med 2001;45:630–634. 20. Hedges LV. Distribution Theory for Glass’s Estimator of Effect Size and Related Estimators. J Educ Behav Stat 1981;6:107–128. 21. Hentschke H, Stüttgen MC. Computation of Measures of Effect Size for Neuroscience Data Sets. European Journal of Neuroscience 2011;34:1887–1894. 22. Vaughan JT, Garwood M, Collins CM, Liu W, DelaBarre L, Adriany G, Andersen P, Merkle H, Goebel R, Smith MB, Ugurbil K. 7T Vs. 4T: RF Power, Homogeneity, and Signal‐to‐noise Comparison in Head Images. Magn Reson Med 2001;46:24–30. 23. Schick F. Whole‐body MRI at High Field: Technical Limits and Clinical Potential. European Radiology 2005;15:946–959. 24. Jones CK, Schlosser MJ, Zijl V, C.m P, Pomper MG, Golay X, Zhou J. Amide Proton Transfer Imaging of Human Brain Tumors at 3T. Magn Reson Med 2006;56:585–592. 25. Lundbom N. Determination of Magnetization Transfer Contrast in Tissue: An MR Imaging Study of Brain Tumors. AJR 1992;159:1279–1285. 26. Mishra AM, Reddy SJ, Husain M, Behari S, Husain N, Prasad KN, Kumar S, Gupta RK. Comparison of the Magnetization Transfer Ratio and Fluid‐attenuated Inversion Recovery Imaging Signal Intensity in Differentiation of Various Cystic Intracranial Mass Lesions and Its Correlation with Biological Parameters. Journal of Magnetic Resonance Imaging 2006;24:52–56. 27. Jin T, Wang P, Zong X, Kim S‐G. MR Imaging of the Amide‐proton Transfer Effect and the pH‐insensitive Nuclear Overhauser Effect at 9.4 T. Magnetic Resonance in Medicine 2012:n/a–n/a. 28. Mougin OE, Coxon RC, Pitiot A, Gowland PA. Magnetization Transfer Phenomenon in the Human Brain at 7 T. NeuroImage 2010;49:272–281. 29. Zhou J, Zijl PCM van. Chemical Exchange Saturation Transfer Imaging and Spectroscopy. Prog Nucl Mag Res Spec 2006;48:109–136. 30. Zaiß M, Schmitt B, Bachert P. Quantitative Separation of CEST Effect from Magnetization Transfer and Spillover Effects by Lorentzian‐line‐fit Analysis of Z‐ spectra. Journal of Magnetic Resonance 2011;211:149–155. 31. Henkelman RM, Huang X, Xiang QS, Stanisz GJ, Swanson SD, Bronskill MJ. Quantitative Interpretation of Magnetization Transfer. Magn Reson Med 1993;29:759–766. 32. Zhou J, Wilson DA, Sun PZ, Klaus JA, van Zijl PCM. Quantitative Description of Proton Exchange Processes Between Water and Endogenous and Exogenous Agents for WEX, CEST, and APT Experiments. Magn Reson Med 2004;51:945–952.

141 Chapter 5

33. Henkelman RM, Stanisz GJ, Graham SJ. Magnetization Transfer in MRI: a Review. NMR in Biomedicine 2001;14:57–64. 34. Scholz TD, Ceckler TL, Balaban RS. Magnetization Transfer Characterization of Hypertensive Cardiomyopathy: Significance of Tissue Water Content. Magnetic Resonance in Medicine 1993;29:352–357. 35. Sun PZ, Wang E, Cheung JS. Imaging Acute Ischemic Tissue Acidosis with pH‐ sensitive Endogenous Amide Proton Transfer (APT) MRI—Correction of Tissue Relaxation and Concomitant RF Irradiation Effects Toward Mapping Quantitative Cerebral Tissue pH. NeuroImage 2012;60:1–6.

142 Magnetization exchange effects and qT1 mapping in tumor patients at 7 T

143

Chapter 6. Multimodal tract-based analysis in ALS patients at 7 T

Daniel Polders *1, Esther Verstraete, MD *2, René CW Mandl, PhD 3, Martijn P van den Heuvel, PhD 3, Jan H Veldink, MD, PhD 2, Peter Luijten, PhD 1, Leonard H van den Berg, MD, PhD 2, Johannes Hoogduin, PhD 1,2

* authors contributed equally

1 Department of Radiology, University Medical Centre Utrecht, Utrecht, The Netherlands 2 Department of Neurology, University Medical Centre Utrecht, Rudolf Magnus Institute of Neuroscience, Utrecht, The Netherlands 3 Department of Psychiatry, University Medical Centre Utrecht, Rudolf Magnus Institute of Neuroscience, Utrecht, The Netherlands This chapter is in preparation for submission to Amyotrophic Lateral Sclerosis

I must confess that, at that time, I had absolutely no knowledge of the slowness of the relaxation processes in the ground state, processes Alfred Kastler physicist, 1902 – 1984, Nobel Prize in Physics in 1966 "for the discovery of Hertzian resonances in atoms".

145 Chapter 6

Abstract

Objective: To explore the value of additional MR contrasts in elucidating the decrease in fractional anisotropy (FA) as has been observed in the corticospinal tracts (CST) of patients with amyotrophic lateral sclerosis (ALS).

Methods: Eleven patients and nine healthy control subjects were scanned at 3 T and 7 T MRI. Whole brain and tract specific comparison was performed of both diffusion weighted (3 T), quantitative T1 (qT1), magnetization transfer (MT) and amide proton transfer weighted (APTw) imaging (7 T).

Results: Whole‐brain comparison using histogram analyses showed no significant differences between patients and controls. Measures along the CST showed a significantly reduced FA together with a significantly increased diffusivity perpendicular to the tract direction in patients as compared to controls. In addition, patients showed a small but significant increase in MTR values within the right CST. No significant changes were observed in qT1 and APTw values.

Conclusions: Our findings, based on a multimodal approach, revealed that the decrease in FA is most likely caused by an increased diffusivity perpendicular to the CST. This finding, together with the increase in MTR might indicate that an increase of free liquid spins rather than demyelination is the primary cause of observed FA decrease in the CST in patients with ALS.

146 Multimodal tract‐based analysis in ALS patients at 7 T

Introduction

The upper motor neurons in the brain and the lower motor neurons in the brainstem or spinal cord are connected through the corticospinal tract (CST), the great white matter ‘highway’ of the motor system. This tract has consistently been found to show degenerative effects in patients with amyotrophic lateral sclerosis (ALS), both in post‐mortem and in imaging studies (1–3). Most notably, diffusion tensor imaging (DTI) measures have demonstrated reduced fractional anisotropy (FA) in the CST (4–15).

The central assumption is that FA reflects white matter integrity, as intact axonal and myelin boundaries will restrict diffusion perpendicular to the white matter fibers and thus increase FA, while loss of white matter integrity will reduce diffusion restriction and thus lower FA. However, besides white matter integrity, DTI measures are influenced by multiple factors including: crossing fibers, fiber re‐organization, increased membrane permeability, destruction of intracellular compartments, and glial alterations (16,17). Which of these factors are the cause of reduced FA in the CST in ALS, might be elucidated by application of additional magnetic resonance imagine (MRI) contrasts.

Other MRI contrasts which might increase our insight into the degenerative changes within the CST in patients with ALS include quantitative T1 mapping

(qT1 mapping), magnetization transfer contrast (MTC; expressed using the magnetization transfer ratio, MTR) and chemical exchange saturation transfer

(CEST). It has been shown that the image intensities found in qT1 mapping can be used as a marker for the degree of myelination in disease and during brain development (18–20). MTR reflects the exchange between water bound to macromolecules and the unbound water fraction. Correlations of MTR with neuronal integrity or the degree of myelination have been shown (21). For example, patients with multiple sclerosis have been reported to show a global decrease in MTR (22). CEST is a relatively new imaging modality; a particular type of magnetization transfer experiment focused on measuring exchange of protons between specific solutes and free water. CEST can be used to study endogenous exchanging protons, such as amide protons resonating at +3.5 ppm from water, resulting in amide proton transfer weighted imaging, APTw

147 Chapter 6 imaging. Amide protons are abundantly present in protein backbones and peptides, and APTw imaging is expected to reflect physiological changes in this specific proton pool in case of disease. It has been suggested that APTw imaging may be sensitive for myelination (23) and it would be of interest to see if APT measurements can contribute to the further understanding of ALS pathology.

In this study we investigated whether measures of qT1, MTR and APTw imaging, applied at ultra high field (7 tesla), might help to disentangle which factors contribute to the FA reduction in the CST of patients with ALS and provide more detailed in‐vivo tissue characterization.

Methods

Subjects Eleven patients with ALS (mean age 55.5 years; range 35‐71; 8 males and 3 females) and nine age and gender matched healthy controls (mean age 54.2 years; range 36‐67; 7 males and 2 females) were included in this study. Patients, diagnosed according to the El Escorial criteria, were recruited from the ALS outpatient clinic of the University Medical Center Utrecht, excluding subjects with a history of brain injury, epilepsy, psychiatric illness and other neurodegenerative diseases. Demographic and clinical characteristics are provided in Table 1, including the functional impairment as measured using the ALSFRS‐R and the subscores for bulbar (B score, items 1, 2 and 3), upper limb (UL score, items 4 and 5) and lower limb (LL score, items 8 and 9) functioning. The local Medical Ethics Committee for Research in Humans approved this study and written informed consent was obtained from all subjects, in concordance with the Declaration of Helsinki.

MRI Hardware DTI data was acquired on a 3 T whole body MR scanner (Philips Medical Systems, Best, the Netherlands). The body coil and 8 channel head coil (Nova Medical Inc., Burlington, MA, USA) were used for signal transmission and reception, respectively.

148 Multimodal tract‐based analysis in ALS patients at 7 T

Quantitative T1, MT, and APTw imaging measurements were performed at a 7 T whole body MR scanner (Philips Medical Systems, Cleveland, USA). A quadrature birdcage transmit head coil (Nova Medical Inc., Burlington, MA, USA) was used in combination with a receive‐only coil (Nova Medical Inc.,

Burlington, MA, USA). A B0 mapping sequence was included, covering the brain volume. This was then used to optimize the shim settings up to the third order.

Table 1 Demographic and clinical characteristics of the patients age site of TTD DD ALSFR B UL LL gender (yr) onset (mnths) (mnths) EE S‐R PR score score score 1 F 65 LL 11 15 prob 39 0.6 11 6 3 2 M 51 LL 22 33 prob 41 0.2 12 8 3 3 M 47 UL 2 16 poss 40 0.5 11 3 7 4 M 35 UL 9 45 prob 31 0.4 11 3 2 5 M 39 UL 2 42 poss 38 0.2 12 3 6 6 M 58 LL 11 25 prob‐LS 44 0.2 12 8 5 7 M 63 LL 12 25 prob 42 0.2 12 7 3 8 F 61 UL 12 18 def 38 0.6 9 5 7 9 F 60 UL 11 21 prob 46 0.1 12 6 8 10 M 61 UL 10 15 prob 40 0.5 12 6 5 11 M 71 UL 3 17 prob‐LS 31 1.0 12 2 4 55.5 9.5 39.1 0.4 24.7 (10.8) (11.2) (5.8) (4.7) (0.3) Abbreviations:TTD = time to diagnosis; DD =disease duration F = female; M = male; B = bulbar; UL = upper limbs; LL = lower limbs; EE = El Escorial category; poss = possible; prob‐ LS = probable laboratory supported; prob = probable; def = definite; ALSFRS‐R = revised ALS functional rating scale; PR = progression rate (48 – ALSFRS‐R / disease duration). The bottom row of the table provides mean values and standard deviations (between brackets).

DTI and fiber tracking DTI measurements were acquired using methods described earlier (10,24). Calculation of the DTI parameters was performed using the ExploreDTI toolbox (25). After brain extraction, DTI volumes were corrected for residual eddy current and motion artifacts by aligning the diffusion weighted images to the b = 0 s/m² image using 3D affine registration and reorientation of the B‐matrix (26). Tensor values were estimated using the RESTORE algorithm (27). Finally,

149 Chapter 6

DTI‐based parameter volumes were transformed to MNI‐space. We further investigate the effects in fractional anisotropy (FA) by looking at the changes in longitudinal diffusivity (Dlong) and transversal diffusivity (Dtrans).

Fiber streamlines were selected by placing seed regions of interest (ROIs) in both the left or right motor tract at the level of the pons. Fiber tracking was then performed by deterministic streamline fiber tractography (28) from 125 seed points distributed homogeneously over each voxel in the seed ROI. The following settings for streamline reconstruction were applied: minimal FA = 0.2, maximal angle = 20°, streamline step size = 1mm, minimal/maximal fiber length = 40/500 mm. Additional selection criteria were placed as an inclusion‐ ROI at the level of the primary motor cortex to select those streamlines that connected both the pons level and cortical level ROIs. Two exclusion‐ROIs were placed to prevent inter‐hemispheric crossover and connections to the cerebellum, see figure 1.

Figure 1: Location of DTI tract selection regions of interest and average fiber bundles. The motor tracts of the CST were identified on a directionally colored FA map at the level of the pons. Panels: a) blue indicates the seed region, placed at an axial slice at the level of the pons, from which fiber tracking was initiated, b) an AND region (green) is placed directly on the motor cortex, two NOT regions (red) prevent inclusion of fibers that c) connect to the cerebellum and d) the other hemisphere. e) the resulting white matter fiber tract selected with these ROIs. f) the subject‐average fiber curves (blue) and group‐average curves based on these (left and right, red).

150 Multimodal tract‐based analysis in ALS patients at 7 T

T1 mapping

A multi‐slice T1 mapping sequence based on the sequences presented by

Ordridge and later Clare (29,30) was implemented. In this T1 mapping approach, an inversion pulse is played out after which all slices in the volume are acquired successively using slice‐selective 90° excitations and EPI read‐outs. During the following repetitions, the slice ordering is shifted, so all slices are acquired at different inversion times. A slice‐shift of 2 sslices wa applied, so that in 23 repetitions, all 46 slices were sampled at 23 different time points after inversion, allowing for robust and precise estimation of the longitudinal relaxation time.

A single‐shot EPI sequence forms the basis of the T1 mapping method. The acquired and reconstructed matrix of 224×224 covered a field of view (FOV) of 224×224 mm. The 46 slices had a thickness of 1.5 mm with a 0.5 mm slice‐gap, covering 91.5 mm. The phase encoding direction was set to be anterior‐ posterior, with the fat‐shift direction towards the anterior. The sequence was accelerated using a SENSE acceleration factor of 3.6 in the phase encoding direction, a half‐scan factor of 0.609, resulting in an EPI factor of 65. Fat suppression was accomplished by spectral inversion of the fat signal (SPIR) before each slice excitation. TR and TE were 10 s and 8.3 ms, respectively. Inversion was achieved by a non‐selective adiabatic inversion. Inversion times for odd slices ranged from 20 ms to 5036 ms, and even slices from 134 to 5150 ms. The total scan duration for this sequence was 4 minutes and 10 seconds.

Relaxation times were estimated by fitting the reordered and polarity restored volumes to an expression that models the inversion curve as a function of the longitudinal relaxation time T1 and signal intensity at full equilibrium I0, see supplementary material for more information.

MT and APTw imaging MT and APTw imaging were conducted using the pulsed steady state method as previously described by Jones et al. (31). One unsaturated and 17 saturated volumes with different saturation frequencies were acquired using a segmented 3‐D EPI readout scheme. Each volume was acquired in ~190 repetitions.

151 Chapter 6

Saturation was accomplished by short off resonance saturation pulses before each excitation pulse, building up a steady state while acquiring the outer regions of k‐space. The saturation pulse had a duration of 50 ms at 1.8μT nominal B1 amplitude. Excitation was achieved using an 18.5° binomial pulse for water only excitation. The scan was accelerated with SENSE = 2.5× in the phase encoding direction (AP), leading to an EPI factor of 15. The acquired resolution (2.00×2.13×2.00 mm) was reconstructed to 2 mm3 isotropic and covered an FOV of 224×224×100 mm3 (LR×AP×FH). The Z‐spectrum was sampled at the following off‐resonance frequencies: ‐1200, ‐1100, ‐1000, ‐900, ‐ 800, ‐150, ‐75, 0, 75, 150, 800, 900, 1000, 1100, 1200 Hz relative to the water resonance. After correction for direct water saturation and residual B0 variations, the MTR was determined at a saturation offset of +4 ppm (~1200 Hz), and APTw images were calculated by determining the MTR asymmetry at ±3.5 ppm (±1043 Hz), see supplementary materials for details.

Quantitative comparison

For quantitative comparison, qT1, MTR, and APTw volumes were registered to each subjects’ FA map by rigid body registration using a normalized mutual information cost function and tri‐linear interpolation. For the MT based volumes, the 4 ppm saturated image was used in the registration procedure and the resulting transformation matrix was used forl al following parameter maps.

Averaged normalized histograms were compared to evaluate the MR contrasts between patients and controls on a whole brain level. The histograms were calculated over all voxels in the brain volume, excluding those that were obviously affected by signal voids due to air‐tissue boundaries, such as near the nasal cavities and ear‐canals, as observed in EPI based methods. Included in these histograms were grey matter and the ventricles, so a contribution of cerebro‐spinal fluid (CSF) is also expected. After normalizing to their total count, the average histograms for patients and controls were calculated and plotted with the accompanying standard deviations.

The goal of the tract‐based analysis was to be able to compare quantitative MRI measures along the specified tract, while allowing for inter‐subject variation of each tract 3D trajectory. Following the method described in (32), each subject’s

152 Multimodal tract‐based analysis in ALS patients at 7 T reconstructed CST (consisting of a collection of ~150 streamline trajectories) was averaged to a single curve. The values of the different MRI contrasts of interest of the voxels crossed by the fiber trajectories were mapped to the nearest corresponding point on the curve. As a result, each subject’s curve reflects the averaged values along that curve. Similarly, the group average curves were calculated by taking calculating the average curve over all subjects. The resulting two averaged curves correspond to the group‐average left and right corticospinal tracts. The compared left and right tracts consisted of 22 curves from patients and 18 from healthy controls, spanning a distance of about 8 cm, in 2 mm steps. We report the median values (± inter‐quartile ranges) of measures of FA, Dlong, Dtrans, qT1, MT, and APTw along this combined curve.

Statistical analysis The entire brain volumes were compared between patients and controls by testing for difference between the mean values for each subject, using a two‐ tailed t‐test. The multiple measures, for each modality, along the left and right CST sections were compared between patients and controls by using a linear mixed ‐effects model, including group assignment (patient or healthy control) as a factor. Age and gender were included as covariates in case this improved the maximum likelihood estimation. Additionally, for each separate point along the group averaged tract, the distributions of patients and controls were compared by testing for difference of the medians using a Mann–Whitney–Wilcoxon or Rank test. The effect size was calculated by calculating the difference between the medians. For all analyses, a statistical threshold of p ≤ 0.05 was considered statistically significant.

Results

Whole brain histogram analysis At the whole brain level, based on histogram analysis, none of the MR contrasts revealed any significant differences between patients and controls. Figure 2 shows the whole brain histograms for each MR contrast, with an example slice of a single subject. A difference in the MT peak locations for patients (0.19) and

153 Chapter 6 controls (0.15) can be observed. However, the whole brain average MT values did not show significant differences between patients and controls.

Corticospinal tract measures We report the values of several MR contrasts measured along the fiber tractography streamlines of the corticospinal tract. Because the coverage of the DTI volume is larger than that of the other contrasts, we focus here on the region of the CST where all acquired volumes overlapped. In the presented results, we show the median values over the left and right group averaged corticospinal tracts each consisting of 41 points, spanning roughly 8 cm in the feet‐head direction.

The tract specific measures along the CST revealed, a significantly reduced FA in patients compared to controls along the right CST (p = 0.01) and a trend‐like reduced FA along the left CST (p = 0.09).Figure 3a shows the FA measures for patients and controls along the tract, significantly reduced FA was found both in the subcortical region (right) and more caudally (posterior limb of the internal capsule, left and right). Diffusivity transverse to the largest direction of diffusion, Dtrans shows a significant increase along the right CST (p = 0.04). The measures over the left CST are not significantly different in patients compared to controls (p = 0.19), however, two consecutive points along the CST were significantly higher in patients compared to controls (Figure 3b). Longitudinal diffusivity, Dlong, measures are not significantly different along the CST (left p = 0.35; right p = 0.46), indicating that the change in FA originates from a relative increase in Dtrans in the patient group (Figure 3c).

154 Multimodal tract‐based analysis in ALS patients at 7 T

Figure 2: Average whole brain histograms for FA, qT1, MT, and CEST asymmetry parameters. Blue and pink indicate the mean with standard deviations for patients and controls, respectively. a) Fractional Anisotropy, b) quantitative T1 values, c) magnetization transfer ratios, and d) MTR asymmetry at 3.5 ppm off‐resonance saturation.

155 Chapter 6

Figure 3: Quantitative MR measures along the group averaged motor tracts, sampled every 2 mm. Results plotted for group averaged left and right motor tracts. Blue and red indicate the

156 Multimodal tract‐based analysis in ALS patients at 7 T median value and inter‐quartile ranges for the patient and control group, respectively. Points highlighted with * show significant differences (Mann–Whitney–Wilcoxon test, P<0.05 two tailed.). Plots highlighted with * in the upper right corner show significant differences over the complete CST section. a) FA, indicating a decrease in FA values for the patient group toward the caudal part of the CST. The average difference between the medians of patients and controls is ‐ 0.035 over this section of the CST. b) Dtrans, showing a relative increase in patients compared to controls in the same caudal section of the CST. c) Dlong, showing no consistent differences between patients and controls. d) qT1, showing the effect of the selected bundle starting in the cortical grey matter (right side of the plot), traversing the reasonably homogeneous WM and then crossing deep lying grey matter. There are no significant differences between patients and controls. e) MTR, showing a trend along the entire bundle which was significant for the right CST, with patients displaying higher MT values than controls. The average difference in medians over this section of the CST is 0.02. f) APTw, showing no significant changes between patients and controls in the lower half of the segment,d an a single significant point towards the rostral part of the tract. No significant changes in qT1 (Figure 3d) values were found (left p = 0.57; right p = 0.36). The MTR measures (Figure 3e) demonstrated a significant increase in patients compared to controls along the right CST,(p = 0.05), again the changes in the left CST were not significant (p = 0.24). As can be observed in the graph, the measures show largely overlapping ranges for patients and controls, limiting the reliability of group comparison. APTw scores along the averaged fiber bundles, showed no significant changes between patients and controls (left p = 0.54; right p = 0.21) (Figure 3f).

We did not find a significant correlation using Pearson’s correlation testing between imaging measures (averaged along the CST) and clinical markers.

Discussion

In this study, quantitative MR parameters were compared between ALS patients and age‐matched healthy controls, both in a ‘whole brain’ analysis and specifically along the CST. We found a significant decrease in FA along the CST, consistent with previous studies(10,33,34). Increased diffusivity perpendicular to the CST (Dtrans) was found underlying the decrease in FA. Additionally, exploratory data, acquired on ultra high field MRI, showed a significant increase in MTR in the right CST in the patient group. The other MRI contrasts, qT1 and APTw imaging, did not reveal significant changes.

157 Chapter 6

We observed a reduced FA both subcortical and in the posterior limb of the capsula interna, which is consistent with previous findings (10,33). To gain more insight in the underlying changes in the brain tissue causing reduced FA in the CST, other diffusion‐based parameters were assessed. Decomposing the FA values into the diffusivities parallel and transverse to the CST direction revealed a significantly increased Dtrans and no significant changes in Dlong in patients with respect to controls. An increase in the observed Dtrans indicates a loss of diffusion restriction perpendicular to the main direction of diffusion. Diffusivity changes were asymmetrical with predominant changes within the right CST, which has been repeatedly observed in ALS (35–37). It has been suggested that the right hemisphere is more vulnerable to the neurodegenerative process (38).

MTR is considered an indicator for the capacity of water bound to macromolecules in (nervous) tissue to exchange magnetization with unbound water molecules (23). In white matter tissue, MTR is mainly determined by the myelin content and, to a lesser extent, the number of or gliosis (39). Demyelination has been demonstrated to cause a reduced MTR, however, in this study we found a significant increase of MTR in the right CST. This finding is different from what we expected and from what has been previously reported (40,41), therefore we need to interpret these data cautiously. As illustrated by Figure 6c the data shows largely overlapping variation and the effect size is small. Also, there is relative little experience with magnetization transfer data on 7 T and it has been suggested that MTR in white matter is more variable at 7 T compared to 3 T (23). However, since MTR is sensitive for the magnetization transfer between macromolecular and liquid protons, an increase of the liquid fraction in tissue has the potential to increase MTR (21). This hypothesis is supported by the finding of increased transverse diffusion, in addition, accumulation of pathological protein aggregates (macromolecules) ‐ as is well known to occur in ALS ‐ might also result in an increased overall magnetic transfer. The correspondence of predominant changes in the right CST for diffusivity characteristics and MTR measures might as well increase the credibility of our findings. The combination of decreased FA and increased MTR in the CST of patients is remarkable since in general a decreased FA (e.g.

158 Multimodal tract‐based analysis in ALS patients at 7 T due to demyelination) is related to a decreased MTR. Therefore, we further assessed the relationship between FA and MTR by assessing the correlation of these two parameters within individual subjects. This analysis confirmed a positive correlation between these two modalities, indicating that at a group level it is unlikelyt tha demyelination is the primary cause of reduced FA in patients. Summarizing, an increased tissue liquid fraction, potentially in combination with protein accumulation, might have caused the increase in MTR values (42). These effects apparently outweigh some degree of demyelination, which is supported by findings in post‐mortem studies in ALS (43,44). Based on these data, pathological changes which lead to an increase in the liquid fraction such as proliferation of glial cells and extracellular matrix expansion (45–47) are more likely to be the primary cause of reduced FA within the CST of patients with ALS.

CEST and quantitative T1 imaging measurements did not show significant differences between patients and controls along the CST. APTw imaging – a variant of CEST ‐ at high field has been suggested as a noninvasive biomarker of white matter pathology, potentially providing complementary information to other MRI methods in current clinical use (48). Grey matter is known to show lower APTw contrast compared to white matter, potentially due to fewer membrane‐associated (extracellular, cytosolic, and transmembrane) proteins such as those found in myelin (49). More research on the effects of brain tissue pathology and the APTw contrasts is needed to further elucidate its dynamics in health and disease. For ALS, the sensitivity of APTw imaging to white matter pathology appears to be lower compared to DTI measures such as FA. No changes in T1 contrast were found, which is contrary to previously published results showing that T1 weighted imaging with MTC enhancement allowed for the visualization of CST lesions (50,51). T1 contrast has also shown to be related to the degree of myelination and the amount of bulk water (52). For example, lesions in MS show shortened T1 values indicating plaque formation while global histogram analysis indicated elongated T1 values, pointing at diffuse demyelination (53,54). To rule out effects in qT1 that might have been missed in the tract specific analysis due to misregistration of the qT1 parameter maps to their corresponding FA volumes, the whole brain volumes were compared

159 Chapter 6 using an additional VBM analysis to highlight significant differences (data not shown). This additional analysis did not yield any significant results. Together with our results using MTC, the lack of effects in APTw and qT1 imaging might lend further supoort that demyelination is not the main cause of FA decrease found in the CST of ALS patients. (46,47)

Given the exploratory nature of this study, there are limitations to which these results need to be interpreted. First, the sample size of patients and controls employed in this study is small and the variation of the measurements between subjects is relatively large, hampering the detection of subtle differences. Second, the modeling of MTR based methods, especially the calculation of APTw images is still a very active field of research and it is expected that future developments will improve both the sensitivity and reproducibility of this new method and that more factors influencing or contributing to this MR contrast will be elucidated.

To conclude, we explored multiple MR contrasts at ultra‐high field in relation to FA reduction in the CST of patients with ALS. MR contrasts were selected based on their sensitivity to neuronal loss and/or demyelination of white matter. A reduced FA in the CST of patients with ALS was confirmed and further showed to originate from an increase in Dtrans. In addition, we found an increase in MTR in the right CST, which makes demyelination as cause of reduced FA less likely. Increased magnetization transfer due to more free liquid spins is in accordance with the observed changes in diffusivity, particularly an increase in Dtrans. The other contrasts, qT1 and APTw imaging, did not reveal any significant changes in ALS. Future studies are needed to validate our findings in larger groups of patients and further advances in both MR technology and neuroscience are needed to increase insights into the dynamics of these MR contrasts in health and disease.

Acknowledgements

The authors like to thank all participants for their contribution to this study and dr. Alexander Leemans for helpful discussions and use of the ExploreDTI toolbox.

160 Multimodal tract‐based analysis in ALS patients at 7 T

References

1. Smith MC. Nerve Fibre Degeneration in the Brain in Amyotrophic Lateral Sclerosis. J Neurol Neurosurg Psychiatry. 1960 Nov 1;23(4):269–82. 2. Hecht M., Fellner F, Fellner C, Hilz M., Neundörfer B, Heuss D. Hyperintense and hypointense MRI signals of the precentral gyrus and corticospinal tract in ALS: A follow‐up examination including FLAIR images. J Neurol Sci. 2002 Jul 15;199(1– 2):59–65. 3. Turner MR, Agosta F, Bede P, Govind V, Lulé D, Verstraete E. Neuroimaging in amyotrophic lateral sclerosis. Biomarkers in Medicine. 2012;6(3):319–37. 4. Toosy A, Werring D, Orrell R, Howard R, King M, Barker G, et al. Diffusion tensor imaging detects corticospinal tract involvement at multiple levels in amyotrophic lateral sclerosis. J. Neurol. Neurosurg. Psychiatry. 2003 Sep 1;74(9):1250–7. 5. Karlsborg M, Rosenbaum S, Wiegell M, Simonsen H, Larsson H, Werdelin L, et al. Corticospinal tract degeneration and possible pathogenesis in ALS evaluated by MR diffusion tensor imaging. Amyotroph. Lateral. Scler. Mot. Neuron Disord. 2004 Sep;5(3):136–40. 6. Wang S. Amyotrophic Lateral Sclerosis and Primary Lateral Sclerosis: The Role of Diffusion Tensor Imaging and Other Advanced MR‐Based Techniques as Objective Upper Motor Neuron Markers. Annals of the New York Academy of Sciences. 2005 Dec;1064(1):61–77. 7. Blain CRV, Williams VC, Johnston C, Stanton BR, Ganesalingam J, Jarosz JM, et al. A longitudinal study of diffusion tensor MRI in ALS. Amyotroph. Lateral. Scler. 2007 Jan;8:348–55. 8. Wong JCT, Concha L, Beaulieu C, Johnston W, Allen PS, Kalra S. Spatial Profiling of the Corticospinal Tract in Amyotrophic Lateral Sclerosis Using Diffusion Tensor Imaging. J Neuroimaging. 2007 Jul 1;17(3):234–40. 9. Agosta F, Pagani E, Petrolini M, Caputo D, Perini M, Prelle A, et al. Assessment of White Matter Tract Damage in Patients with Amyotrophic Lateral Sclerosis: A Diffusion Tensor MR Imaging Tractography Study. Am. J. Neuroradiol. 2010 Sep;31(8):1457–61. 10. Verstraete E, van den Heuvel MP, Veldink JH, Blanken N, Mandl RC, Hulshoff Pol HE, et al. Motor Network Degeneration in Amyotrophic Lateral Sclerosis: A Structural and Functional Connectivity Study. Zhan W, editor. PLoS ONE. 2010 Oct 27;5:e13664. 11. Senda J, Kato S, Kaga T, Ito M, Atsuta N, Nakamura T, et al. Progressive and widespread brain damage in ALS: MRI voxel‐based morphometry and diffusion tensor imaging study. Amyotroph. Lateral. Scler. 2011 Jan;12:59–69. 12. Keller J, Vymazal J, Ridzoň P, Rusina R, Kulišt’ák P, Malíková H, et al. Quantitative brain MR imaging in amyotrophic lateral sclerosis. MAGMA. 2011 Mar 15;24(2):67– 76. 13. Canu E, Agosta F, Rive N, Sala S, Prelle A, Caputo D, et al. The Topography of Brain Microstructural Damage in Amyotrophic Lateral Sclerosis Assessed Using Diffusion Tensor MR Imaging. Am. J. Neuroradiol. 2011 Aug;32(7):1307–14.

161 Chapter 6

14. Zhang Y, Schuff N, Woolley SC, Chiang GC, Boreta L, Laxamana J, et al. Progression of white matter degeneration in amyotrophic lateral sclerosis: A diffusion tensor imaging study. Amyotroph. Lateral. Scler. 2011 Nov;12(6):421–9. 15. Rose S, Pannek K, Bell C, Baumann F, Hutchinson N, Coulthard A, et al. Direct evidence of intra‐ and interhemispheric corticomotor network degeneration in amyotrophic lateral sclerosis: An automated MRI structural connectivity study Neuroimage. 2012 Feb 1;59(3):2661–9. 16. Acosta‐Cabronero J, Williams GB, Pengas G, Nestor PJ. Absolute Diffusivities Define the Landscape of White Matter Degeneration in Alzheimer’s Disease. Brain. 2010 Feb 1;133(2):529–39. 17. Beaulieu C. The basis of anisotropic water diffusion in the nervous system ‐ a technical review. NMR in Biomedicine. 2002;15(7‐8):435–55. 18. Glasser MF, Van Essen DC. Mapping Human Cortical Areas In Vivo Based on Myelin Content as Revealed by T1‐ and T2‐Weighted MRI. J. Neurosci. 2011 Aug 10;31(32):11597–616. 19. Paus T, Collins D., Evans A., Leonard G, Pike B, Zijdenbos A. Maturation of white matter in the human brain: a review of magnetic resonance studies. Brain Research Bulletin. 2001 Feb;54(3):255–66. 20. Barkovich A. Magnetic resonance techniques in the assessment of myelin and myelination. Journal of Inherited Metabolic Disease. 2005 May 19;28(3):311–43. 21. Henkelman RM, Stanisz GJ, Graham SJ. Magnetization transfer in MRI: a review. NMR in Biomedicine. 2001 Apr 1;14(2):57–64. 22. Pike G. Magnetization transfer imaging of multiple sclerosis. The Italian Journal of Neurological Sciences. 1997 Jun 21;18(6):359–65. 23. Mougin OE, Coxon RC, Pitiot A, Gowland PA. Magnetization transfer phenomenon in the human brain at 7 T. NeuroImage. 2010 Jan 1;49(1):272–81. 24. Verstraete E, Veldink JH, Mandl RCW, van den Berg LH, van den Heuvel MP. Impaired Structural Motor Connectome in Amyotrophic Lateral Sclerosis. He Y, editor. PLoS ONE. 2011 Sep 2;6:e24239. 25. Leemans A, Jeurissen B, Sijbers J, Jones D. ExploreDTI: a graphical toolbox for processing, analyzing, and visualizing diffusion MR data. Proceedings 17th Scientific Meeting, International Society for Magnetic Resonance in Medicine, Honolulu. 2009. p. 3537. 26. Leemans A, Derek K. Jones. The B‐matrix must be rotated when correcting for subject motion in DTI data. Magnetic Resonance in Medicine. 2009;61(6):1336–49. 27. Chang L, Jones DK, Pierpaoli C. RESTORE: Robust estimation of tensors by outlier rejection. Magn Reson Med. 2005 May 1;53(5):1088–95. 28. Basser PJ, Sinisa Pajevic, Carlo Pierpaoli, Jeffrey Duda, Akram Aldroubi. In vivo fiber tractography using DT‐MRI data. Magnetic Resonance in Medicine. 2000;44(4):625–32. 29. Ordidge RJ, Gibbs P, Chapman B, Stehling MK, Mansfield P. High‐speed multislice T1 mapping using inversion‐recovery echo‐planar imaging. Magn Reson Med. 1990 Nov;16(2):238–45. 30. Clare S, Jezzard P. Rapid T1 mapping using multislice echo planar imaging. Magn Reson Med. 2001 Apr;45(4):630–4.

162 Multimodal tract‐based analysis in ALS patients at 7 T

31. Jones CK, Polders D, Hua J, Zhu H, Hoogduin HJ, Zhou J, et al. In vivo three‐dimensional whole‐brain pulsed steady‐state chemical exchange saturation transfer at 7 T. Magn Reson Med. 2012;67(6):1579–89. 32. Mandl RCW, Schnack HG, Luigjes J, van den Heuvel MP, Cahn W, Kahn RS, et al. Tract‐based Analysis of Magnetization Transfer Ratio and Diffusion Tensor Imaging of the Frontal and Frontotemporal Connections in Schizophrenia. Schizophrenia Bull. 2010 Jul 1;36(4):778 –787. 33. Graaff MM van der, Sage CA, Caan MWA, Akkerman EM, Lavini C, Majoie CB, et al. Upper and extra‐motoneuron involvement in early motoneuron disease: a diffusion tensor imaging study. Brain. 2011 Apr 1;134(4):1211–28. 34. Filippini N, Douaud G, Mackay CE, Knight S, Talbot K, Turner MR. Corpus callosum involvement is a consistent feature of amyotrophic lateral sclerosis. Neurology. 2010 Nov 2;75(18):1645–52. 35. Cosottini M, Pesaresi I, Piazza S, Diciotti S, Cecchi P, Fabbri S, et al. Structural and functional evaluation of cortical motor areas in Amyotrophic Lateral Sclerosis. Experimental Neurology. 2012 Mar;234(1):169–80. 36. Agosta F, Pagani E, Rocca M a., Caputo D, Perini M, Salvi F, et al. Voxel‐based morphometry study of brain volumetry and diffusivity in amyotrophic lateral sclerosis patients with mild disability. Human Brain Mapping. 2007;28(12):1430–8. 37. Kassubek J, Unrath A, Huppertz H‐J, Lulé D, Ethofer T, Sperfeld A‐D, et al. Global brain and corticospinal tract alterations in ALS, as investigated by voxel‐ based morphometry of 3‐D MRI. Amyotroph. Lateral Scler. Other Motor Neuron Disord. 2005 Dec;6(4):213–20. 38. Chen Z, Ma L. Grey matter volume changes over the whole brain in amyotrophic lateral sclerosis: A voxel‐wise meta‐analysis of voxel based morphometry studies. Amyotrophic Lateral Sclerosis. 2010 Dec;11(6):549–54. 39. Stanisz GJ, Kecojevic A, Bronskill MJ, Henkelman RM. Characterizing white matter with magnetization transfer and T2. Magnetic Resonance in Medicine. 1999 Dec 1;42(6):1128–36. 40. Kato Y, Matsumura K, Kinosada Y, Narita Y, Kuzuhara S, Nakagawa T. Detection of pyramidal tract lesions in amyotrophic lateral sclerosis with magnetization‐transfer measurements. Am J Neuroradiol. 1997 Sep 1;18(8):1541–7. 41. Tanabe JL, Vermathen M, Miller R, Gelinas D, Weiner MW, Rooney WD. Reduced MTR in the corticospinal tract and normal T2 in amyotrophic lateral sclerosis. Magn Reson Imaging. 1998 Dec;16(10):1163–9. 42. Rocha AJ da, Maia ACM, Valério BCO. Corticospinal Tract MR Signal‐Intensity Pseudonormalization on Magnetization Transfer Contrast Imaging: A Potential Pitfall in the Interpretation of the Advanced Compromise of Upper Motor Neurons in Amyotrophic Lateral Sclerosis. AJNR Am J Neuroradiol. 2012 May 1;33(5):E79– E80. 43. Iwanaga K, Hayashi S, Oyake M, Horikawa Y, Hayashi T, Wakabayashi M, et al. Neuropathology of sporadic amyotrophic lateral sclerosis of long duration. J Neurol Sci. 1997 Mar 10;146(2):139–43.

163 Chapter 6

44. Ikeda K, Akiyama H, Arai T, Ueno H, Tsuchiya K, Kosaka K. Morphometrical reappraisal of motor neuron system of Pick’s disease and amyotrophic lateral sclerosis with . Acta Neuropathol. 2002;104(1):21–8. 45. Hughes JT. Pathology of amyotrophic lateral sclerosis. Adv Neurol. 1982;36:61–74. 46. Philips T, Robberecht W. Neuroinflammation in amyotrophic lateral sclerosis: role of glial activation in motor neuron disease. Lancet Neurol. 2011 Mar;10(3):253–63. 47. Ince P, Highley J, Kirby J, Wharton S, Takahashi H, Strong M, et al. Molecular pathology and genetic advances in amyotrophic lateral sclerosis: an emerging molecular pathway and the significance of glial pathology. Acta Neuropathol. 2011;122(6):657–71. 48. van Zijl PCM, Yadav NN. Chemical exchange saturation transfer (CEST): What is in a name and what isn’t? Magn Reson Med. 2011;65(4):927–48. 49. Dula AN, Asche EM, Landman BA, Welch EB, Pawate S, Sriram S, et al. Development of chemical exchange saturation transfer at 7T. Magn. Reson. Med. 2011;66(3):831–8. 50. Da Rocha AJ, Oliveira ASB, Fonseca RB, Maia ACM, Buainain RP, Lederman HM. Detection of Corticospinal Tract Compromise in Amyotrophic Lateral Sclerosis with Brain MR Imaging: Relevance of the T1‐Weighted Spin‐Echo Magnetization Transfer Contrast Sequence. Am J Neuroradiol. 2004 Oct 1;25(9):1509–15. 51. Carrara G, Carapelli C, Venturi F, Ferraris MM, Lequio L, Chiò A, et al. A Distinct MR Imaging Phenotype in Amyotrophic Lateral Sclerosis: Correlation between T1 Magnetization Transfer Contrast along the Corticospinal Tract and Diffusion Tensor Imaging Analysis. Am J Neuroradiol 2012; 33:733‐739. 52. Fatouros PP, Marmarou A. Use of magnetic resonance imaging for in vivo measurements of water content in human brain: method and normal values. Journal of Neurosurgery. 1999;90(1):109–15. 53. Neema M, Stankiewicz J, Arora A, Dandamudi VSR, Batt CE, Guss ZD, et al. T1‐ and T2‐Based MRI Measures of Diffuse Gray Matter and White Matter Damage in Patients with Multiple Sclerosis. Journal of Neuroimaging. 2007;17:16S–21S. 54. Vrenken H, Geurts JJG, Knol DL, Van Dijk LN, Dattola V, Jasperse B, et al. Whole‐ Brain T1 Mapping in Multiple Sclerosis: Global Changes of Normal‐Appearing Gray and White Matter. Radiology. 2006 Sep 1;240(3):811–20. 55. Deoni SCL. Quantitative Relaxometry of the Brain. Topics in Magnetic Resonance Imaging. 2010 Apr;21:101–13. 56. Tofts, P.S., Steens, C.A., van Buchem, M.A. MT: Magnetization Transfer. In: Tofts, P.S., editor. Quantitative MRI of the Brain: Measuring Changes Caused by Disease. John Wiley & Sons Ltd.; 2003. p. 258–98. 57. Zhou J, Payen J‐F, Wilson DA, Traystman RJ, van Zijl PCM. Using the amide proton signals of intracellular proteins and peptides to detect pH effects in MRI. Nat Med. 2003;9(8):1085–90.

164 Multimodal tract‐based analysis in ALS patients at 7 T

Supplementary material

Longitudinal relaxation time T1

The longitudinal relaxation time T1 is a central MR parameter that reflects the capability of the spin‐lattice to relax longitudinal magnetization back to equilibrium. In normal tissue, the speed of relaxation is directly dependent on the amount and nature of interactions of the bulk water with the surrounding lattice. In free water such as CSF (around 4500 ms at 7 T), T1 is long and in structured tissue such as myelinated white matter T1 is short (around 1100 ms at 7 T). By quantitatively determining the longitudinal relaxation time, as opposed to weighting an image with an unknown amount of T1 effect, we can discriminate small changes in T1 and thus identify small changes due to pathology (55). To calculate the T1 per voxel, the signal intensity after inversion of the magnetization can be sampled at different time points. This is called the inversion curve. In this paper it was modeled by the following expression:

 Tt 1   TTR 1    0  21  eeItI 

Where I(t) is the measured signal intensity at inversion time t, I0 is the fitted voxel intensity at the fully relaxed state, T1 is the longitudinal relaxation time, and TR is the repetition time of the sequence. The difference between the data and this expression was minimized using a Levenberg‐Marquardt optimization within the following parameter space: T1: [0 ‐ 10000] ms, I0: [0 – 10 × Ihighest], where Ihighest is the highest signal intensity of the sampled curve.

Magnetization transfer based measures Magnetization transfer imaging is a technique that probes the macro‐molecular environment of bulk water protons by assessing the amount of magnetization transfer from the bound water to the bulk water pool. This is done in a saturation transfer experiment, where the amount of signal reduction directly reflects the amount of bound water molecules (and therefore macromolecular content). To compare the amount of magnetization transfer between subjects, the saturation effect is expressed in the magnetization transfer ratio, which quantifies the amount of signal suppression due to magnetization transfer. The

165 Chapter 6 amount of magnetization transfer contrasct is directly depended on the amount of myelination of the tissue, with areas that are strongly myelinated showing large MTR values such as the corpus callosum, while deep white matter, sub‐ cortical white matter u‐fibers and grey matter areas displaying decreasing MTR values (56).

Because the off‐resonance frequency is relatively close to the water resonance, it is expected that direct water saturation has a significant effect on the total observed saturation. To correct for this, we fit and remove a Lorentzian lineshape from the measured saturation curves as follows.

In our experiments, the saturated volumes were normalized to the unsaturated volume and corrected for residual B0 offsets and direct water saturation by fitting the middle offset frequencies (i.e.: ±150, ±75, and 0 Hz) to describe a Lorentz lineshape:

 w2  1,,  AfwAL   sat  2   4 0  ffw sat 

Where A is the fitted amplitude of the Lorentz peak, w is the fitted line width at half minimum, f0 is the fitted peak center which corresponds to the water resonance frequency for that voxel, and fsat is the off resonance frequency of the saturation pulse. The signal intensities at the frequencies of interest were then found by spline interpolation of the acquired data. For each voxel, the MTR at a given saturation frequency fsat was calculated as follows:

 fS satsat  fMTR sat 1  fL sat S0

Where S( fsat) is the signal intensity with a saturation pulse at ~1200 Hz off resonance, normalized to the unsaturated image S0, and L(fsat) is the fitted Lorentz peak height for that frequency.

Another, more specific, mechanism for measuring magnetization exchange is by sensitizing for chemical exchange using a method called chemical exchange saturation transfer (CEST). This method has gained much traction in the last

166 Multimodal tract‐based analysis in ALS patients at 7 T few years, because it offers a fundamental insight in the chemical nature of the tissue, probing the concentration and exchange rates of endogenous amide protons (48). CEST measurement are performed by repeating MT measurement with varying off‐set frequencies of the saturation pulse. This results in a sampling ofe th Z‐spectra, which is roughly symmetric around the water resonance, except for those frequencies where the exchanging protons resonate. By means of asymmetry analysis, i.e. taking the difference between saturated image at the resonance frequency of interest and its opposite counterpart frequency, chemical transfer effects can be distinguished from the conventional MT effects. In this work, we focus on the exchanging amide protons, which resonate at +3.5 ppm from the water resonance. The resulting images are called APTw images. For many applications amide protons are of interest, due to their increased natural abundance in pathology and appropriate exchange rates for saturation experiments (57). In our experiments, APTw volumes were calculated by calculating the difference in MTR at +3.5 ppm and ‐3.5 ppm as follows:

APTw  MTRasym  5.3 ppm  MTR 5.3  MTRppm  5.3 ppm

167

Chapter 7. Conclusions

Reasoning draws a conclusion, but does not make the conclusion certain, unless the mind discovers it by the path of experience. Roger Bacon, natural philosopher, 1214–1294

169 Chapter 7

170 Conclusions

The aim of this thesis was to explore the possibilities of performing quantitative MR in human brain using 7 tesla MRI. This chapter recapitulates the major results from the methods included in this thesis and discusses their feasibility for use in clinical studies. The final section suggests some directions for future research, highlighting additional quantitative methods that have not been discussed in this thesis but might also gain from application at 7 T.

Quantitative T1 mapping

In Chapter 2, a highly efficient technique was implemented to map the longitudinal relaxation time constant T1 in a clinically relevant scan time of 4 minutes 10 seconds. It was also shown that this measurement is very precise

(within 1.5 to 5% of the estimated T1 value) and has excellent scan‐rescan reproducibility (about 1% over selected regions of interest). A method was developed to assess the uncertainty of estimated T1 values on a subject‐by‐ subject basis. Because this approach used the wild bootstrap method for estimation of uncertainty values, no additional measurements were necessary. This means that quality assurance could be performed on each subject by (automated) post‐processing, to monitor data quality continuously.

When this method was applied to a small group of intra‐cranial tumor patients, strong effects in T1 were observed (Chapter 5). Quantitative T1 (qT1) maps allowed for excellent delineation of tumor boundaries and good contrast between necrotic and tumor tissue.

This method for quantitative T1 mapping was also applied in a small cohort of ALS patients. While diffusion based parameters showed significant changes between patients and controls, no significant effect was observed in qT1

(Chapter 6). This begs the question whether the qT1 method is as sensitive as diffusion parameters for micro‐structural effects. It is difficult to answer this question without extensive knowledge of the underlying system. As the etiology of ALS is unknown, this might explain the absence of significant T1 changes in this limited group of subjects.

171 Chapter 7

Considering these results and the observation that the estimated T1 parameter maps are robust to the B1 inhomogeneities at 7 T, it was concluded that quantitative T1 mapping is a reliable tool for clinical research.

Chemical exchange saturation transfer

Chemical exchange saturation transfer (CEST) is a novel MR methodology that has gained much traction in the last 5 years. In Chapter 3, a method was introduced that makes CEST possible in clinical time frames; the pulsed steady state CEST method. This method is highly efficient, because of the interleaved approach to water saturation and signal readout. However, precise knowledge of the local B1 amplitude is necessary for the saturation effect to approach a reliable steady state. This is a necessity when this method is to be used for the modeling of physiological parameters such as pH and proton transfer rates. This remains an issue at the current 7 T system.

An implementation of the pulsed steady state CEST method was applied in two patient groups, tumor and ALS patients. In tumor patients (Chapter 5), large effects were observed at amide proton resonances, but also at the opposite resonance frequency. These opposite effects were hypothesized to be the result of Nuclear Overhauser Effects (NOE) and it remains to be elucidated which of these effects determine the final APTw effect in this pathology. Additionally, strong correlations with T1 were observed, suggesting that these factors need to be taken into account when applying the pulsed steady state CEST method in patient studies. A full characterization of the measured system, in terms of relaxation constants, B1 transmit and receive amplitudes, and B0 homogeneity is advised for full interpretation of the saturation effects.

In Chapter 6, an exploratory study in 11 ALS patients and age matched controls was performed, applying both conventional magnetization transfer (MT) and amide proton transfer weighted (APTw) imaging, in an effort to further elucidate the structural changes that might underlie this disease. However, neither MT nor APTw images showed significant differences between patients and controls. Again, this might be an issue of insufficient sensitivity of these methods, or the lack of a measurable signal in this small and heterogeneous

172 Conclusions group of subjects. As long as no clear etiology for ALS is known, this question cannot be answered.

Considering these results, it can be concluded that CEST measurements can be successfully applied at 7 T, with great potential for tissue characterization. The increased sensitivity of the pulsed steady state method at 7 T has shown that there are multiple factors that contribute to the saturation effect and care should be taken during interpretation. Further research in both molecular exchange mechanisms and B1 techniques will benefit the modeling and measurement of CEST data.

Diffusion tensor imaging

Diffusion tensor imaging (DTI) is an inherently SNR sensitive technique and it has a lot to gain from the improved SNR levels at 7 T. The results from Chapter 4 show that this is indeed the case. In areas where SNR is optimal at 7 T, the uncertainties in fitted DTI parameters drop significantly. This is relevant for clinical research because it makes interpretation of DTI parameters more straightforward and reduces the amount of subjects needed to achieve sufficient statistical power.

However, this increase in SNR is currently not homogeneously distributed over the imaging volume. This is largely due to the inhomogeneities in the B1 field, which prevent the accurate application of refocusing pulses, a necessity for DTI acquisitions. The inhomogeneity makes current 7 T DTI only usable for clinical research in specific regions of the brain. For wider clinical use, homogeneous full‐brain coverage is often necessary. The current developments in multi transmit RF methods and technology all focus on solving this problem. When B1 inhomogeneity is sufficiently addressed, DTI will be one of the first techniques to benefit from application at 7 T.

Another factor that limits the performance of DTI at high field is the strength and speed of the gradients in the MRI system. In diffusion measurements strong gradients are applied to achieve diffusion weighting. The amount of time spent applying the gradients directly determines the echo time of the measurement, with longer echo times resulting in more T2 relaxation and signal

173 Chapter 7 loss. In addition, echo planar imaging was used to acquire the images. The duration of the EPI readout train directly influences the amount of blurring due to T2* relaxation, which is much increased at 7 T. Improved gradient performance by using a separate gradient insert for brain applications enables the acquisition of DTI images at increased resolution (less T2 signal loss and less

T2* blurring). This way, partial volume effects are reduced and the reliability of DTI measurements is increased, both in tractography applications and in quantitative diffusivity measures to monitor tumor treatments.

DTI offers unique anatomical information, i.e. the volumetric information of subject specific white matter tracts. DTI data acquired at a 3 T MRI scanner was used to perform the tract specific analysis in Chapter 6, where the corticospinal tracts were used as a framework to observe the dynamics of other MRI contrasts acquired at 7 T.

Suggestions for future research

The ability to excite spins in a homogeneous manner over the whole imaging volume is a necessity for many imaging sequences and even more so for sequences that rely on saturation or homogeneous inversion. Two out of three of the methods that have been investigated in this thesis depend strongly on the homogeneity of the B1 field. It is expected that multi‐transmit approaches to improve the B1 homogeneity will solve or at least greatly improve the current B1 related problems. At this moment, B1 inhomogeneity is the major factor preventing DTI to be used at the 7 T routinely. For CEST the problem is maybe less severe, APTw images can be created without the loss of too much signal. However, successful quantification and modeling of the CEST effect will be impossible without precise knowledge of the local B1 field behavior. Research focusing on improving B1 homogeneity will prove most useful for many MR scientists, as field strengths are ever increasing and the B1 problem will only become more prohibitive at higher fields.

Another suggestion for further research focuses on the effects observed in CEST imaging. Only the smallest tip of the iceberg of possible exchange mechanisms has been observed and many questions remain unanswered. For instance, if the

174 Conclusions

NOE effects arise from interactions with aliphatic protons, does that mean that it could be a sensitive marker for myelin, where such protons are abundant? If so, it might prove an additional marker for white matter characterization.

As mentioned above, quantitative T1 mapping is now a relatively mature technique. Its short scan times and possibilities for data quality monitoring make it an ideal candidate for inclusion in larger clinical studies. Especially in cases where diffuse pathology or slow disease progression are expected, this method will allow for robust collection of information of one of MRI’s central parameters.

This thesis focused only on three quantitative methods. Of course there are many more methods that could gain from the increased SNR available at 7 T. Here, two other methods are mentioned that deserve additional attention.

The first is arterial spin labeling, which strives to quantify one of the brains most important physiological processes: the transport of blood to and from the brain. By labeling the inflowing blood, many physiological parameters are quantified. An increase in SNR will benefit the modeling and improve our understanding of the processes involved. Combined with recent advances in multi‐channel transmission of RF, selective labeling of specific blood vessels might improve regional ASL applications.

The second quantitative method that has gained interest lately is that of quantitative functional MRI, or calibrated fMRI. In conventional fMRI, the measure of interest is that of statistical significance to find the locations in the braint tha show activation. Calibrated fMRI aims to quantify the effect size of the neuro‐vascular coupling itself, to obtain a measure of the strength of the activation. As the fMRI signal differences are in the order of a few percent of the baseline signal, improvements in SNR can make a large difference in the quality of these measurements.

Conclusion

In conclusion, this thesis describes the implementation of quantitative MR methods in the human brain at 7 T. By highlighting the drawbacks and

175 Chapter 7 advantages of the increased field strength, the use of 7 T MRI for quantitative measurements in clinical research was demonstrated. Inhomogeneities in the transmitted RF field limit the feasibility of methods that rely on the application of homogeneous RF pulses. The increased SNR at this high field strength enables rapid acquisition of high quality imaging volumes suitable for quantitative modeling while at the same time decreasing the uncertainty in fitted parameters.

176

Chapter 8. Conclusies

Redeneren brengt een conclusie, maar dat maakt deze nog niet zeker, behalve als het verstand haar ontdekt op het pad van ervaring. Roger Bacon, natuurfilosoof, 1214–1294

177 Chapter 8

178 Conclusies

Het doel van dit proefschrift was om de mogelijkheden van kwantitatieve magnetische resonantie beeldvorming toegepast in het menselijk brein (Engels: magnetic resonance imaging, MRI) bij een veldsterkte van 7 tesla te onderzoeken. In dit hoofdstuk passeren de belangrijkste resultaten van de onderzochte methodes in dit proefschrift de revue en wordt de haalbaarheid in klinische studies bediscussieerd. In het laatste onderdeel van dit hoofdstuk worden een aantal aanvullende kwantitatieve methoden gesuggereerd voor toekomstig onderzoek. Deze methoden worden niet behandeld in dit proefschrift, maar zouden evenzeer baat kunnen hebben bij toepassing op 7 tesla.

Kwantitatieve T1 bepaling en visualisatie

In hoofdstuk twee is een efficiënte methode geïmplementeerd die het mogelijk maakt om de longitudinale relaxatie tijd T1 te bepalen in een klinisch haalbare meettijd van 4 minuten en 10 seconden. In dat hoofdstuk is ook aangetoond dat deze methode erg precies is (met een onzekerheid tussen 1.5 en 5% van de bepaalde T1 waarde) en ook hoogst reproduceerbaar (met een variatie van ongeveer 1% binnen geselecteerde regio’s). Er is een methode ontwikkeld om de onzekerheid in de geschatte T1 waarde te bepalen met één meting. Deze methode gebruikt van de ‘wild‐bootstrap’ aanpak om de onzekerheid te bepalen en heeft hierdoor geen extra metingen nodig om een ondergrens van de onzekerheid te bepalen. Door deze methode samen met automatische beeldverwerking toe te passen is het mogelijk de kwaliteit van de data continu te controleren.

Deze methode om de T1 relaxatie tijd te bepalen is toegepast in een kleine groep patiënten met intra‐craniële hersentumoren en hierbij werden grote effecten in de bepaalde T1 waarden geobserveerd (Hoofdstuk 5). Kwantitatieve T1 beelden maakten een duidelijke afbakening van tumorgrenzen mogelijk en visualiseerden een goed contrast tussen necrotische gebieden en tumor.

De methode om de relaxatietijd T1 te bepalen was ook toegepast in een kleine groep patiënten met amyotrofische laterale sclerose (ALS). Hoewel metingen die gebaseerd waren op diffusie significante verschillen lieten zien tussen

179 Chapter 8 patiënten en gezonde controles, was dit niet het geval voor de kwantitatieve T1 metingen (Hoofdstuk 6). Het is de vraag of kwantitatieve T1 bepaling even gevoelig is voor de microscopische structurele effecten die plaats vinden binnen deze patiënten als diffusie parameters. Het is moeilijk om deze vraag te beantwoorden zonder uitgebreide kennis over het onderliggende systeem. Omdat de etiologie van ALS vooralsnog onbekend is, zou dit een afwezigheid van effecten in T1 metingen kunnen verklaren.

Deze resultaten in ogenschouw genomen en opmerkend dat de geschatte kwantitatieve T1 waarden vrij ongevoelig zijn voor de bestaande B1 inhomogeniteiten op 7 T, concluderen we dat kwantitatieve T1 bepaling en visualisatie een robuuste meting is die geschikt is voor het gebruik in klinisch onderzoek.

Overdracht van magnetische onderdrukking door chemische uitwisseling

De overdracht van magnetische onderdrukking door chemische uitwisseling (Engels: chemical exchange saturation transfer, CEST) is een nieuwe MR methodologie die de laatste 5 jaar erg in opkomst is. In hoofdstuk 3 is een nieuwe implementatie van deze techniek geïntroduceerd die het mogelijk maakt om CEST te meten binnen klinisch relevante scan ‐tijden: de ‘pulsed steady‐state CEST’ methode. Deze methode is zeer tijds‐efficient, omdat de acquisitie van de beeldinformatie wordt afgewisseld met het opleggen van magnetische onderdrukking. Exacte kennis van de amplitude van het B1 veld is echter nodig om een betrouwbare en robuste hoeveelheid CEST effect te meten. Dit is nodig als het doel is om deze meting te gebruiken als de basis voor modellen voor fysiologische parameters zoals zuurgraad en uitwisselingssnelheden van protonen. De amplitude en homogeniteit van het B1 veld is nog een een onopgelost probleem op het huidige 7 T MR systeem.

De pulsed steady‐state CEST methode is toegepast op patiënten met intracraniële tumoren of ALS. In patiënten met tumoren (Hoofdstuk 5) zijn grote CEST effecten gezien bij de resonantie frequenties van amide protonen, maar ook op de tegenoverliggende frequenties. Deze tegenoverliggende

180 Conclusies effecten zijn toegeschreven aan zo genaamde nucleaire Overhauser effecten (NOE) en het blijft de vraag welke rol deze effecten precies spelen in de uiteindelijke bepaling van amide proton transfer effecten in pathologie. Daarbij zijn sterke correlaties met de longitudinale relaxatietijd T1 geobserveerd, wat suggereert dat ook deze parameter in het oog gehouden dient te worden als CEST wordt toegepast in patiënten studies. Een volledige karakterisatie van het gemeten systeem, in termen van relaxatie constanten, B0 homogeniteit en B1 zend‐ en ontvangstamplitudes is van belang voor een volledige interpretatie van de verzadigingseffecten in CEST metingen.

In hoofdstuk 6 is een verkennende studie in 11 patiënten met ALS en gezonde controles beschreven. Hierbij zijn zowel conventionele magnetisatie overdrachtsmetingen (Engels: magnetization transfer, MT), als amide proton uitwisselingsmetingen (Engels: amide proton transfer, APT) toegepast, met als doel het ALS ziektebeeld op te helderen. Echter, MT noch APT‐gewogen beelden resulteerden in significante verschillen tussen patiënten en gezonde proefpersonen. Ook hierbij is het de vraag of deze methodes niet gevoelig genoeg zijn om verschillen behorend bij ALS op te pikken of dat deze heterogenep groe individuen te klein was, of dat er simpelweg geen meetbaar effect is in patiënten met ALS. Zo lang er niet meer bekend is over de etiologie van ALS, is het niet mogelijk hier uitsluitsel over te geven.

Deze resultaten in ogenschouw genomen, kunnen we concluderen dat CEST metingen succesvol kunnen worden toegepast binnen 7 T klinische studies en een groot potentieel bieden voor weefselkarakterisatie. De verhoogde gevoeligheid als gevolg van het sterkere magnetisch veld heeft laten zien dat er meerdere factoren zijn die bijdragen aan het gemeten CEST effect en waar rekening mee gehouden dient te worden in de interpretatien va deze data. Verdere ontwikkelingen in de theorie van zowel moleculaire uitwisselings effecten en B1 technieken zullen de modellering van CEST data verbeteren.

Diffusie tensor beeldvorming

Diffusie tensor beeldvorming (Engels: diffusion tensor imaging, DTI) is een techniek die gevoelig is voor de signaal‐ruis verhouding (Engels: signal to noise

181 Chapter 8 ratio, SNR) en heeft daarom veel te winnen bij de verbeterde SNR van de 7 T MRI scanner. De resultaten uit Hoofdstuk 4 laten zien dat dit inderdaad het geval is. In regio’s waar de SNR winst op de 7 T optimaal is gerealiseerd, worden significant kleinere onzekerheden in DTI parameters gemeten. Dit is relevant voor klinisch onderzoek omdat het de interpretatie van deze DTI parameters eenduidiger maakt en de benodigde hoeveelheid proefpersonen verminderd om statistisch significante resultaten te behalen.

Deze verbeterde SNR is met de huidige techniek echter niet homogeen verdeeld over het beeldvolume. Dit komt grotendeels door de inhomogeniteiten van het

B1 veld op 7 T, wat de effectiviteit van refocusserings pulsen hindert, iets dat essentieel is voor succesvolle DTI metingen. Door deze inhomogeniteiten zijn de huidige 7 T DTI metingen binnen klinisch onderzoek momenteel alleen mogelijk zijn in bepaalde regio’s in het brein. Om een bredere toepassing in klinisch onderzoek mogelijk te maken, is een homogene dekking van het gehele brein een noodzaak. De huidige wetenschappelijke ontwikkelingen op het vlak van B1 technieken die gebruik maken van meerdere zendkanalen richten zich vooral op het oplossen van dit B1 probleem. Op het moment dat dit voldoende is verholpen, is DTI een van de technieken die als eerste de vruchten zal plukken van toepassing op de 7 T.

Een andere factor die het gebruik van DTI op hogere veldsterktes hindert is de magnetische kracht en schakelsnelheid van het gebruikte gradiëntsysteem in de MRI scanner. Tijdens DTI metingen worden grote magnetische gradiënten gebruikt om zogenaamde diffusie weging te bewerkstelligen. De hoeveelheid tijd die nodig is om voldoende diffusieweging te verkrijgen beïnvloedt direct de echo‐tijd van de meting, waarbij langere echo‐tijden resulteren in meer T2 weging en er signaal verloren gaat.r Daa komt nog bij dat de signaal uitleesmethode ook gevoelig is voor de gebruikte gradiënten. Sterkere en snellere gradiënten maken het mogelijk om de uitleestijd en de hoeveelheid signaal verlies te beperken. Door verbeterde gradiëntsystemen te gebruiken, zoals een specifieke gradiënt‐insert voor brein toepassingen, kunnen DTI metingen versneld worden, waardoor er minder signaal verloren gaat. Deze signaalwinst kan dan worden toegepast om de resolutie van DTI metingen te vergroten. Dit vermindert de effecten van partiele voxel vulling door

182 Conclusies verschillende soorten weefsel, wat van belang kan zijn bij zowel witte stof baan reconstructie studies als bij het kwantitatief bepalen van diffusieparameters voor bijvoorbeeld tumorkarakterisatie.

DTI levert wetenschappers unieke anatomische informatie over het menselijk brein, namelijk de volumetrische ligging van persoonspecifieke witte stof verbindingen. DTI metingen uitgevoerd op eenI 3 T MR scanner zijn in Hoofdstuk 6 gebruikt om de cortico‐spinale witte stof verbindingen als raamwerk voor verdere analyses te gebruiken. Hierdoor konden andere MRI parameters langs deze essentiële witte stof baan bestudeerd worden.

Suggesties voor toekomstig onderzoek

De mogelijkheid om spins homogeen over het hele beeldvolume te aan te slaan is een essentiële voorwaarde voor veel MRI beeldvormings methoden, vooral voor die sequenties die gebruik maken van signaal onderdrukking of homogene inversie van het signaal. Twee van de drie methodes onderzocht in dit proefschrift zijn sterk afhankelijk van de homogeniteit van het B1 veld. Het valt te verwachten dat zenden met meerdere kanalen de homogeniteit van het opgelegde B1 veld zal verbeteren. Op dit moment zijn B1 problemen de belangrijkste factor die klinisch onderzoekers ervan weerhouden om DTI grootschalig op de 7 T toe te passen. Voor de CEST metingen zijn de problemen ogenschijnlijk minder nijpend, aangezien APT‐gewogen beelden opgenomen kunnen worden zonder al te veel problemen. Om CEST succesvol te modelleren en er relevante fysiologische parameters uit te bepalen, is echter precieze kennis van het B1 veld nodig. Onderzoek dat zich richt op het oplossen van het B1 probleem zal uiterst waardevol blijken voor onderzoekers die werken aan MRI op hoge veldsterktes, nu en in de toekomst, aangezien de verwachting is dat B1 problemen enkel zullen verergeren met de alsmaar toenemende veldsterktes.

Een andere suggestie voor vervolgonderzoek richt zich op de verschillende effecten die zijn opgemerkt tijdens CEST metingen. Op dit moment is slechts het kleinste topje van de ijsberg van mogelijke uitwisselingsmechanismen bekend en er zijn veel vragen nog onbeantwoord. Bijvoorbeeld: als signalen door NOE afkomstig zijn van interacties met alifatische protonen, betekent dat

183 Chapter 8 dan ook dat NOE een gevoelige marker zou kunnen zijn voor myelinestructuren waar deze protonen in grote mate voorkomen? Als dat het geval is, zouden CEST metingen die zich richten op NOE een additionele methode voor witte stof karakterisatie kunnen zijn.

Zoals hierboven al vermeld werd, is kwantitatieve T1 bepaling inmiddels een relatief volwassen techniek. Door de voorgestelde methode met korte scan‐ tijden en mogelijkheden voor (geautomatiseerde) data kwaliteitsmonitoring is het een ideale kandidaat om opgenomen te worden in grotere klinische studies. Specifiek in gevallen waar ziektebeelden zich diffuus door het brein laten zien, of waar ziekte zich langzaam over de tijd ontwikkelt is dit een robuste methode om kwantitatieve informatie over een van de centrale MRI parameters te verzamelen.

Dit proefschrift richtte zich slechts op drie mogelijke kwantitatieve methoden en natuurlijk zijn er vele andere methodes die baat zouden kunnen hebben van de verhoogde SNR op 7 T. Hier worden twee andere methoden genoemd die meer aandacht verdienen.

De eerste is een methode die ernaar streeft om één van de belangrijkste fysiologische processen in het brein te kwantificeren: het transport van bloed van en naar de hersenen. Met arteriële spin markering (Engels: , ASL) wordt een gedeelte van het inkomende bloed magnetisch gemerkt, waardoor verschillende fysiologische parameters gemodelleerd kunnen worden. Een verhoogde SNR zal deze modellering, en het inzicht in de processen die de uitwisseling van bloed beïnvloeden, verbeteren. Nieuwe ontwikkelingen op het gebied van zenden over meerdere kanalen maken preciezere selectieve markering van bloedvaten in regionale ASL mogelijk.

De tweede methode die de afgelopen tijd de interesse heeft gewekt is die van kwantitatieve functionele MRI (fMRI), ook wel gecalibreerde fMRI genoemd. In conventionele fMRI wordt de significantie van de verschillen in MRI signaal vertaald naar locale neuronale activiteit. Het doel van gecalibreerde fMRI is om het effect van de totale neurovasculaire koppeling te modelleren en zo de mate van activatie te bepalen (in plaats van enkel wel/geen activatie). Aangezien de verschillen in fMRI signaal maar een paar procent bedragen, kunnen

184 Conclusies verbeteringen in de SNR grote verschillen maken in de kwaliteit van deze metingen en de modellering ervan.

Conclusie

In dit proefschrift is de implementatie van een aantal kwantitatieve MRI methoden in het menselijk brein op 7 T besproken. Door de voor‐ en nadelen van het verhoogde magnetisch veld te benoemen, is het gebruik van 7 T MRI voor kwantitatieve metingen in klinisch onderzoek beschreven. De inhomogeniteiten van het RF transmissie veld stellen een limiet aan de haalbaarheid van methodes die erg afhankelijk zijn van homogene B1 velden. De verhoogde SNR op hoog magnetisch veld maken de acquisitie van beelden van hoge kwaliteit mogelijk, waardoor kwantitatieve modellering geoorloofd is, en de onzekerheid van de geschatte parameters vermindert.

185

Curriculum Vitae

187

188

Daniël Louis Polders was born on Saturday the 28th of November 1981 in Haarlem, The Netherlands. After high school, he started his Bachelors degree in molecular sciences at Wageningen University and Research Center in 2000. He completed his BSc with a thesis project at the biophysics lab, supervised by Frank Vergeldt and Henk van As. Taking a liking to the biophysics field and MR methods in particular, he continued his education with an MSc in molecular sciences, specializing in physical chemistry and performing his master thesis at the biophysics lab again, this time supervised by Bart Venne and Henk van As. During this program, he performed a six‐month research project in the group of prof. Paul Callaghan at Victoria University in Wellington, New Zealand. In 2007 he received his MSc degree. He then started a PhD program at the newly formed 7 T research group at the University Medical Centre Utrecht where he was supervised by prof. Peter Luijten and Hans Hoogduin. This project focused on the development of quantitative MRI methods at the 7 T platform. During his PhD program, he spent 2.5 months at the Kennedy Krieger Institute / Johns Hopkins hospital, in the group headed by prof. Peter van Zijl. In this short visit he worked with Craig Jones to develop a CEST method for use at the 7 T system.

Currently Daniel holds a post‐doc position at the NKI – Antony van Leeuwenhoek hospital (the Dutch cancer institute), at the department of radiotherapy lead by prof. Marcel van Herk. He is supervised by Uulke van der Heide in a project that investigates the uncertainties involved in low‐dose rate brachytherapy in cancer.

189

Dankwoord

He aha te mea nui o te ao? He tangata! He tangata! He tangata! -What is the most important thing in the world? It is people! It is people! It is people! sir Paul Callaghan, MR scientist 1947-2012

191

192

Het dankwoord van een proefschrift is zonder twijfel het meest gelezen hoofdstuk. Hoe exact of beta het onderzoek ook mag zijn geweest, het lijkt erop dat deze sociale kant van de wetenschap het meest gewaardeerd wordt. Hoewel het onderwerp van dit proefschrift misschien ver van de meeste mensen afstaat, is dat niet het geval voor het proces van wetenschap bedrijven. Het illustreert mooi dat klinisch onderzoek bovenal een sociale onderneming is. We proberen gezamenlijk meer van het menselijk lichaam en z’n gebreken te begrijpen. Klinische wetenschap wordt dan ook niet op een sociaal eiland bedreven, maar is stevig ingebed in een netwerk van wetenschappers, artsen, technici, en natuurlijk patiënten. Dit proefschrift is hier geen uitzondering op en in dit hoofdstuk zal ik proberen het sociale landschap te schetsen waarin ik dit werk heb mogen volbrengen. Een dergelijke schets uit zich in grote lijnen en grove vormen, waarin de mooiste details helaas niet tot uitdrukking komen. Houdt dat dus in het achterhoofd als je je naam hier niet expliciet terugvindt. Zonder jouw bijdrage was mijn tijd bij het UMCU niet hetzelfde geweest!

Allereerst wil ik alle patiënten en gezonde vrijwilligers bedanken voor hun onbaatzuchtige inzet en enthousiasme voor wetenschappelijk onderzoek. Zonder jullie gezindheid om een uur of meer in een smalle koker te liggen had dit proefschrift nooit tot stand kunnen komen. Bedankt!

Hans, Peter, een promotie onderzoek is voor de promovendus een ontdekkingsreis in een nieuw land en het is de taak van de (co‐) promotoren om als gids op te treden. De moeilijkheid ligt hem in het vinden van de juiste balans tussen het bij de hand nemen van de promovendus en hem zijn eigen weg laten vinden. Ik ben blij dat ik onder jullie hoede altijd vrij gelaten werd om de interessante aspecten van 7 T MRI op .te zoeken Dat maakte deze reis veelzijdig en altijd interessant. Nog blijer ben ik dat jullie mij uiteindelijk de weg terug hebben laten zien, zodat er een mooi afgerond reisverslag over geschreven kon worden.

Sylvia, zonder iemand die de procedures, agenda’s, bureau’s en computers in de gaten houdt, kan geen enkele onderzoeker functioneren, laat staan een promovendus. Bedankt voor je zorgen om een werkzame infrastructuur voor onze groep te verzorgen.

193

Mariëlle, Alexander L, Jeroen H, Rene M, Nick, in jullie rol als officieuze part‐ time begeleiders waren jullie vooral motivator en een bron van inspiratie. Bedankt voor het brainstormen, het ping‐pongen van half uitgedachte ideëen, en het gewoonweg aanmoedigen. Dat geeft een promovendus hoop in bange dagen.

Doke, Dolf, het is een flink bouwwerk geworden, die studie van mij. Een hoge toren die niet begint bij de universiteit of zelfs de middelbare school, maar een fundering nodig heeft die al veel eerder is opgebouwd. Die fundering kan alleen tot stand komen met een stabiele thuissituatie waarin je de ruimte en de prikkeling wordt gegeven om je ten volste te ontwikkelen. Bedankt daarvoor.

Peter v Z, Craig, my visit to the Kennedy Krieger Institute was very brief, much too short to complete a research project. Nevertheless, the people at the F. M. Kirby Research Center for Functional Brain Imaging made it a welcoming and exciting place to visit and I have fond memories of my collaboration with you all.

Catalina, Vincent, Jeroen, Fredy, Jaco, Dennis, Natalia, als 7 T‐ers van het eerste uur zijn jullie de collegaʹs die ik het meest heb leren kennen en waarderen. Door part‐ time te gaan werken heb ik een normaal 4‐jarige promotie weten te rekken tot bijna vijf jaar, maar nu is het dan toch echt tijd om verder te gaan. Ik heb mijn relaties met jullie altijd veel meer als vriendschappelijk dan simpel professioneel beschouwd en zal de eindeloze discussies en vrimiboʹs missen.

Irene, Anouk, Wouter, Wiebe, Joep, Ronald, Alex, Tijl, het is goed om te zien dat de 7 T groep geen moeite heeft om nieuwe aanwas te vinden zonder aan de kwaliteit te hoeven toornen. In de tijd dat ik jullie heb leren kennen, ben ik er achter gekomen dat jullie zowel inhoudelijk als sociaal je mannetje weten te staan. Bedankt voor de gezelligheid!

Mandy, Mies, Anja, Esther V, Laura, Jill, Bertine, Nolan, de grote kracht van de 7 T groep is de samenkomst van technisch en klinisch onderzoek. Tijdens het ontwikkelen van nieuwe methoden is het essentieel om feedback te krijgen over welke resultaten ook klinisch interessant zijn. Bedankt voor het geven van die terugkoppeling, en voor de tijd die jullie namen om mijn statistische kennis uit

194

te breiden. Buiten dat, het was ook erg gezellig om bij de dokters te gaan buurten en misschien wat lekkers te bietsen!

Astrid, Hanneke, Anna, Alessandro, Alexander, eigenlijk horen jullie bij de afdeling radiotherapie, hoewel ik nooit zo goed begrepen heb waarom daar zoʹn strenge scheidslijn tussen zou moeten zitten. In de praktijk merken we daar gelukkig maar weinig van, behalve dan misschien de RT een stuk realiteitszin toevoegt die bij radiologen soms ontbreekt. Bedankt voor het oprekken van mijn referentiekader, en het helpen met het verhelderen van de concepten van RF techniek en wiskunde die mij in eerste instantie boven de pet gingen.

Hugo, Ingmar, Bart vd B, Michel en andere soldeerknapen. Het is een zeldzame situatie, een groep die zowel grote patientenstudies uitvoert als zelf actief is om de benodigde apparatuur te ontwikkelen en fabriceren. Ik hoop dat jullie evenveel over MRI hebben geleerd als ik van jullie over schakelingen, Smith charts en spoelen.

Geert, Gerdien, Bart L, Jurriaan, Esther A, Janneke E, Micha, Auke, Lianne, Jeroen, Karen, Henry, we kennen elkaar nu al zo lang, dat een hiaat van maanden in het onderling contact geen enkel probleem is. Bedankt voor jullie veelzijdige vriendschap die ongedwongen en vertrouwd is.

Janneke, mijn gebrek aan overzicht en planning deed jou af en toe afvragen of er ooit een einde zou komen aan mijn promotietraject. Ik weet dat zonder jouw steun en organisatievermogen het in ieder geval nog langer had geduurd. Ik heb me altijd gesterkt gevoeld door de gedachte datr er in iede geval iemand was die er niet aan twijfelde of ik het in me had om mijn promotie af te maken, ook als ik dat zelf helemaal niet zo zeker wist. Zonder jou was dit proefschrift er niet geweest. Bedankt.

Lucas en Simone, het was fijn om tijdens mijn promotietraject de tijd te kunnen nemen om veel thuis te zijn tijdens jullie eerste jaren. Ik hoop dat jullie trots zijn op je papa als je dit boekje later nog eens bekijkt. Ik ben in ieder geval trots op jullie.

195

197