HIGH-FIELD MAGNETIC RESONANCE FINGERPRINTING FOR MOLECULAR MRI
by
CHRISTIAN EDWIN ANDERSON
Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy
Dissertation Advisor: Dr. Christopher Flask
Department of Biomedical Engineering
CASE WESTERN RESERVE UNIVERSITY
August, 2018
CASE WESTERN RESERVE UNIVERSITY SCHOOL OF GRADUATE STUDIES
We hereby approve the thesis/dissertation of
Christian E. Anderson
Candidate for the degree of Doctor of Philosophy*.
Committee Chair Nicole Steinmetz Committee Member Christopher Flask Committee Member Xin Yu Committee Member Vikas Gulani Committee Member Connie Piccone
Date of Defense April 12, 2018
*We also certify that written approval has been obtained for any propriety material contained therein
i
Dedication I want to first dedicate this work to Alex. She was always a willing partner for getting free food and a tolerant audience for practicing oral presentations.
I’d also like to dedicate this work to my family: Erik, Theresa, Paige, and Nolan. Thank you for your support and always being there for me.
ii
Table of Contents List of Tables ...... vi List of Figures ...... vii Acknowledgements ...... viii Abstract ...... 1 Chapter 1: Introduction to MRI ...... 3 1.1 The MRI Signal ...... 4 1.1.1 Physical Source of MRI Signal ...... 4 1.1.2 RF Excitation and Detection ...... 6 1.1.3 Magnetic Relaxation in MRI ...... 8 1.2 Spatial Localization and Image Formation ...... 10 1.2.1 Selective Excitation ...... 11 1.2.2 Magnetic Field Gradients and k-space ...... 12 1.2.3 Relationship Between k-space and Image Space ...... 15 1.2.4 The MRI Pulse Sequence...... 17 1.2.5 The Signal-to-Noise Ratio...... 20 1.3 Quantitative MRI ...... 22
1.3.1 T1 and T2 Mapping ...... 23 1.4 Paramagnetic MRI Contrast Agents ...... 25 1.4.1 MR Effects of Exogenous MRI Contrast Agents ...... 25 1.4.2 Relaxivity Constants: r1 and r2 ...... 26 1.4.3 Common (and Uncommon) MRI Contrast Agents ...... 27 1.5 Preclinical MRI at High Fields ...... 29 1.5.1 Animal Challenges ...... 29 1.5.2 Hardware Challenges ...... 31 1.6 Dissertation Overview ...... 32 Chapter 2: Fundamentals of Magnetic Resonance Fingerprinting ...... 34 2.1 Overview of MR Fingerprinting ...... 34 2.1.1 MRF Image Acquisition ...... 34 2.1.2 Creation of the MRF Dictionary ...... 37 2.1.3 Quantification via Dictionary Matching ...... 40 2.2 Reduction of MRF Acquisition Time ...... 41 2.2.1 Acceleration of the MRF Acquisition ...... 42
iii
2.3 High-Field Preclinical Magnetic Resonance Fingerprinting ...... 44 2.4 Summary ...... 45 Chapter 3: Regularly Incremented Phase Encoding – MR Fingerprinting ...... 47 3.1 Introduction ...... 47 3.2 Methods ...... 48 3.2.1 RIPE-MRF Encoding ...... 48 3.2.2 MRF Design and Quantification ...... 49 3.2.3 In vivo RIPE-MRF ...... 50 3.3 Results ...... 53 3.3.1 In vivo Studies ...... 53 3.4 Discussion ...... 58 3.5 Conclusions ...... 61 3.6 Supporting Methods ...... 62 3.6.1 MRF Sequence ...... 62 3.6.2 MRF Dictionary and Quantification: ...... 63 3.6.3 Phantom MRF Studies ...... 63
3.6.4 Conventional MRI Measurement of T1 and T2 Relaxation Times: ...... 64 3.7 Supporting Figures ...... 65 Chapter 4: Dual Contrast - Magnetic Resonance Fingerprinting (DC-MRF) ...... 67 4.1 Introduction ...... 67 4.2 Methods ...... 70 4.2.1 Multiple Contrast Agent Relaxation Model ...... 70 4.2.2 In vitro DC-MRF Assessments at 3 T ...... 73 4.3 Results ...... 76 4.3.1 60 MHz Relaxometry: Multiple Contrast Agent Relaxation Model Validation .. 76 4.3.2 3 T DC-MRF: Simultaneous Assessment of Two Paramagnetic MRI Contrast Agents 79 4.4 Discussion ...... 83 4.5 Conclusions ...... 89 Chapter 5: Summary and Future Directions ...... 90 5.1 Future Directions for MRF in Molecular MRI ...... 90 5.1.1 In vivo Application of DC-MRF ...... 90 5.1.2 Opportunities for DC-MRF Technical Development ...... 92
iv
5.1.3 Potential Applications of MRF for Molecular MRI ...... 93 5.2 Future Directions for High-Field Magnetic Resonance Fingerprinting ...... 95 5.2.1 Quantitative Opportunities in High-Field Preclinical MRF ...... 95 5.2.2 Technical Developments for High-Field Preclinical MRF ...... 97 5.3 Conclusion ...... 98 Bibliography ...... 99
v
List of Tables
Table 4.1…………………………………………………………………………………72 Table 4.2…………………………………………………………………………………80
vi
List of Figures
Figure 1.1………………………………………………………………………………..5 Figure 1.2………………………………………………………………………………..7 Figure 1.3………………………………………………………………………………..9 Figure 1.4………………………………………………………………………………12 Figure 1.5………………………………………………………………………………15 Figure 1.6………………………………………………………………………………17 Figure 1.7………………………………………………………………………………19 Figure 1.8………………………………………………………………………………20 Figure 1.9………………………………………………………………………………21 Figure 1.10..……………………………………………………………………………23 Figure 1.11..……………………………………………………………………………30 Figure 2.1………………………………………………………………………………35 Figure 2.2………………………………………………………………………………40 Figure 2.3………………………………………………………………………………43 Figure 3.1………………………………………………………………………………49 Figure 3.2………………………………………………………………………………50 Figure 3.3………………………………………………………………………………53 Figure 3.4………………………………………………………………………………54 Figure 3.5………………………………………………………………………………55 Figure 3.6………………………………………………………………………………56 Figure 3.7………………………………………………………………………………57 Figure 3.8………………………………………………………………………………58 Figure S3.1……..………………………………………………………………………65 Figure S3.2………………………..……………………………………………………65 Figure S3.3………………………..……………………………………………………66 Figure 4.1………………………………………………………………………………77 Figure 4.2………………………………………………………………………………78 Figure 4.3………………………………………………………………………………78 Figure 4.4………………………………………………………………………………80 Figure 4.5………………………………………………………………………………81 Figure 4.6………………………………………………………………………………82 Figure 4.7………………………………………………………………………………83 Figure 4.8………………………………………………………………………………86 Figure 5.1………………………………………………………………………………91
vii
Acknowledgements
I would like to start by thanking the BME department and the MSTP for
supporting my PhD training. Over the course of my training I have had the opportunity to
work with many incredible people. Dr. Connie Piccone has been an invaluable mentor
and member of my Thesis Committee. Her clinical insight ensured I stayed focused on
the ultimate goal of helping patients. The MR research group at CWRU has provided me
with the opportunities, support, and resources to grow and develop as a scientist. Mark
Griswold, Nicole Seiberlich, Vikas Gulani, and the many post-docs and students have all
at some point challenged me to think more creatively about my research questions.
Drs. Xin Yu and Nicole Steinmetz have been incredible collaborators and
members of my Thesis Committee. Along with Dr. Susann Brady-Kalnay, their labs have provided me with tremendous amounts of support in method development, experimental design, and providing resources for experiments. They’ve helped foster a collaborative environment that allowed me to accomplish more than I could have alone. I’d also like to acknowledge the Small Animal Imaging Research Core staff, Bernie, Mike, Chunying
and Shannon. Their incredible and tireless work is a large part of what makes our
research successful and I am incredibly fortunate to have benefited from their help.
Finally, I’d like to acknowledge my advisor and mentor, Chris. He has guided and
challenged me to become a better scientist and person. His commitment and investment
in me and my training has been substantial and I cannot say thank you enough.
viii
High-Field Magnetic Resonance Fingerprinting for Molecular MRI by CHRISTIAN E. ANDERSON
Abstract
Magnetic Resonance Imaging (MRI) is a powerful imaging modality providing
exceptional soft-tissue contrast with widespread use in clinical and preclinical
applications. Exogenous paramagnetic contrast agents are often used during MRI studies
to improve tissue visualization and disease detection. The utility of MRI contrast agents
has led to the development of novel MRI contrast agents with improved magnetic
properties, specific targeting to disease markers, and responsiveness to physiological
conditions in vivo. These new agents are used in preclinical molecular MRI studies to
sensitively identify disease, evaluate tissue function, and report on the tissue
microenvironment. However, current molecular MRI studies are limited to single agent
use despite the existence of situations where two or more agents would provide useful
and complementary information. This is primarily due to the inability of contrast
enhanced MRI studies to uniquely identify individual agents when two or more agents
are used simultaneously.
In this work, we propose using the multi-property quantification of the novel
Magnetic Resonance Fingerprinting (MRF) technique as a method for uniquely quantifying the local concentration of two co-administered paramagnetic MRI contrast agents. MRF is a wholly new way to perform quantification of MRI-specific tissue
1
properties and has demonstrated the ability to simultaneously quantify multiple tissue
properties with high temporal resolution and robustness to imaging artifacts. First, we
describe the Regularly Incremented Phase Encoding-MRF (RIPE-MRF) method for improving the motion resistance of preclinical MRF scans. RIPE-MRF uses a variable acquisition scheme that suppresses the appearance of motion artifacts in the resulting quantitative MRF tissue property maps. Then, we present an extension to the fundamental
MRI contrast equations that enables simultaneous calculation of the concentration of two co-administered MRI contrast agents. We term this method Dual Contrast-Magnetic
Resonance Fingerprinting (DC-MRF) and show how it enables unambiguous analytical calculation of multiple agent concentrations. This work demonstrates how DC-MRF can be used in multi-agent molecular MRI studies creating numerous opportunities for subvoxel analysis of tissue fractions, MRI-based imaging of genetic expression, and measurement of enzyme activity.
2
Chapter 1: Introduction to MRI
Medical imaging has become an integral part of medical care providing non-
invasive assessments of human anatomy and function. Magnetic Resonance Imaging
(MRI) is a powerful imaging modality capable of producing images with exceptional soft-
tissue contrast without the use of ionizing radiation. Contrast in an MRI image can be due
1 2 to MRI-specific tissue properties (such as T1 and T2 ), tissue microstructure (diffusion ), or functional status (blood oxygenation level dependent contrast3) allowing multiple
unique assessments of disease in a single imaging session. In addition, MRI contrast agents
can be administered to improve the desired contrast making MRI an extremely versatile imaging modality with widespread use in both clinical and preclinical imaging.
The work presented herein seeks to develop new MRI-based methods for use in
novel quantitative molecular MRI applications and studies and this chapter will present the underlying concepts that form the foundation of the proposed work. It will start by describing the source of the MRI signal and how magnetic fields can be used to generate images with spatial information. This will lead to a description of quantitative MRI which is the process of estimating MRI-specific property values via MRI experiments. Finally, paramagnetic MRI contrast agents will then be briefly presented and the chapter will close with an introduction to preclinical MRI at high fields.
Before the details of MRI are described, it is important to outline the basic components of the MRI scanner. The most prominent feature is the main magnet constructed as a solenoid of superconducting wire designed to generate a constant, static, homogeneous magnetic field. The gradient coils are located within the main magnetic field and are used to create the spatially varying magnetic fields necessary to perform generate
3
spatially localized images (instead of non-localized NMR spectroscopic measurements).
The sample and radiofrequency (RF) coil(s) are also placed inside the main magnetic field
as well as inside the gradient coils. This RF coil is responsible for generating and/or
detecting the time varying magnetic fields used to interact with the sample. While there are
additional cables and hardware for signal conduction, amplification, and processing, the
main magnet, gradient coils, and RF coils are the components most relevant to this work.
1.1 The MRI Signal
1.1.1 Physical Source of MRI Signal
Magnetic Resonance Imaging derives its name from the phenomena of nuclear
magnetic resonance (NMR). NMR results from the behavior of nuclei with a net magnetic moment and angular momentum, referred to as spins, in the presence of magnetic fields.
When these spins are placed in a static magnetic field, they will tend to align with the magnetic field resulting in an excess of spins aligned in the same direction creating a non- zero net magnetic moment (Fig. 1.1). This spin excess, or net magnetization (M0), can be
4 calculated based on the temperature (T) and magnetic field strength (B0) :
= (1.1) 6 2 2 𝜌𝜌0𝛾𝛾 ℎ 𝑀𝑀0 𝐵𝐵0 𝜋𝜋𝜋𝜋𝜋𝜋 In Equation 1.1 k is Boltzmann’s constant, h is Planck’s constant, ρ0 is the total number of spins, and γ is the gyromagnetic ratio and it is a physical property of the nucleus being investigated. Any nucleus having spin can be detected and imaged using MRI but the vast majority of experiments utilize water protons due to their high natural abundance in biological samples.
4
Figure 1.1: Alignment of spins with and without the presence of a static magnetic field (B0). Without the B0 field the spins are randomly distributed and produce no net magnetization (left panel). In the presence of the B0 field the spins will tend to align with the field creating a net magnetization vector (right panel) with a magnitude calculated by Equation 1.1.
M0 and B0 from Equation 1.1 are often treated as constants in an MRI experiment
and they represent the equilibrium values of the net magnetization and the main magnetic
field, respectively. However, during an MRI experiment the net magnetization, M(t), and
the effective magnetic field, B(t), are time varying with appropriate vector components.
Conventionally, spins are treated in aggregate such that M(t) (with M(0) equal to M0)
possesses x-, y-, and z-components. When describing M(t), the z-component, Mz, is referred
to as the longitudinal magnetization and the x- and y-components are combined as the transverse magnetization, Mxy. Similarly, during an MRI experiment additional magnetic
fields are frequently added to the B0 field creating a dynamic magnetic field vector, B(t).
During an MRI experiment, the Bloch equations1 (neglecting relaxation effects, discussed later) can be used to describe the behavior of M(t) in the presence of an external magnetic field, B(t):
5
( ) = ( ) × ( ) (1.2) ��⃗ 𝑑𝑑𝑀𝑀 𝑡𝑡 � ��⃗ �⃗ The implication of Equation 1.2 is that𝑑𝑑 the𝑑𝑑 magnetization𝛾𝛾𝑀𝑀 𝑡𝑡 𝐵𝐵 𝑡𝑡 vector will experience a torque
due to the applied magnetic field and this will result in M(t) precessing about the axis of
B(t). In the case, where the only field applied is the external field, B0, assumed to be along
the z-axis (longitudinal axis), it can be found that spins (and by extension M(t)) precess
about the longitudinal axis with a characteristic frequency proportional to B0:
= (1.3)
𝜔𝜔0 𝛾𝛾𝐵𝐵0 This precession rate, ω0, is the Larmor frequency and it is also the resonant
frequency for a given nucleus. Any aspect of the MRI experiment that has a fundamental
frequency equal to ω0 fulfills the resonance condition and is deemed “on-resonant”. In light of this we can differentiate between two different frames of reference for observing NMR dynamics: the lab frame and the rotating reference frame. In the lab frame all frequencies are observed relative to 0 Hz precession about the z-axis and spins are seen to precess at
ω0. In the rotating reference frame frequency observations are made relative to precession
about the z-axis at ω0 so that on resonant frequencies and spins appear to precess at 0 Hz.
Descriptions in this work are assumed to be in the rotating reference frame unless stated to be in the lab frame.
1.1.2 RF Excitation and Detection
The presence of spins in an external magnetic field creates a net magnetization vector, M(t), but it does not produce a detectable MRI signal because there is no net precession of M(t) (although individual spins are precessing). An MR detectable signal is generated by rotating M(t) away from the longitudinal axis and into the transverse (x-y)
6
plane creating a non-zero Mxy component of M(t). Rotation of M(t) is accomplished through the application of a second, time-varying magnetic field generated by the RF coil.
This additional magnetic field, referred to as the B1 field or RF pulse, is typically applied
on-resonance to most efficiently rotate the spins away from the longitudinal axis.
Conventionally, the B1 field is applied orthogonal to the longitudinal axis causing M(t) to
4 precess about the B1 field with a frequency ω(t) and a total rotation angle φ(t) :
( ) = ( ) (1.4)
1 𝜔𝜔( 𝑡𝑡) = 𝛾𝛾𝐵𝐵 ( 𝑡𝑡) (1.5) 𝑡𝑡 𝑟𝑟𝑟𝑟𝑟𝑟 The rotation angle, φ(t), is often referred𝜙𝜙 𝑡𝑡 to as∫ 0the𝜔𝜔 excitation𝑡𝑡′ 𝑑𝑑𝑑𝑑′ angle or flip angle (denoted as
α) since it is the angle used to “excite” the sample or “flip” magnetization in to the
transverse plane. The process of RF excitation is shown schematically in Figure 1.2 and
results in a non-zero value of Mxy proportional to sin(α).
Figure 1.2: RF excitation in the rotating reference frame. Application of a B1 field will cause the M0 vector to rotate from position 1 to position 2 through an angle, α (frame 2). Once M0 has been rotated by α degrees the B1 field is removed and the result is a non-zero transverse component of the magnetization (Mxy). Shown is α=90° so that the entire magnetization vector is in the transverse plane so that Mz equals zero and Mxy equals M0 after excitation (ignoring relaxation effects).
After the magnetization is rotated into the transverse plane and the B1 field is turned off, M(t) will precess about the z-axis due to the B0 field (Equation 1.2) creating a time varying magnetic field. This time-varying magnetic field is detectable via the voltage it will induce in the receiver RF coil according to Faraday’s Law of Induction:
7
( ) = (1.6) 𝑑𝑑𝑀𝑀𝑥𝑥𝑥𝑥 𝑡𝑡 𝜀𝜀 − Where ε is the electromotive force induced on 𝑑𝑑coil𝑑𝑑 by the precessing magnetization. It is
important to note that the magnitude of ε is proportional to Mxy and that the detected signal,
s(t), is due to the entire volume of excited spins:
( ) ( ) ( ) (1.7) −𝑖𝑖𝑖𝑖 𝑡𝑡 𝑡𝑡 𝑥𝑥𝑥𝑥 𝑠𝑠 𝑡𝑡 ∝ �𝑉𝑉 𝑀𝑀 𝑟𝑟 𝑒𝑒 𝑑𝑑𝑑𝑑 1.1.3 Magnetic Relaxation in MRI
It could be inferred from the previous section that after RF excitation both the
longitudinal and transverse magnetization should return to their equilibrium values of M0
and 0, respectively. This return to equilibrium occurs as two separate processes, referred
to as longitudinal and transverse relaxation, having two different time constants associated
with them. Recalling the Bloch equations1 from before, they can be modified to incorporate relaxation terms:
= ( ) × ( ) + 1 ( ) 1 (1.8)
��⃗ 0 𝑧𝑧 𝑥𝑥𝑥𝑥 𝑑𝑑𝑀𝑀� ��⃗ �⃗ 1 2 𝑑𝑑𝑑𝑑 𝛾𝛾𝑀𝑀 𝑡𝑡 𝐵𝐵 𝑡𝑡 �𝑇𝑇 𝑀𝑀 − 𝑀𝑀 − �𝑇𝑇 𝑀𝑀 This equation introduces the two fundamental MR specific time constants
important for understanding the magnetic behavior of spins during an NMR/MRI
experiment, the T1 and T2 relaxation times. While the first term in the right-hand side of
the equation is the familiar precession term described in Equation 1.2, the second and third
terms describe magnetic relaxation. We can look exclusively at the relaxation terms to generate differential equations for the longitudinal and transverse components of
magnetization:
8
= 1 ( ) (1.9) 𝑧𝑧 𝑑𝑑𝑀𝑀 0 𝑧𝑧 �𝑑𝑑𝑑𝑑 �𝑇𝑇1 𝑀𝑀 − 𝑀𝑀 = 1 (1.10) 𝑥𝑥𝑥𝑥 𝑑𝑑𝑀𝑀 𝑥𝑥𝑥𝑥 � − 2 𝑀𝑀 Solving these for their respective components𝑑𝑑𝑑𝑑 gives�𝑇𝑇 analytical expressions for the behavior
of Mz and Mxy:
= (0) + (1 ) (1.11) 𝑡𝑡 𝑡𝑡 − �𝑇𝑇1 − �𝑇𝑇1 𝑧𝑧 𝑧𝑧 0 𝑀𝑀 𝑀𝑀 𝑒𝑒= (0𝑀𝑀) − 𝑒𝑒 (1.12) 𝑡𝑡 − �𝑇𝑇2 𝑥𝑥𝑥𝑥 𝑥𝑥𝑥𝑥 The equation for Mz confirms that𝑀𝑀 our magnetization𝑀𝑀 𝑒𝑒 will recover back toward M0, doing
so at a rate proportional to T1 through a process referred to as spin-lattice relaxation (Fig
1.3), and that Mxy will decay to zero at a rate proportional to T2 due to spin-spin relaxation
(Fig 1.3). These exponential models can be used to predict the magnitude of the
magnetization vector components (Mz, Mxy) in any MRI experiment for a given pair of T1
and T2.
Figure 1.3: Examples of T1- and T2-relaxation as described by the Bloch equations (Equation 1.11 and 1.12) normalized to M0=1. Evaluation of the Bloch equations reveals the exponential behavior of the magnetization due to T1- and T2-relaxation. In this example the longitudinal magnetization was inverted to -M0 to better demonstrate T1-relaxation. It can be seen how T1-relaxation recovers longitudinal magnetization, Mz, for use in RF excitation while T2-relaxation results in decay of Mxy and causes signal loss.
9
It is important to note that transverse relaxation can be accelerated by the presence
of spatial heterogeneities in the B0 field (ΔB0). Field inhomogeneities create different precession frequencies across a voxel resulting in a differential phase accumulation, Δφ:
= (1.13)
0 This phase accumulation dephases the∆ 𝜑𝜑spins𝛾𝛾 ∆causing𝐵𝐵 𝑡𝑡 a reduction in the net transverse
magnetization vector accelerating signal decay. The rate of decay due to field
inhomogeneity is governed by T2′ and can be combined with T2 into a net transverse decay
term often referred to as T2*:
1 1 1 = + (1.14) ∗ ′ � 2 � 2 � 2 T2′ can be written in terms of ΔB𝑇𝑇 0 allowing𝑇𝑇 T2*𝑇𝑇 decay to be related to the field
inhomogeneity:
= (1.15) ′ 1 𝑇𝑇2 1 𝛾𝛾∆𝐵𝐵0 = + (1.16)
∗ 0 2 2 𝛾𝛾∆𝐵𝐵 By definition T2* is always shorter than𝑇𝑇 T2𝑇𝑇 and can severely reduce the available signal in
the presence of inhomogeneous fields.
1.2 Spatial Localization and Image Formation
To this point, the signal detected by the RF coil contained no information about the
relative location of each spin. In order to generate an image, spatial information must be
encoded into the signal so that each spin contributing to the signal can be spatially
localized. In addition, the currently described application of an RF pulse will excite the
entire sample, but the majority of imaging studies are only concerned with a specific area
of the sample requiring some form of selective excitation to image only the area of interest.
10
Both spatial localization and selective excitation involve some form of spatial encoding
which can be accomplished in MRI through the use of magnetic field gradients. Gradients,
G(r,t), create spatially varying magnetic fields that cause protons at different positions, r,
to have different precession frequencies:
( , ) = + ( , ) (1.17)
0 Spatially dependent precession 𝜔𝜔frequencies𝑟𝑟 𝑡𝑡 𝛾𝛾� can𝐵𝐵 then𝑟𝑟𝑟𝑟 be𝑟𝑟 used𝑡𝑡 � to localize protons within the
sample allowing imaging of specific regions and formation of images.
1.2.1 Selective Excitation
Slice selection is the process of restricting RF excitation (and subsequently
detectable signal) to only a subset of spins within the sample. Selective excitation is
accomplished through the combination of an amplitude modulated B1 field (shaped RF pulse) and a magnetic field gradient. The use of a shaped RF pulse introduces frequency dependent excitation that causes only spins within a certain range of precession frequencies
(the pulse bandwidth with frequency range ω1 to ω2 in Fig. 1.4) to rotate into the transverse
plane leaving spins outside this range unperturbed. As previously mentioned, application
of a gradient will create a range of precession frequencies, ω(r,t), across the sample according to Equation 1.17. Playing both fields at the same time grants control over the area of the sample that will be excited and generate signal. The size and location of this area can be modified via the amplitude of the gradient field, G(r,t), or the timing and shape of the RF pulse. A Fourier transform (FT) of the pulse envelope reveals the bandwidth and relative frequency dependent excitation for a given pulse (Fig. 1.4) allowing selection of
the appropriate pulse duration and shape for the MRI experiment. The benefit of selective
excitation is that any further gradient manipulation will only affect the excited spins.
11
Figure 1.4: Frequency selective excitation profile of different RF-pulses. B1(t) describes the amplitude modulation of the B1 magnetic field in the rotating reference frame. The slice profile is the resulting spectrum from the Fourier transform of the pulse envelope and reveals the bandwidth (highlighted by the gray box, range from ω1 to ω2) and relative excitation efficiency at different precession frequencies. The rect pulse has substantial side-lobe excitation while sinc pulses have a sharp frequency response preferred for selective excitation. 1.2.2 Magnetic Field Gradients and k-space
While slice selection provides an element of spatial localization, the signal within
this subset of spins still has no spatial information. Gradients can provide spatial
information to the excited spins by creating spatially dependent precession frequencies in
a process referred to as frequency encoding. This is similar to slice selection except there
is no additional B1 field and the spatially dependent precession frequencies, ω(r,t), are
along a separate dimension (the x-axis for illustrative purposes):
12
( , ) = + ( , ) (1.18)
0 Here each spin with a different x coordinate𝜔𝜔 𝑥𝑥 will𝑡𝑡 have𝛾𝛾�𝐵𝐵 its own𝑥𝑥𝑥𝑥 unique𝑥𝑥 𝑡𝑡 � precession frequency and this will cause each spin to accumulate phase, φ, relative to ω0 and proportional to
Δω(x,t):
( , ) = ( , ) (1.19)
∆𝜔𝜔 𝑥𝑥 𝑡𝑡 𝜔𝜔 𝑥𝑥 𝑡𝑡 − 𝜔𝜔0 ( , ) = ( , ) (1.20) 𝑡𝑡
𝜑𝜑 𝑥𝑥 𝑡𝑡 �0 ∆𝜔𝜔 𝑥𝑥 𝑡𝑡′ 𝑑𝑑𝑑𝑑′ Defining a variable, kx(t), to describe the phase accumulation due to the application of
Gx(t):
( , ) = ( , ) (1.21) 𝑡𝑡 ′ ′ 𝑘𝑘𝑥𝑥 𝑥𝑥 𝑡𝑡 � 𝛾𝛾𝐺𝐺𝑥𝑥 𝑥𝑥 𝑡𝑡 𝑑𝑑𝑡𝑡 The signal equation can be modified to incorporate0 this spatial encoding:
( ) = ( ) ( , ) (1.22) −𝑖𝑖𝑘𝑘𝑥𝑥 𝑥𝑥 𝑡𝑡 𝑥𝑥𝑥𝑥 𝑠𝑠 𝑡𝑡 �𝑉𝑉 𝑀𝑀 𝑥𝑥 𝑒𝑒 𝑑𝑑𝑑𝑑 In this way spins with different precession frequencies will contribute to the overall signal
and their frequency information is stored in s(t).
As Gx is applied spins will accumulate phase and kx will continue to get larger. We can imagine traversing a range of kx values from k-xmax to k+xmax (Figs. 1.6, 1.7) with the
signal at each kx representing the total phase accumulation across all of the spatially
varying precession frequencies. This provides spatial encoding in a single dimension and
the same process can be applied in two- and three-dimensions to give multi-dimensional
encoding. The data space with values for signal at different combinations of kx, ky, and kz
is known as k-space5. The underlying spatial frequency information that generated the k-
13
space data, s(kx,ky,kz), can be recovered via the FT with the resulting spectrum showing the relative amplitude of each precession frequency, known as the MRI image.
The classic scheme for traversing and filling k-space is a Cartesian pattern that directly samples k-space at evenly spaced grid points in each of the imaging dimensions6.
For two-dimensional imaging two encoding gradients, Gx and Gy, are combined in a
rectilinear fashion to generate a grid of sample points at every pair of kx and ky. This grid
of data can be converted directly to an image using the fast Fourier transform (FFT).
Cartesian imaging is widely applied because it is simple to implement, robust to system
imperfections, and fast to compute for image generation.
Instead of collecting points on a grid, points can be collected in any number of
patterns (called trajectories)7–10; most common of which are the spiral and radial
trajectories seen in Figure 1.5. Images can be generated from these non-uniformly sampled points in two ways: one is to distribute the data from the sampled points to a predefined grid through the process of regridding11,12, or two, applying the non-uniform FFT directly
to the sampled points13. Despite the added computational complexity, non-Cartesian
trajectories are popular because of their multiple advantages over Cartesian sampling14. As
a note, hybrid trajectories have been proposed that capitalize on the advantages of both
Cartesian and non-Cartesian sampling schemes15–17.
14
Figure 1.5: Example k-space trajectories. The Cartesian trajectory acquires k-space points on a rectilinear grid making it simple to acquire data and generate an image using the fast Fourier transform. Radial and spiral imaging collect samples along spokes or spiral arms and can be more challenging to implement. They also require additional computation steps to generate images but provide several advantages that make them appealing under different imaging conditions.
1.2.3 Relationship Between k-space and Image Space
Data in an MRI experiment is acquired in k-space and converted into images using the Fourier transform. When k-space is sampled at discrete points with a fixed spacing, Δk, the Nyquist-Shannon sampling theorem defines the bandwidth of the resulting FT frequency spectrum (the image) as:
= 1 (1.23)
For an MRI image, BW defines the range of precession𝐵𝐵𝐵𝐵 � ∆frequencies𝑘𝑘 included in the image.
Since the precession frequencies are distributed in space, this BW corresponds to a physical dimension, referred to as the image field of view (FOV), and describes the size of the image in a given direction (Fig. 1.6):
1 = (1.24)
𝑟𝑟 𝐹𝐹𝐹𝐹𝐹𝐹 𝑟𝑟 The range of precession frequencies found in an MRI�∆𝑘𝑘 experiment is limited by the amplitude of the encoding gradients (Equation 1.18) and size of the object being imaged.
This defines the BW of the image and sets the maximum spacing (minimum sampling rate)
15
between points, Δkmax, that will generate the smallest FOV containing the entire object
within the image. This is called the Nyquist sampling rate and is the rate needed to avoid
introducing aliasing into the signal.
Equation 1.24 demonstrates that the FOV can be increased or decreased by
changing the spacing between points in k-space. When Δk is equal to Δkmax, the image
FOV contains the entire range of precession frequencies, and therefore the entire object,
inside the image. If Δk is larger than Δkmax the FOV will not be large enough to contain the
entire object being imaged and will result in the highest precession frequencies (edges of
the object) aliasing back into the image. Conversely, Δk can be smaller than Δkmax
increasing the FOV so that it is larger than the object being imaged. It should be noted that
images are also described by their resolution, Δx, which is related to the number of points
acquired in k-space (Nk, Fig. 1.6):
1 = (1.25)
∆𝑥𝑥 𝑚𝑚𝑚𝑚𝑚𝑚 = �𝑘𝑘 × (1.26)
𝑚𝑚𝑚𝑚𝑚𝑚 𝑘𝑘 This states that for a fixed sampling𝑘𝑘 rate, 𝑁𝑁Δk, the∆𝑘𝑘 number of acquired points can be increased or decreased as needed to generate the desired image resolution.
16
Figure 1.6: Relationship between k-space sampling and image properties. Changing the kmax will alter the image resolution (Δx,y, Equation 1.25) and altering the Δk will change the image FOV (Equation 1.24).
While acquiring data with Δk equal to Δkmax ensures alias-free images, a Δk that is
larger than Δkmax can be used to acquire fewer data points and accelerate the image
acquisition process. This is referred to as undersampling, with the ratio of data between a
fully sampled and undersampled scan described by the undersampling or acceleration
factor, R. Undersampling both reduces scan time and increases aliasing proportionally with increasing R resulting in images acquired more quickly but increasingly corrupted by aliasing. Methods have been developed that use additional information to generate alias free images from highly undersampled acquisitions allowing dramatic reductions in imaging time without a loss of image quality18,19.
1.2.4 The MRI Pulse Sequence
RF excitation, slice selection, and spatial encoding are the three components that
comprise the imaging kernel and make up every MRI experiment. Their order and timing
during an experiment is referred to as the MRI pulse sequence and is described schematically in the sequence diagram (Fig. 1.7). This diagram defines two important
17
timing parameters for MRI: the repetition time (TR) and the echo time (TE). TR is the time
between subsequent RF pulses and usually determines when the sequence moves to the
next phase encoding step with the total image acquisition time, Tacq, defined as:
= × (1.27)
𝑎𝑎𝑎𝑎𝑎𝑎 𝑃𝑃𝑃𝑃 where NPE is the number of phase encoding𝑇𝑇 𝑁𝑁steps. TE𝑇𝑇𝑇𝑇 is the time between the midpoint of
the RF pulse and the formation of an echo for each phase encoding step. An echo refers to
the time when any gradient induced phase is completely refocused and the maximum
possible signal is available for that particular encoding step (Fig. 1.7).
The most basic pulse sequence that will generate an image is seen in Figure 1.7. It
consists of an RF pulse combined with a slice select gradient and two spatial encoding
gradients (phase and frequency encoding). This is called a gradient echo (GRE) sequence
because only gradients are used to refocus the magnetization and generate an echo20. If we add a second RF pulse to refocus any dephasing due to B0 inhomogeneity it becomes a spin echo (SE) sequence21 (Fig. 1.8). GRE imaging can be performed very quickly using small flip angles, short TE, and short TR to maximize the available signal. SE sequences have
more timing restrictions limiting their acceleration but are insensitive to field
inhomogeneities22. The properties of GRE and SE imagining typically results in GRE
23 images being either T1- or T2*-weighted while SE images can be either T1- or T2-
weighted.
18
Figure 1.7: Gradient echo pulse sequence and corresponding traversal of k-space. An RF pulse (with arbitrary flip angle α) is combined with a gradient (GS) to perform selective excitation and is followed by spatial frequency (GR) and phase (GPE) encoding. The gray GPE bar represents a single phase encoding step and this phase encoding step will result in moving around k-space as shown. Dashed lines in k-space represent movement without data acquisition and the solid line is when data is acquired. The echo time (TE) is when the frequency encoding gradient induced phase is refocused and the trajectory crosses kR = 0. This process is repeated using different phase encoding steps (GPE) for each repetition (TR). If TR ≤ T2 then the magnetization eventually approaches a steady state. Almost all MRI sequences are based on either gradient echo or spin echo sequences.
Different contrast within these sequences can be accomplished by adding/modifying the
RF pulses, adding/removing gradients, and changing the TE or TR. It is worth pointing out
an interesting state emerges if RF pulses are applied with only short (< T2) intervals
between them20,24. Provided no additional signal manipulation is performed, most spins
will have a non-zero value of Mxy at the end of a TR. This means the Mxy after the next RF
pulse will depend on both Mz and the remaining Mxy from the previous TR. Eventually, T1
recovery of Mz, T2 decay of Mxy, and the applied RF pulse reach a dynamic equilibrium so that the magnetization at each TE is equal to the one before it. At this point the magnetization is in a so-called steady state and is referred to as steady state with free
25–27 precession imaging (SSFP) . SSFP generates images with a combination of T1- and T2-
19
weighting and has become a very popular imaging method due to its soft-tissue contrast, rapid imaging times, and relatively high signal.
Figure 1.8: The spin echo pulse sequence. This is similar to the gradient echo sequence with the additional requirements of α=90° and a refocusing pulse with α=180° applied at a time, TE/2. This combination forces the TE to be twice the time between the 90° and 180° pulses for the refocusing pulse to rephase ΔB0 inhomogeneities eliminating the effect of T2′ on the signal.
1.2.5 The Signal-to-Noise Ratio
In MRI the quality of an image will be related to the amount of signal available. An important metric for evaluating the amount of signal and quality of the resulting images is the signal-to-noise ratio (SNR). SNR compares the intensity of the image signal to the signal variation due to background noise effectively describing the extent to which noise is apparent in an image. A high SNR will result in an image that clearly represents the data
20
of interest while an image with low SNR will have key features obscured by noise (Fig.
1.9). The SNR/voxel is quantitatively related to various imaging parameters4:
(1.28)
𝑎𝑎𝑎𝑎𝑎𝑎 𝑥𝑥 𝑦𝑦 𝑧𝑧 Equation 1.28 shows that the SNR𝑆𝑆𝑆𝑆𝑆𝑆 in⁄ 𝑣𝑣𝑣𝑣our𝑣𝑣𝑣𝑣 image𝑣𝑣 ∝ ∆ 𝑥𝑥is∆ directly𝑦𝑦∆𝑧𝑧�𝑁𝑁 related𝑁𝑁 𝑁𝑁 to𝑁𝑁 the∆𝑡𝑡 number of spins
within a voxel (image resolution dimensions Δx, Δy, Δz), the number of points acquired in
each k-space dimension (Nx, Ny, Nz), times the experiment is repeated (Nacq), and inversely
related to the data sampling rate (1/Δt).
Figure 1.9: Comparison of images with different SNR. The high SNR image clearly displays image features of the liver and associated anatomy while the low SNR has key features and tissue boundaries obscured by noise. In order preserve SNR in an MRI experiment, and by extension image quality, there
must be a balance between imaging time and spatial resolution. For example, doubling the
resolution in all dimensions would require a 64x increase in imaging time to maintain the
same SNR. This SNR limitation becomes problematic when imaging applications call for
high resolutions or rapid scan times. To account for this, high field scanners have been
developed to increase the SNR available for imaging. These stronger B0 fields (3.0 up to
7.0 Tesla for humans28 and from 7.0 up to and exceeding 11.7T29 for animals) create a
larger net magnetization vector (≈signal, Equations 1.1 and 1.7) that can be used to improve
21
the spatial resolution of the image without compromising the quality of the image. While
stronger magnetic fields have numerous benefits for MRI they unfortunately present their
own set of challenges30 (see Section 1.5.2).
1.3 Quantitative MRI
The majority of MR images are qualitative “weighted” images where the intensity
value of a pixel has no quantitative meaning and images are described in relative terms
prohibiting quantitative descriptions of the underlying tissue. Despite the prevalence of
weighted imaging in MRI, it is usually possible to directly estimate the values of the
underlying MRI-specific properties and describe disease with a quantitative MR-based
measurement. This provides a quantifiable basis for the identification of healthy and
diseased tissue enabling disease characterization and tracking of progression.
Conventional MRI quantification methods fit acquired data to a model of the
expected signal behavior with Equations 1.11 and 1.12 being examples of models used to
estimate T1 and T2, respectively (Fig. 1.10). To generate pixel-based maps of a specific
property, images weighted for the desired tissue property (such as T1- or T2-weighted
images) are acquired with different amounts of property weighting. Then, on a pixel-by-
pixel basis, the signal values in each pixel are plotted against the property weighting value
(left panel, Fig. 1.10). A curve is fit to the data by iteratively applying different parameter
combinations to the signal model31 until the parameter set that best describes the
experimental data is found (right panel, Fig. 1.10). The best fit is defined as the parameter
set that optimizes the chosen cost function in relation to the acquired data. Curve fitting is
powerful because it can have almost unlimited range and precision for estimating the parameters of interest but can be time-consuming to identify the best fits.
22
Figure 1.10: Example of T1 estimation using curve fitting. The left panel shows two T1-relaxation curves with different T1 values and added noise. These curves are sampled at known time points (NP=11) and a curve is fit to the sampled values using the appropriate model (Equation 1.11 for T1). This curve optimizes the cost function that measures the difference between the sample points and the fitted curve. After fitting the result is an estimate for the T1 value that generated the sampled points. One of the primary limitations to quantitative MRI methods in clinical and research
applications is their extended acquisition times. Imaging time in a quantitative MRI
experiment is related to Equation 1.27 and can be modified to include the number of
property weighting values acquired, NP:
= × × (1.29)
𝑎𝑎𝑎𝑎𝑎𝑎 𝑃𝑃𝑃𝑃 𝑃𝑃 The increase in imaging time is𝑇𝑇 proportional𝑁𝑁 𝑇𝑇to𝑇𝑇 the 𝑁𝑁number of points along the curve requiring a balance between acquisition time and parameter estimation accuracy. It can be
seen how accurate estimation of a single tissue property can be a time consuming process.
This becomes even more time intensive as more properties are estimated limiting the use
of quantitative MRI in clinical applications.
1.3.1 T1 and T2 Mapping
T1 and T2 relaxation times have long been investigated as potential quantitative
markers of disease and tissue structure32,33. Due to this potential value, strategies for
accelerating quantitative MRI have been some of the most active areas of MRI research.
23
Conventional methods for the estimation of T1 and T2 repeatedly sample the exponential
relaxation of the magnetization vector. For rapid estimation of T2 or T2*, this process is
straightforward simply requiring the acquisition of multiple spin or gradient echoes during a single transverse relaxation process after a 90-degree excitation pulse34,35. Rapid
36,37 38 estimation of T1 requires a preparation pulse, such as an inversion or saturation pulse,
to enable sampling of the magnetization vector as it undergoes T1 relaxation. The Look-
36–40 Locker method is a classic T1 mapping strategy that acquires images at multiple different weighting values during a single longitudinal relaxation process drastically reducing the quantification time. Alternative quantification methods use segmented acquisitions41, signal intensity ratios42,43, and magnetization preparation combined with
44,45 rapid imaging methods to accelerate quantification of T1 and T2.
Both T1 and T2 have shown value as quantitative markers of disease in a variety of
tissues16,46–51. The often complementary information provided by these measurements has
48,52–57 made simultaneous measurement of T1 and T2 a goal of quantitative MRI .
Simultaneous multi-property mapping provides several benefits over the mapping of each parameter individually in part due to using the same underlying images to generate both
the T1 and T2 property maps. This eliminates much of the uncertainty that arises from serial
quantification due to movement and physiological changes during in vivo imaging. This
attribute is particularly important in situations with rapidly changing states, such as
dynamic contrast enhanced MRI, where the nature of the signal may change substantially
over the course of a scan. Additionally, the resulting maps are inherently co-registered
allowing direct pixel-wise comparisons reducing variation introduced by region of interest
analysis. These quantitative multi-property techniques also create the opportunity for new
24
multi-parameter analyses that may have been difficult to impossible using serial acquisition
of single tissue properties.
1.4 Paramagnetic MRI Contrast Agents
Injection of exogenous MRI contrast agents has become a key feature of clinical and research MRI studies, improving sensitivity to disease58–60 through local magnetic interactions. These agents use a paramagnetic atom/molecule to modulate the T1 and T2
relaxation time constants of the local protons resulting in dynamic changes in the image signal and contrast. A detailed discussion of contrast agent physics is beyond the scope of this work and can be found elsewhere61,62, but it is essential for this work to understand the
basic relationship between MRI contrast agents and T1 and T2 relaxation rates.
1.4.1 MR Effects of Exogenous MRI Contrast Agents
When the paramagnetic ion in an MRI contrast agent interacts with water protons,
the T1 and T2 of the protons change according to established concentration-dependent
relaxivity equations61,63:
1 1 = + [ ] (1.30)
1 �𝑇𝑇1 �𝑇𝑇10 𝑟𝑟 𝐴𝐴 1 = 1 + [ ] (1.31)
2 � 2 � 20 𝑟𝑟 𝐴𝐴 In these equations T10 and T20 are the𝑇𝑇 T1 and𝑇𝑇 T2 values in the absence of the contrast agent
(pre-contrast), r1 and r2 are the relaxivity properties of the contrast agent for T1 and T2
respectively, and [A] is the concentration of the agent. These equations allow prediction of
the observed T1 and T2 in the presence of a contrast agent (post-contrast) as a function of
concentration. The r1 and r2 are physical properties of the agent, [A], that describe the
relative efficiency with which an agent modifies T1 or T2 and their ratio provides insight to
25
the relative changes that will be seen in the resulting images. As a general rule, agents with
a small r2/r1 are used as so-called positive contrast T1 agents because the contrast
enhancement is pronounced in T1-weighted images. Alternatively, agents with a large r2/r1
are typically used as negative contrast T2 agents due to the change in contrast seen on T2-
64 weighted images . As seen in Equations 1.30 and 1.31, all contrast agents impact both T1
and T2. Therefore, while a specific agent may predominantly impact one relaxation time
(e.g., T1), there will still be a corresponding change in the other relaxation time (T2).
1.4.2 Relaxivity Constants: r1 and r2
Each MRI contrast agent has a unique pair of relaxivity values, r1 and r2, making
their estimation crucial for understanding and predicting the agent’s effect on image contrast. Estimation is typically done by creating solutions with different known agent concentrations and then measuring T1 and T2 for each of the concentrations. A plot of R1
(1/T1) and R2 (1/T2) versus concentration is created and a linear fit to Equations 1.30 and
1.31 yields the r1 and r2 value for that particular contrast agent. These in vitro estimates of r1 and r2 are used to characterize the agent as either a T1 or T2 agents and serve as a guide
for development and optimization of new agents.
The caveat when estimating r1 and r2 is that the values for relaxivity are sensitive
to several experimental factors65 including the magnetic field strength66, temperature64,67,
the molecular structure of the chelation molecule68–70, and macromolecular interactions71–
73. These effects interact in complex ways making prediction of relaxivity and image
enhancement under a new set of experimental conditions challenging. This is particularly
important for in vivo imaging where there may be substantial differences between in vitro
and in vivo relaxivity due to temperature, protein binding, and other micro-environmental
26
changes (pH, enzyme activity, etc.)74–77. Despite the variable effects on relaxivity in vivo,
in vitro estimation of these parameters remains an essential tool for design and selection of
MRI contrast agents67,68,78–81.
A detailed discussion of the chemistry and interactions of MRI contrast agents is
beyond the scope of this introduction. Excellent reviews and discussions of agent design
and the factors that determine relaxivity can be found in reviews61,63,82 and several works review the factors that control relaxivity under experimental conditions62,83.
1.4.3 Common (and Uncommon) MRI Contrast Agents
The most commonly used MRI contrast agents come in two flavors: gadolinium
based agents and iron based agents. Gadolinium based contrast agents (GCAs) are typically
small molecules that have a Gd(III) ion chelated with a carbon backbone designed to
improve stability, reduce toxicity, and control clearance of the gadolinium ion83. Iron agents combine super-paramagnetic iron oxide (SPIO) particles with larger nano-structures aimed at reducing clearance and controlling aggregation of the particles. The reason for these two main flavors has to do with the differences in the r2/r1 ratio between GCAs and
SPIOs. GCAs typically have in vitro r2/r1 ratios of 1-2 classifying GCAs as positive contrast agents with increased signal on T1-weighted imaging. Conversely, SPIOs can have r2/r1
ratios >100 making them negative contrast agents due to signal dropout from excessive signal decay64. Due to their complementary nature, these two classes of MRI contrast agents have provided the vast majority of contrast enhanced clinical and research scans64,84,85.
The contrast agents regularly used clinically and in research rely on passive
methods of accumulation in tissues to generate contrast enhancement. This leads to
27
relatively inefficient and non-specific tissue enhancement requiring large agent doses while
providing little additional information about the tissue itself. As a result, research has
attempted to improve the capabilities of MRI contrast agents by developing more efficient
65,79,80,86 81,87–92 agents with improved r2/r1 ratios , specific disease targeting , and
variable/activatable relaxivities dependent on the local tissue microenvironment93,94.
Researchers have explored replacing the Gd ion in GCAs with other lanthanide ions such
as Dy68,81, Ho95,96, or Eu97–100 to modulate the relaxivity while maintaining a similar
pharmacokinetic profile101. Similarly, SPIOs can be formulated with different sized oxide
102–104 cores resulting in different r2/r1 ratios . Other agents have incorporated highly specific
targeting moieties to identify disease specific proteins, cell surface receptors, and other disease associated molecular species78,81,89,90,92,105. Agents have even been developed that
are sensitive to pH80,100,106–108, enzyme activity109–111, and oxidation state94,98.
These novel agents have the potential to outperform current agents by generating
tissue enhancement with smaller doses, sensitively detecting and identifying disease, and
providing information about the tumor microenvironment. Each of these agents provides a
unique piece of information about the disease, and situations could be envisioned where
using more than a single agent would provide valuable information. The combination of a
targeted agent and an agent designed to identify hypoxia could be co-administered to
specifically detect the presence of disease and then understand the oxygenation state of different disease areas. Alternatively, agents could be combined to evaluate drug delivery and therapeutic effect potentially allowing early assessment of response to treatment.
Unfortunately, these unique agents are restricted to single agent preclinical applications
limiting their benefit to patients.
28
1.5 Preclinical MRI at High Fields
Although MRI research is considered safe in human subjects, preclinical MRI has
become a valuable research tool primarily due to the opportunity for researchers to perform
unique studies not possible in human subjects. Animal MRI studies can be performed under tightly controlled experimental conditions using well understood animal models of disease and confirmed with gold-standard validation methods47,74,78,81,112–116. Quantitative MRI measurements can be associated with changes in tissue structure117, protein levels47, and
genetic expression90 enabling a better understanding of the biological mechanisms causing
the changes observed with MRI. Measurements made with MRI are also used as a non- invasive means for scientists to evaluate new therapeutic interventions or monitor disease progression in these animal models114,118–120. These preclinical MRI studies have helped
establish the underlying pathophysiologic processes that result in changes to MRI tissue
properties and aided in the fundamental understanding of different diseases.
1.5.1 Animal Challenges
Preclinical MRI has substantial value to the research community, but imaging of
animals creates an SNR limitation that must be accounted for. Animals are many orders of
magnitude smaller than humans requiring higher image resolutions to resolve key features
in preclinical MRI images. This reduction in SNR is highlighted by comparing the
resolution in a typical human experiment (1 × 1mm in plane resolution) to the resolution
often needed to image a mouse (0.1 × 0.1 mm in plane). According to Equation 1.28 this
will result in the SNR/voxel being reduced by a factor of 100 simply due to the change in
resolution. To account for this animals are often imaged using larger B0 fields and special
29
RF coils designed to maximize available signal but preclinical MRI, particularly in smaller
animals such as mice, is still limited by SNR.
In addition to the SNR issues limitations, the physiology of animals presents unique challenges. While anesthetized in the scanner animals still experience rapid, periodic motion due to breathing (~60 breaths/minute) and cardiac pulsatility (>200 beats/minute)38,113,115. The motion during an MR scan causes shifts in the k-space data
resulting in ghosting artifacts that can severely degrade the image quality (Fig. 1.11).
Motion is also problematic in high resolution images because even small movements result
in structures shifting multiple pixels causing image blurring and mis-registration between
scans. It is possible to acquire data during the quiescent period of motion using gating but
this extends scan times in the already typically longer preclinical experiments17,55,121.
Gating also interferes with quantitative imaging which often requires measurements at
known time intervals. Several methods have been proposed to correct motion artifacts, such
as retrospective motion correction122 or oversampling the center of k-space123, but these
require precise monitoring or advanced processing for accurate correction of artifacts.
Figure 1.11: Motion artifacts in preclinical MRI. The image on the left was acquired using respiratory gating which results in a reduction in motion artifacts compared to the ungated image on the right. Gating increased imaging time by a factor of 2.5 substantially increasing the image acquisition time.
30
1.5.2 Hardware Challenges
The high fields and stronger gradients necessary in preclinical imaging have several
30 disadvantages that make imaging challenging . Larger B0 fields are an important way of
increasing the available signal but they exacerbate the effect of ΔB0 proportionally with
the value of the B0 field. This shortens T2* (Equation 1.16) leading to excessive signal
decay making short echo times extremely important in high-field T2*-sensitive gradient
echo imaging. Methods that minimize the time between RF excitation and the echo time124,
such as ultra-short echo time imaging55, are valuable in preclinical MRI for preserving
signal and observation of short T2/T2* tissues.
The effect of ΔB0 also impacts the ability of preclinical researchers to use certain imaging methods. Balanced SSFP imaging is an incredibly efficient imaging method generating very high SNR/unit time125,126 but suffers from areas of low SNR (bands) due
to ΔB0 effects. On high field systems with substantial field inhomogeneities, bSSFP
imaging experiences intense banding artifacts often obscuring the anatomy of interest. One
solution has been to acquire multiple images in different states and combine them to create
a band-free image127. The drawback to this is imaging time increases proportionally with
the number of image states to be acquired. Banding can be eliminated by using a spoiling
gradient to generate images without sensitivity to ΔB0 but these images overall have a
lower SNR.
The imaging gradients used in preclinical imaging are often much stronger than
those used on human scanners (400 mT/m vs. 40 mT/m) increasing the presence of gradient induced artifacts. Rapidly turning the gradients on and off during imaging creates time-
varying magnetic fields inducing eddy currents that cause spins to accumulate unwanted
31
phase. This phase results in image artifacts and is best visualized via the N/2 ghost that is
generated during EPI imaging128. In non-Cartesian imaging these eddy current induced
phase accumulations can degrade the image quality without causing obvious artifacts and
are worse with increasing gradient strength129. These challenges require balancing high
resolution imaging with generating high quality, artifact-free images using powerful
gradients.
Decreased homogeneity of the B1 field is another disadvantage faced by high-field
MRI. Typically, MRI experiments assume the RF excitation field is homogeneous across
a sample with each pixel experiencing the same flip angle. This is frequently not the case
on human scanners and the increased field strengths used in preclinical imaging make
130 generation of a homogenous B1 field even more difficult subsequently increasing spatial variations in signal. Non-uniform B1 fields can be detrimental to quantitative MRI methods since these methods often rely on accurate excitation angles to estimate MR tissue properties. This dependence has led to the development of imaging methods either
131 designed to be insensitive to B1 or which measure the B1 field and correct for field
inhomogeneity132–134. Parallel transmission is a promising new method for creating a homogeneous B1 field through the combined use of multiple independent RF excitation
fields135.
1.6 Dissertation Overview
The preceding chapter provided an overview of the concepts necessary to
understand the work presented in the following chapters. Chapter 2 will introduce the
Magnetic Resonance Fingerprinting (MRF) method in a separate chapter since this serves
as a core aspect of this work. The general acquisition framework will be introduced along
32
with a discussion of how to generate sensitivity to different tissue properties. That chapter
will also discuss how to accelerate the MRF acquisition and applications of preclinical MR fingerprinting. Detailed derivations and descriptions were deliberately left out of these introductory chapters. For more detailed reviews the reader is referred to the references for review articles and books on MRI physics.
After Chapters 1 and 2, work will be presented in Chapter 3 on the use of acquisition
schemes for the suppression of motion artifacts in preclinical MRF. Chapter 4 will detail
the development of the Dual Contrast – Magnetic Resonance Fingerprinting method and
how it enables the quantification of multiple MRI contrast agents within a single scan. A
model extending Equations 1.30 and 1.31 will be validated using in vitro validation
experiments in Chapter 4. The dissertation will close with a discussion of future directions
and conclusions from this work.
33
Chapter 2: Fundamentals of Magnetic Resonance Fingerprinting
Magnetic Resonance Fingerprinting (MRF) was introduced by Ma et al.136 as a method for simultaneous quantification of the MRI-specific tissue properties T1 and T2,
off-resonance frequency (ΔB0), and proton density (M0). The original work used time- varying MRI acquisition sequence parameters, a dictionary-based template matching scheme, and a highly undersampled k-space trajectory to generate accurate in vivo property
maps in less than fifteen seconds. Since this initial report, the method has been extended to
high-field preclinical scanners by Gao et al.137 and is experiencing rapid development138–
140. This chapter will describe the basic framework of the MRF method for general
implementation followed by a more detailed discussion of property sensitivity, temporal
acceleration, and MRF in high-field and preclinical applications.
2.1 Overview of MR Fingerprinting
During an MRF experiment acquired image data is compared to a predictive model in order to quantify unknown MRI-specific properties. Key features of the MRF method are the data acquisition scheme and quantification by comparing the acquired data to a
“matched” dictionary of simulated data. This dictionary-based quantification creates a link between the simulated and acquired data allowing assignment of the known property values from the dictionary to the unknown property values of the acquired data. In this way MRF maps each property of interest generating a quantitative representation of the underlying
MRI data.
2.1.1 MRF Image Acquisition
The experimental data in an MRF experiment consists of a large number, n
(typically >500), of dynamically acquired MRI images. Each image is collected with a
34
fixed imaging kernel and a different combination of acquisition sequence parameters (e.g., repetition time (TR), flip angle (FA); Fig. 2.1) resulting in a set of acquired images with
variable contrast (Fig. 2.2). In order to better understand the changing contrast, the signal
from a given voxel (volume element) can be plotted over the entire MRF image set
revealing the signal evolution profile for that voxel. These voxel-based evolution profiles depict the coherent buildup of signal over time and are the fundamental component of the
MRF experiment.
Figure 2.1: Example schematic of an MRF pulse sequence. The top panel shows the MRF acquisition sequence with each image acquired using a unique combination of acquisition parameters (TR and FA in this example) and a constant imaging kernel (gradient spoiled SSFP). Example TR and FA patterns are shown in the bottom panel with each combination (1 to n) being used to acquire a single image (n=1024). Additionally, in this sequence an inversion preparation is used to enhance sensitivity to T1. The entire pattern is repeated so that all phase encoding steps are acquired using each of the 1024 combinations of TR and FA. A similar figure is presented in Figure 3.2.
The first step in designing an MRF imaging sequence is to select an imaging kernel
for data acquisition. This imaging kernel controls the inherent sensitivity of the signal to
different MRI-specific properties (e.g., T1, T2, diffusion) with the resulting MRF signal
evolution profiles being a function of these properties. Changing the structure of the
imaging kernel will modify the sensitivity of an MRF sequence and substantial work has 35
been done to understand the sensitivity of different imaging kernels and sequences131,141–
143. One example of how the sensitivity of an imaging kernel can be modified is the addition
of a spoiling gradient to eliminate the sensitivity of steady state with free precession
144 145 imaging (SSFP) to ΔB0 while increasing sensitivity to T2* . Further, the position of this spoiling gradient in the sequence can increase or decrease the sensitivity to diffusion146,147 demonstrating the ability of the imaging kernel to determine property sensitivity during an MRF experiment.
After the imaging kernel has been selected, numerous degrees of freedom can be
exploited when choosing which acquisition sequence parameters to vary during an MRF
experiment. Repetition time and flip angle were already mentioned as potential variable
parameters but, in theory, almost any aspect of the sequence could be varied. Yet, the
choice of variable acquisition parameters is restricted by both MRI hardware and
simulation limitations. In practice, the user attempts to vary parameters that will improve
MRF’s ability to simultaneously estimate the MRI properties of interest54,145,148. For
example, in SSFP imaging the SNR and contrast-to-noise ratio depend on the T1, T2, and
FA with different amounts of T1 and T2 weighting based on the applied FA, particularly in
the transient state26. This signal weighting dependence on FA is a primary reason why it is
often varied during MRF experiments that utilize an SSFP imaging kernel. One unsolved
question is determining the optimal parameter variation patterns that will lead to the most
efficient estimation of parameters of interest. Current methods rely on an understanding of
the MRF imaging sequence, simulation studies, and empirical evidence to identify optimal
patterns of variation but work is being done to identify metrics that can be used to determine
the optimal imaging parameters for MRF quantification139.
36
It is important to mention that magnetization preparation148,149 can be used to
enhance/modify the property sensitivity in an MRF experiment. A simple example is the
addition of an inversion pulse at the beginning of an MRF acquisition that inverts the
magnetization vector from +M0 to –M0. After inversion, the magnetization vector will relax
150 back towards its equilibrium state of +M0 using predominantly T1 relaxation . This
inversion recovery process is sampled during the MRF experiment creating a component
of the profile that is heavily dependent on T1 resulting in improved estimation of T1. A
149,151 152 similar process can be done using T2 or diffusion preparation blocks to generate
signal differences during the preparation that evolve over the course of the MRF
experiment. Preparation blocks can be used to enhance the inherent property sensitivity of the imaging kernel or add new sensitivity to an MRF experiment.
2.1.2 Creation of the MRF Dictionary
MRF quantification utilizes a dictionary-based matching scheme with the dictionary serving as a database of the foreseeable MRF signal evolution profiles that could be generated by the MRF imaging sequence being implemented. The dictionary is populated by ideal, theoretical signal evolution profiles simulated using the established
Bloch equations1, the known variation of MRI acquisition sequence parameters (e.g., FA,
TR, etc.), and a predetermined set of possible MRI-specific property combinations (e.g.,
T1, T2, ΔB0). The dictionary simulation is designed to recreate the actual MRF experiment
so that the dictionary entries “match” the acquired profiles. However, some reasonable
simplifications are made regarding performance of the applied radiofrequency pulses,
magnetic field gradients, diffusion, and other effects in order to reduce simulation complexity and time. Additionally, the MRF dictionary should contain profiles spanning
37 the entire scope of potentially acquired profiles, but the range and number of property combinations included in the dictionary is often restricted to maintain reasonable dictionary generation and quantification times.
Under certain experimental conditions the dictionary simulation assumptions will not accurately reflect the actual acquisition parameters used during an MRF experiment.
This situation commonly occurs when there is imperfect application of the MRI acquisition sequence parameters. Typically, the dictionary simulation assumes the entire sample being imaged experiences uniform application of the MRI acquisition sequence parameters.
However, the B1 field and gradient performance often deviate from ideal
138,153–156 performance . The B1 field suffers from spatial heterogeneity resulting in regional deviations between the actual and nominal flip angle reducing the accuracy of the dictionary simulation. Additionally, variable gradient performance can introduce image artifacts, distortions, and unwanted gradient spoiling which can also cause the acquired signal evolution profiles to deviate from the ideal simulated profiles. Fortunately, hardware issues can be assessed prior to the experiment and often accounted for during either dictionary creation or image reconstruction. This has reduced the impact of hardware imperfections on the quantitative accuracy of MRF but underscores the need to understand the performance of the MR system being utilized.
When designing the simulation that will be used to populate the MRF dictionary, only a subset of MRI-specific tissue properties is included in the dictionary simulation. The selection of properties is made based on the known sensitivity of the MRF imaging sequence with all other properties assumed to have zero effect on the signal evolution (e.g.,
T1, T2 included, diffusion assumed to be 0). This allows reduction in simulation complexity
38
and time but may reduce the accuracy of the simulation. For example, MRI sequences are
typically very sensitive to the underlying T1 and T2, but can be sensitive to diffusion under
147 certain circumstances . If diffusion is assumed to be negligible and only T1 and T2 are
explicitly included in the dictionary simulation, any diffusion effects will cause signal
attenuation similar to T2 decay likely resulting in an “apparent” T2 estimate that is shorter
157 than the native T2 . It is therefore important to keep in mind the sequence sensitivity and
dictionary simulation assumptions when analyzing MRF measurements. Fortunately, when
chosen intelligently, these simulation assumptions introduce only negligible differences
between the simulated and acquired data and quantitative accuracy is preserved.
In addition to using a subset of properties in the simulation, only a finite number of
property values can be used to simulate entries in the dictionary. This restricts the possible
values in the property maps to the limited and discrete set of property values used during
generation of the MRF dictionary. Ideally, the range of values in the dictionary will be
large enough to account for unexpected results but the resolution fine enough to identify
differences between tissue and disease. Property values can be added to the dictionary to
increase the dictionary range or resolution, but each additional value generates a new
dictionary entry and increases the dictionary size, simulation, and quantification time. This
requires balancing quantification and computation time when designing the MRF
dictionary. Several methods have been proposed to overcome the computation
limitations158–161 allowing creation of dictionaries with increased range and resolution without a corresponding increase in computation time.
39
2.1.3 Quantification via Dictionary Matching
The final step in an MRF experiment is estimating the MRI-specific properties of
interest and generating maps of their values. In MRF this is done by individually comparing each voxel’s acquired MRF signal evolution profile to the dictionary and finding the
dictionary entry that best matches the acquired profile. This process is referred to as
“matching” and identifies a best match by optimizing a cost function evaluated between the acquired profile and the dictionary profile. Once the match is made, the known combination of MRI properties used to simulate the matched dictionary entry (i.e., T1 and
T2 relaxation times) are assigned to the voxel being investigated providing an estimate of
the MRI properties that generated the acquired profile. This process is repeated for all
voxels of interest resulting in a set of inherently co-registered MRI property maps (depicted
schematically in Fig. 2.2).
Figure 2.2: Schematic describing the MRF quantification method. An MRF experiment begins by acquiring a series of images with variable contrast (Figure 2.1). Then, for a given voxel of interest, the signal from that voxel is taken from all the images to generate an acquired MRF profile. This profile is compared to the dictionary profiles which were simulated using the Bloch equations designed with sensitivity to the tissue properties of interest (T1 and T2 in this example). Identification of a “best match” allows assignment of the T1 and T2 used to simulate the best
40
matched profile to the voxel that generated the acquire profile. This is repeated for all voxels of interest to create maps of each property. The acquired images are often corrupted by artifacts but if the MRF sequence is implemented properly the resulting maps should be free of the artifacts in the underlying images. A beneficial feature of the MRF method is that the quality of the MRF maps does
not necessarily depend on the quality of the images used to generate the map (Fig. 2.2).
Unwanted features present in the acquired images, such as noise, aliasing, and other
artifacts, are suppressed in the resulting MRF maps as long as they are incoherent in
time154. Since the dictionary simulation does not include noise or aliasing, these incoherent components of the acquired profile are effectively ignored and a match is identified based on the coherent signal evolution. This creates the opportunity to acquire very low-quality
images yet still produce artifact-free maps provided the coherent parts of the signal are
given sufficient time to evolve and generate sensitivity to the properties of interest.
2.2 Reduction of MRF Acquisition Time
Despite the use of fast imaging methods capable of acquiring a single image in a
matter of seconds, fully sampling all of the hundreds to thousands of images in an MRF
experiment can take several minutes for a single slice. For example, an MRF experiment
that acquired 64 phase encoding steps (PE) for n=500 images using an average TR of 7.5
ms would require 4 minutes of imaging time:
= × ( ) (2.1) 𝑛𝑛 𝑎𝑎𝑎𝑎𝑎𝑎 𝑇𝑇 𝑃𝑃𝑃𝑃 �1𝑇𝑇𝑇𝑇 𝑛𝑛 This relatively poor temporal resolution limits the use of MRF in clinical and dynamic
studies where short imaging times are necessary.
41
Equation 2.1 states, that to improve the temporal resolution of an MRF experiment
either faster imaging methods must be used (reducing TR) or less data should be acquired
(fewer PE or n). While either of these strategies could be used to reduce the imaging time,
the choice is not necessarily straight forward due to the implications of each option.
Reducing the number of images or shortening the TR may not allow the MRF signal to
evolve sufficiently to generate sensitivity to the desired tissue properties, while acquiring fewer PE lines introduces aliasing into the underlying MRF images reducing image quality.
Reducing the total MRF acquisition time without compromising the quantitative accuracy of the MRF method requires careful balancing of each of these effects.
2.2.1 Acceleration of the MRF Acquisition
Conventional MRF improves temporal resolution by sampling fewer PE lines instead of changing the TR or number of images. This choice sacrifices underlying image quality for reduced scan time while still allowing the coherent buildup of signal that is necessary to generate sensitivity to multiple MRI-specific properties. Image quality is typically sacrificed in favor of a prolonged signal evolution because MRF is primarily concerned with the resulting quantitative maps and not the underlying images. In this situation controlling the appearance of the aliasing artifacts in the images is important for the MRF quantification process (Fig. 2.3). The goal of undersampling in an MRF experiment is to create image aliasing artifacts that are incoherent in time and then suppressed in the resulting MRF maps. Non-Cartesian sampling is frequently used in MRF imaging because the aliasing artifacts appear noise-like and can be modified by changing the sampling pattern162 creating a simple way to generate time varying aliasing artifacts without substantial modification to the MRF acquisition sequence.
42
Figure 2.3: Aliasing artifacts at different acceleration factors. The top image is a fully sampled image for reference. Aliasing artifacts appear differently depending on the k-space sampling trajectory (spiral vs. Cartesian) and the amount of undersampling (R=2, 4, or 8). Additionally, different aliasing patterns will arise depending on the k-space lines used for the reconstruction. Patterns 1 and 2 for the R=8 images use different subsets of the k-space data for reconstruction. Other methods have been proposed to accelerate image acquisition and still generate accurate property maps. These techniques rely on advanced reconstruction methods to generate un-aliased images for quantification and include view sharing153, reconstruction based methods to “denoise” the images163, and parallel imaging164. MRF methods that utilize advanced image reconstruction techniques allow alternative k-space sampling patterns that have less desirable artifact patterns, such as Cartesian sampling, to be used for MRF experiments. This is beneficial in situations where non-Cartesian k-space trajectories may be unavailable or difficult to implement. While current methods enable acceleration factors of R=48 and higher, image reconstruction and quantification remain an active area of research.
43
2.3 High-Field Preclinical Magnetic Resonance Fingerprinting
The MRF method was initially developed and validated in humans on clinical
scanners but has been extended to high-field preclinical scanners for use in animals137.
165 Unfortunately, the issues of B0 field inhomogeneity, periodic physiological motion , and
powerful gradients that corrupt preclinical images persist as issues for preclinical MRF.
Large ΔB0 values cause T2* decay that leaves very little to no signal available resulting in acquired profiles that are dominated by noise. Although the MRF method is relatively insensitive to noise, there still must be some signal available for matching and quantification. Additionally, if T2* is short compared to TR, an SSFP sequence may lose
sensitivity to T2. This can make MRF-based quantification challenging in areas with low
field homogeneity or large differences in susceptibility (i.e. lung, air-tissue interfaces).
The physics of imaging at high fields also changes the dictionary simulation
assumptions that can be made during a high-field preclinical MRF experiment. Signal effects that are minimal in human studies at human field strengths (≤ 3.0 T) routinely become non-negligible during high-field preclinical studies. For example, the B1 field is
increasingly inhomogeneous at high fields making uniform excitation over large areas
particularly challenging130. In addition, the strong gradients required for high resolution
imaging in animals (400 mT/m vs. 40 mT/m) may deviate more from ideal performance
(increased eddy currents) reducing both image and map quality. These gradients can also
dramatically increase the amount of diffusion weighting in an image potentially requiring
incorporation of diffusion into the MRF dictionary simulation. Fortunately, as described
previously, these issues can be assessed prior to performing the MRF experiment and their
effect minimized through appropriate application of MRI acquisition sequence parameters
44 and/or compensation in the MRF dictionary. In general, methods developed for application on human scanners can only be translated to high-field preclinical applications after careful consideration of the differences between the clinical and preclinical environments.
Preclinical MRF-based quantification faces a unique challenge due to the nature of motion artifacts generated by animals. During an MRF scan, animals are anesthetized but they breath continuously at a rate of ~1 Hz and have heart rates of ~200 beats per minute.
This motion is not only periodic, but also relatively rapid creating a high likelihood that artifacts will appear similarly across a large number of images. Under the right conditions these artifacts may manifest as coherent deviations in the acquired MRF profile biasing the matching process and resulting in maps that are corrupted by motion artifacts. Options for correcting/eliminating these motion artifacts are limited due to the fact that MRF requires the simulation to exactly match the acquisition. Any motion correction or gating would need to be incorporated in to the dictionary simulation requiring creation of a new dictionary for each unique motion pattern149. This is further complicated by the need for precise physiological monitoring which can be difficult on high-field preclinical scanners.
Overall, this motion creates a unique challenge requiring novel methods to eliminate their effects on MRF-based measurements.
2.4 Summary
Magnetic Resonance Fingerprinting is a robust method for rapid quantification of multiple MRI-specific tissue properties. Through the dynamic application of variable MRI acquisition parameters, unique magnetization evolution profiles are generated that are a function of the properties of interest. Acquired profiles are matched to a dictionary of simulated profiles with the best matched profile revealing the properties of the acquired
45 profile. Quantification errors arise when the simulation does not match the performance of the scanner or the signal is sensitive to properties not incorporated into the dictionary simulation. MRF relies on the coherent buildup of magnetization and is therefore highly amenable to radical undersampling allowing for rapid acquisition and quantification.
Preclinical MRF faces many challenges inherent to imaging animals on high fields but these can be overcome through design of the MRF experiment.
46
Chapter 3: Regularly Incremented Phase Encoding – MR Fingerprinting
3.1 Introduction
Quantitative preclinical MRI methods can be used to provide objective assessments
of animal models of human disease for comparison with gold standard histological
assessments47,112,113,118,165,166. Unfortunately, quantitative imaging in small animals can be
challenging due to anesthetized respiration rates of 40 to 80 breaths per minute and heart
rates of >200 beats per minute causing blurring and ghosting artifacts that can result in
quantification errors38,47,113–115,167,168. This is complicated because prospective respiratory
and/or cardiac gating to reduce the motion artifacts31,38,121,169,170 may not be practical because it substantially can extend acquisition times17. Intubation and significant motion
restriction115,171 also can help reduce motion artifacts, but may cause increased mortality in animal models with advanced disease. As a result, preclinical respiratory and cardiovascular motion remains a significant challenge for accurate quantification in preclinical MRI studies.
MR Fingerprinting (MRF)136,144 is a quantitative MRI framework that has demonstrated unique properties including insensitivity to patient movement136 and respiratory motion in preclinical imaging studies137. However, this resistance to respiratory
motion only was observed for low to moderate respiration rates. The same preclinical MRF
study also showed that Cartesian MRF is susceptible to artifacts due to cardiac pulsatility.
In this work, we propose the regularly incremented phase-encoding (RIPE)-MRF
methodology172 for use on high field preclinical MRI scanners to suppress both respiratory
and cardiac pulsatility motion artifacts. This new approach utilizes an incremented view
ordering of the k-space phase-encoding, within a fully-sampled Cartesian MRF trajectory,
47
to reduce the impact of motion on the MRF-based T1 and T2 maps. The intention of the
RIPE-MRF view ordering is to add temporal incoherence to the motion artifacts using altered k-space trajectories similar to the temporal incoherence instituted in highly
undersampled non-Cartesian MRF149,154,159,173. Herein, we compare the quantitative and
motion suppression capabilities of this new RIPE-MRF technique to the previously
described in vivo preclinical MRF methodology137.
3.2 Methods
3.2.1 RIPE-MRF Encoding
For this study, two Cartesian MRF methods were implemented on a Bruker Biospec
7 T MRI Scanner (Bruker Inc., Billerica, Massachusetts, USA). The previously described
preclinical MRF method137 will be referred to as standard Cartesian MRF (SC-MRF). The
RIPE-MRF method is derived from the SC-MRF method but has modified the acquisition
order of the phase-encoding lines, as shown in Figure 3.1. The SC-MRF method (Fig. 3.1a)
acquires the same phase-encoding line during each variable repetition time for an entire set
of dynamic MRF images. After acquiring the entire set of MRF data for this single phase-
encoding line, the phase-encoding line is incremented and another set of dynamic MRF
data is acquired. This process is repeated for all phase-encoding lines, starting with the
edge line of k-space and continuing in a sequential fashion. In contrast, the RIPE-MRF
method (Fig. 3.1b) linearly increments the acquired k-space line for each repetition time
throughout the entire set of dynamic MRF images. The result of this phase-encoding
variation is a shuffled set of fully sampled Cartesian k-space time points that can be
reordered to enable image reconstruction and quantification.
48
Figure 3.1: Schematic representation of two different MRF acquisition schemes. (a) SC‐MRF and (b) RIPE‐MRF phase‐encoding schema showing four (out of 1,024) MRF k‐space time points. The red line in each time point dataset is acquired during the first set of dynamic MRF images (1,024 total k‐space lines). The blue line is the phase‐encoding line acquired during the second set of dynamic MRF images (second set of 1,024 k‐space lines). The green line is the third set of acquired MRF phase‐encoding lines. SC‐MRF acquires the same phase‐encoding line during each set of dynamic MRF images, whereas the RIPE‐MRF strategy increments the acquired phase‐encoding line during each set of dynamic MRF images to provide temporal incoherence for respiratory and pulsatile motion artifacts.
3.2.2 MRF Design and Quantification
Both the SC-MRF and RIPE-MRF acquisitions used a fast imaging with steady- state free precession imaging kernel137,144 with patterns of flip angles and repetition times
based on the original MRF method (Ma et. al., Nature 2013136, and Fig. 3.2). Following the
MRF acquisitions, all imaging data were exported to MATLAB (MathWorks, Natick,
Massachusetts, USA) for analysis. Quantification was performed by matching the acquired profiles on a pixel-by-pixel basis to a dictionary of simulated profiles from all logical combinations of longitudinal relaxation time (T1) and transverse relaxation time (T2). For
each pixel the maximum inner product between the acquired profile and the individual
145 dictionary entry yielded the matched T1 and T2 values . In addition, proton density (M0) was estimated as a scale factor between the acquired data and simulated dictionary profile136. More details on the MRF sequence, dictionary, and matching process are described in the Supporting Methods in Section 3.6. In vitro MRF experiments also were
49
performed to compare the accuracy of the T1 and T2 estimates from SC-MRF and RIPE-
MRF with conventional spin echo MRI assessments (details in Supporting Methods,
Section 3.6).
Figure 3.2: Schematic of MRF pulse sequence with TR and FA patterns. The top panel demonstrates the inversion preparation combined with variable flip angles and repetition times played out until FAn and TRn (n=1,024). The slice gradient is unbalanced resulting in the dephasing needed to perform gradient spoiled fast imaging with steady state free precession imaging. This pattern is repeated for all lines of k‐space changing the acquired phase encoding line based on the method implemented (SC‐MRF or RIPE‐MRF; Fig. 3.1). For each TR the phase encoding is perfectly balanced allowing for an arbitrary phase encoding order to be applied.
3.2.3 In vivo RIPE-MRF
All studies were conducted according to protocols approved by the Case Western
Reserve University Institutional Animal Care and Use Committee. In vivo single-slice axial liver scans of wild type female C57BL/6 mice (8-12 weeks of age, n = 5) were acquired sequentially using both SC-MRF and RIPE-MRF methods, resulting in a set of T1, T2, and
M0 maps for each method. The order of the RIPE-MRF and SC-MRF scans was alternated between the animals to avoid potential bias due to the acquisition order (Section 3.7,
Supporting Fig. S3.1). Each mouse was scanned in two separate imaging sessions, using two different levels of anesthesia to evaluate the impact of respiration and heart rate on the two MRF methods. For the first imaging session, respiration was maintained at 45 to 60
50
breaths per minute by adjusting the depth of isoflurane anesthesia (high anesthesia state).
The animals underwent a second imaging session with the respiration rate maintained at
80 to 100 breaths per minute (low anesthesia state). No gating or triggering was utilized so
that it was possible to assess each method’s baseline sensitivity to motion artifacts. Axial
liver images were chosen to provide an MRF dataset with significant respiratory and
pulsatile motion artifacts. Imaging parameters for the in vivo MRF studies were: 3 × 3 cm
field of view, 128 × 128 matrix, and 1-mm slice thickness. The total imaging time for each
mouse imaging session was 1.5 hours (45 minutes each for SC-MRF and RIPE-MRF).
We analyzed the SC-MRF and RIPE-MRF datasets to assess the effect of motion
on the MRF signal evolution profiles. For this qualitative analysis, we utilized the
magnitude of the signal at the center of k-space for the MRF datasets which demonstrates
both sensitivity to motion artifacts as well as limited noise levels in comparison to
individual image voxels. An ideal dictionary profile was plotted with each acquired profile
to illustrate the appearance of deviations due to motion. In addition, composite MRF
images were generated by taking the complex sum of the reconstructed dynamic MRF
images across the entire time domain to visualize the temporal coherence of the motion
artifacts for the RIPE-MRF and SC-MRF methods. In these composite images incoherent noise should add destructively whereas coherent signals will add constructively highlighting any motion artifact coherence through time.
Coherence of the motion artifacts was quantified by manually selecting three regions of interest (ROI) in the background region of the composite images: 1) a region of pulsatility artifacts; 2) a region of respiratory motion artifacts; and 3) an overall artifact
ROI with respiration and pulsatility artifacts. To ensure consistency, ROIs were selected
51
to cover the entire background in the phase-encoding direction for the desired region (see
example ROIs in Section 3.7, Supporting Fig. S3.2). To measure the relative magnitude of
the motion artifacts, mean values for each of the artifact ROIs were divided by the mean
of four separate ROIs, taken in background regions with minimal motion artifacts, to
calculate an artifact-to-noise ratio (ANR)174. The ANR results from the RIPE-MRF and
SC-MRF acquisitions then were compared using unpaired, two-tailed Student’s t-tests.
A similar ROI analysis also was performed on the in vivo MRF-based T1 and T2
maps. Two ROIs were selected within the liver to compare the impact of the motion
artifacts on the mean ROI T1 and T2 values as well as comparing the spatial variation of the
T1 and T2 values within each ROI (measured via standard deviation (SD)). The two areas
were: (1) an area impacted by pulsatility artifact, and (2) an area impacted by respiratory
motion artifact. To ensure consistency of the ROIs selection, the ROI for pulsatile motion
artifact was chosen in the liver directly anterior to the aorta, whereas the ROI for respiratory
motion artifact was chosen to be lateral to the pulsatile motion area (example ROIs in
Section 3.7, Supporting Fig. S3.2). The mean T1 and T2 value, as well as the average spatial
SD of the values within each ROI as a percent of the mean value within the ROI, were
compared between corresponding areas from both methods using unpaired Student’s t-
tests. MRF-based M0 maps were analyzed separately in this study because M0 is a scale factor whereas T1 and T2 are matched parameters. For all experiments, statistical significance was set at p < 0.05.
52
3.3 Results
3.3.1 In vivo Studies
Representative MRF profiles from SC-MRF and RIPE-MRF, along with the dictionary profiles, are shown in Figure 3.3 for both the high and low anesthesia states. The
SC-MRF profiles exhibit regular coherent deviations from the matched dictionary profiles.
For the RIPE-MRF profiles, these deviations are more distributed over the MRF profile and manifest as noise-like spikes. The temporal distribution of these motion-induced variations is consistent between the two anesthesia states, but as expected is more frequent in the low anesthesia state.
Figure 3.3: Comparison of MRF signal profiles. Signal intensity profiles from the center of k‐space for the MRF datasets are shown, with the corresponding dictionary entry overlaid. Profiles from SC‐MRF and RIPE‐MRF for both the (a) high and (b) low anesthesia states highlight differences in the appearance of motion between the methods. SC‐MRF profiles show coherent deviations due to motion, whereas the RIPE‐MRF profiles exhibit noise‐like deviations distributed more evenly throughout the profile. The frequency of these deviations is increased in the low anesthesia state (b), as expected. ANR measurements from the ROI analysis of the in vivo composite MRF images
(Fig. 3.4) are shown in Figure 3.5. ANR from the RIPE-MRF scans was significantly
53
reduced in all ROIs for both anesthesia states in comparison to SC-MRF (*p < 0.005). The
ANR also was reduced for the low anesthesia state compared to the high anesthesia state,
with significant differences seen between anesthesia states for SC-MRF in the respiration and overall motion artifact regions (*p < 0.005), as well as in the overall motion artifact
region for RIPE-MRF (**p < 0.0005).
Figure 3.4: Composite MRF images for both SC‐MRF and RIPE‐MRF. Top and bottom rows are the same images; the top row is windowed to show anatomy; the bottom row is windowed to show artifacts. These composite MRF images were generated by calculating the magnitude of the complex sum of the dynamic MRF images in the time dimension. For these images, temporally coherent signals add constructively resulting in high signal magnitude whereas temporally incoherent signals will add destructively giving lower magnitude. SC‐MRF shows high signal magnitude from artifact in the image background compared to RIPE‐MRF indicating an element of incoherence being added to the artifact in the time domain when using the RIPE‐MRF method. Some residual pulsatility artifacts are seen in the RIPE‐MRF. Additionally, RIPE‐MRF images appear to have less blurring.
54
Figure 3.5: ANRs for SC‐MRF and RIPE‐MRF. An ROI analysis was used on composite MRF images to measure mean ANRs for each MRF acquisition. Three regions corresponding to predominantly pulsatile motion artifacts, predominantly respiratory motion artifacts, and an overall motion artifact were analyzed. Noise ROIs were selected from motion‐free regions of the image background. The RIPE‐MRF method exhibited significantly reduced ANR for both the high and low anesthesia states in the regions of pulsatility (**p < 0.0005, *p < 0.005, respectively), respiration (**p < 0.0005, **p < 0.0005, respectively), and overall (***p < 0.00005, **p < 0.0005, respectively) motion artifacts in comparison to SC‐MRF. The high and low anesthesia states showed significant differences within a method in the area of overall motion artifact for both SC‐ MRF and RIPE‐MRF (*p < 0.005, **p < 0.0005, respectively) and in SC‐MRF in the area of respiration artifact (*p < 0.005).
Representative in vivo MRF-based T1 and T2 maps are shown in Figure 3.6 for both
the high and low anesthesia states with parameter ranges chosen for display purposes.
Consistent with ANR results, the in vivo SC-MRF maps exhibited visible motion artifacts in the phase-encoding direction (blue and green arrows, Fig. 3.6). In contrast, the in vivo
RIPE-MRF maps show visibly reduced artifacts both within the liver and in the background
regions of the MRF maps. The ROI analysis of the MRF maps (Fig. 3.7) demonstrated that
the mean parametric estimates were similar for both methods. The only significant
difference was in mean T1 values in the region of respiration artifact between anesthesia
levels for RIPE-MRF (Fig. 3.7a: *p < 0.05). The spatial SD of the parametric values within
an ROI as a percentage of the mean generally was reduced in the RIPE-MRF maps in
comparison to SC-MRF suggesting improved uniformity of the RIPE-MRF maps. In particular, RIPE-MRF exhibited significantly reduced variation in three out of four regions
55 with pulsatility artifacts (Figs. 3.7c,d: *p < 0.05). The corresponding M0 maps are presented in Section 3.7 as Supporting Figure S3.3; these demonstrate similar artifact patterns as the T1 and T2 maps.
Figure 3.6: Representative T1 (a) and T2 (b) relaxation time maps of a healthy mouse liver. Example axial in vivo SC‐MRF and RIPE‐MRF maps, acquired sequentially in the same imaging session, are shown for both the high and low anesthesia states. Pervasive motion artifacts are seen in the phase‐encoding direction of the SC‐MRF maps. The green arrow highlights an area of predominantly pulsatile motion artifacts in the SC‐MRF scan, whereas the blue arrow shows an area impacted by respiratory motion artifact. These artifacts are minimized in the RIPE‐MRF maps, and the areas appear more homogeneous.
56
Figure 3.7: Liver ROI T1 and T2 measurement analysis. Mean and spatial SD as a percentage of mean value for in vivo liver T1 (a, c) and T2 (b, d) values obtained from the SC‐MRF and RIPE‐ MRF maps in regions of predominantly pulsatile and respiratory motion artifact (green and blue arrows, respectively, in Fig. 3.6) for both high and low anesthesia states. Statistically significant differences in T1 relaxation time were seen between high and low anesthesia states in the area of respiration motion for RIPE‐MRF (*p < 0.05). RIPE‐MRF demonstrates reduced spatial variation of the T1 and T2 estimates as measured by the reduction in the spatial SD of the measurements. For the RIPE‐MRF assessments, the spatial SD was significantly reduced in three out of four T1 comparisons to SC‐MRF (*p < 0.05, **p < 0.005) and one of the four comparisons in T2 (*p < 0.05). The low anesthesia state resulted in increased variation for all comparisons, but only the area of respiration artifact in T2 was statistically significant for both SC‐MRF and RIPE‐MRF (***p < 0.0005, **p < 0.005, respectively). Results of phantom studies showed significant differences between spin echo and
both MRF methods. The MRF methods over-estimated the mean phantom T1 values in comparison to conventional spin echo methods. In addition, the RIPE-MRF method showed small but significant reduction in mean T2 values in comparison to both the SC-
MRF and spin echo methods (Fig. 3.8).
57
Figure 3.8: Phantom results from spin echo, SC‐MRF, and RIPE‐MRF. T1 and T2 maps show similar quantitative values between both MRF methods and reasonable agreement with spin echo. Error bars are shown as the standard deviation of the 5 mean T1 and T2 estimates obtained for each phantom in the in vitro repeatability study. MRF‐based T1 estimates are significantly higher than conventional MRI estimates while MRF‐based T2 estimates are under‐estimated. The phantom results are more consistent between the two MRF methods. (*p < 0.05, **p < 0.01, ***p < 0.005, ****p < 0.0005).
3.4 Discussion
Herein, we describe the RIPE-MRF scheme to suppress respiratory and pulsatile motion artifacts in preclinical MRF applications. The RIPE-MRF approach attempts to
decrease the temporal coherence of motion artifacts by varying the acquired k-space line during the MRF acquisition (Fig. 3.1). Linear increments in the phase-encoding line were added during the dynamic MRF acquisition (Fig. 3.2) resulting in significant reductions in the coherence of motion artifacts (Figs. 3.3, 3.4, 3.5). The application of the modified view ordering to minimize the effect of motion artifacts previously has been used to improve image quality175–178. The goal of this work was to utilize view ordering within the MRF
58
framework to limit the impact of motion artifacts on the resulting quantitative maps.
Subsequently, we show that this new RIPE-MRF trajectory provides a significant reduction in the coherence of motion artifacts, generating in vivo T1 and T2 measurements with
reduced spatial variation (Figs. 3.6, 3.7) compared to previously reported preclinical MRF
methods137.
In vivo application of the RIPE-MRF methodology altered the temporal distribution
of motion artifacts in the MRF profiles (Fig. 3.3). Although the periodic, motion-induced
deviations in the MRF profiles are coherent for SC-MRF, RIPE-MRF deviations are noise- like displacements that are more evenly distributed across the entire MRF signal evolution profile. Composite images (Fig. 3.4) and ANR measurements (Fig. 3.5) show these differences in the image domain. The reordering of the k-space datasets with RIPE-MRF
limited the temporal coherence of the motion artifacts and resulted in significantly reduced
ANR. Importantly, these ANR reductions were consistent for both pulsatility and
respiratory motion artifacts and for both high and low anesthesia states.
Motion-induced artifacts in the MRF signal evolution profiles also resulted in
corresponding alterations in the MRF-based T1 and T2 maps. In vivo RIPE-MRF maps
demonstrated a visible reduction in the appearance of motion artifacts in the liver compared
to SC-MRF maps (Fig. 3.6). The impact of adding temporal incoherence to the motion
artifacts on MRF-based quantification was manifest through reduced spatial variation in
the hepatic T1 and T2 values, as measured by RIPE-MRF in both regions of respiration and pulsatility artifacts (Fig. 3.7). Furthermore, RIPE-MRF resulted in reduced spatial variation
for both levels of anesthesia, indicating that it is effective in reducing the impact of motion
artifacts for a range of physiological states. Both MRF methods resulted in mean hepatic
59
47,171 T1 estimates consistent with previous reports . Liver T2 estimates from both MRF
115,171 methods were reduced in comparison to reported literature values . This T2
underestimation also was observed for the MRF phantom results in comparison to
conventional MRI techniques (Fig. 3.8). T2 underestimation partially may be ameliorated
through corrections to B0 and B1 heterogeneities but was not further explored in this initial study.
The primary advantage of the RIPE-MRF method is eliminating the need for
physiological gating and/or triggering to obtain artifact-free quantitative maps in small
animal imaging. Prior preclinical studies have performed self-navigation and/or motion
correction122,123,179,180, retrospectively identified corrupted data using monitoring
systems167,181, or triggered the acquisition during the quiescent period of motion17,182. These
methods reduce motion artifacts but also significantly increase scan time and/or require
complex post-processing. In contrast, RIPE-MRF provides a straightforward quantification
process utilizing all available data with no restrictions. Additionally, the method is robust
across a range of physiological motion rates reducing the need for precise physiological
controls (i.e., intubation). Higher variation in the MRF maps in the regions of high-
frequency pulsatility artifacts indicate that the frequency of motion artifacts is an important
factor in MRF trajectory design. For example, further reductions in motion artifacts may
be needed to provide reliable MRF maps in regions with high-frequency/rapid motion (e.g.,
pulmonary/cardiac imaging). Therefore, future studies will be needed to investigate
alternative view-ordering methods (e.g., random phase-encoding) and/or gated MRF
149 approaches to provide artifact-free MRF-based T1 and T2 maps .
60
In this initial study, non-Cartesian trajectories were not explored as a potential form of motion suppression. Prior preclinical studies incorporating non-Cartesian trajectories, such as PROPELLER165,183, spiral15,17,117,182, and radial41,184,185 k-space sampling, show
promising motion artifact suppression. However, the resistance of non-Cartesian MRF
methods to motion must be balanced with additional error sources, such as eddy current-
induced trajectory errors186 and off-resonance artifacts prevalent on high-field preclinical
MRI scanners17. From a practical perspective, non-Cartesian trajectories may be difficult or even impossible to implement on some preclinical MRI scanners. Further, this Cartesian
RIPE-MRF approach may be useful for both preclinical and clinical MRF applications in combination with established parallel imaging strategies187. Regardless, future studies will
be needed to thoroughly compare the relative motion artifact resistance of the Cartesian
RIPE-MRF approach with non-Cartesian MRF trajectories.
3.5 Conclusions
In conclusion, we have developed the motion artifact-resistant RIPE-MRF method
to provide reliable multi-parametric quantification for preclinical MRF applications. In this
initial implementation, incrementing the phase-encoding line during the dynamic MRF
acquisition resulted in suppression of both pulsatility and respiratory motion artifacts in the
in vivo MRF-based T1 and T2 relaxation time maps of mouse abdomens. This improved
resistance to motion artifacts was achieved with no change in the MRF acquisition time, as
well as with minimal impact on the accuracy of the T1 and T2 estimates. The RIPE-MRF
method represents a shift in the preclinical quantitative imaging paradigm from attempting
to reduce or eliminate motion artifacts in the underlying images to accepting the presence
of motion artifacts but manipulating them so they are suppressed during quantification.
61
Thus, the RIPE-MRF method serves as a foundation for free-breathing Cartesian MRF in
rodents.
3.6 Supporting Methods
This section provides additional details on the methods presented in Section 3.2.
3.6.1 MRF Sequence
MRF uses a series of variable flip angles and repetition times to generate a series
of images with changing contrast based on the chosen imaging kernel and sequence
sensitivity. Both SC-MRF and RIPE-MRF were based on a fast imaging with steady state
free precession imaging kernel to generate sensitivity to both T1 and T2 and an inversion
preparation was used to enhance the overall sensitivity of the sequence to T1. The sequence
implemented in this work used a sinusoidal pattern of variable flip angles ranging from 0
to 70 degrees and a Perlin noise pattern of variable repetition times varying from 9.5 to 12
ms similar to the pattern used by Ma et al. in the initial MRF implementation136. For this fully sampled Cartesian acquisition, the acquisition parameter patterns were truncated to
1024 time points to reduce acquisition time and to facilitate the phase encoding increments used for the RIPE-MRF method. A schematic diagram of the MRF pulse sequence and TR and FA patterns can be seen in Figure 3.2. After each set of dynamic MRF images was acquired (1024 k-space lines), a 10 second delay was added to allow the magnetization to return to equilibrium prior to the next set of 1024 MRF k-space lines similar to the method proposed by Gao et al137. Prior to the acquisition of the very first set of 1024 k-space lines a dummy set of 1024 k-space lines is performed but not acquired to establish the
“equilibrium magnetization” for the MRF experiment. Following the MRF acquisitions, all data were exported to MATLAB for reconstruction and analysis.
62
3.6.2 MRF Dictionary and Quantification:
MRF quantification was performed in a similar manner to that described by Ma et
al.136 with the method and differences summarized here. Quantification was accomplished
by matching the acquired profiles on a pixel by pixel basis to a single dictionary of
simulated profiles from all logical combinations of T1 (50-3000 ms, increment: 10 ms;
3000-5000 ms, increment: 100 ms; 5000-10,000 ms, increment: 200 ms) and T2 (2-200 ms,
increment: 2 ms; 200-500 ms, increment: 10 ms; 500-800 ms, increment: 50 ms) resulting in 44,667 dictionary entries. All MRF dictionary profiles were simulated in MATLAB based on the Bloch equations using combinations of T1 and T2 to create dictionary entries.
Matched T1 and T2 values for each imaging pixel were identified through template
matching by finding the maximum inner product between the acquired profile and the
individual dictionary entry. In addition, proton density (M0) was estimated as the scale
factor between the acquired MRF data and simulated dictionary profile as described
previously. The same MRF dictionary was used for all in vitro and in vivo experiments.
3.6.3 Phantom MRF Studies
Four phantoms were prepared in NMR tubes (Norell, Inc. Morganton, North
-1 Carolina, USA) by diluting a 1 mol L stock solution of MnCl2 (Sigma-Aldrich, St. Louis,
Missouri, USA) with deionized water to achieve concentrations of 50, 100, 200, and 300
μmol L-1. Axial 2D SC-MRF and RIPE-MRF datasets were acquired with a 2 × 2 cm FOV,
128 × 128 matrix, and 1.5-mm slice thickness to minimize the impact of noise on the data
analysis. Each MRF acquisition (SC-MRF and RIPE-MRF) was identical during imaging
except for the ordering of the phase encoding lines (Fig. 3.1). SC-MRF was repeated one
time on 5 separate days, and RIPE-MRF was repeated one time during a different set of 5
63
separate days (5 repeats for each method, 10 total scan days) to assess the reproducibility
of the method. An ROI analysis was used to calculate the mean T1 and T2 values in each
phantom for each individual imaging session. These mean values were averaged over the
5 different acquisitions and a standard deviation over time was calculated. The phantom
results from both MRF acquisitions were then compared using an unpaired, two-tailed
Student’s t-test. MRF results were also compared to gold-standard spin echo measurements
to evaluate the accuracy of the MRF method (Fig. 3.8).
3.6.4 Conventional MRI Measurement of T1 and T2 Relaxation Times:
The MRF estimates of T1 and T2 were compared with conventional MRI techniques in phantoms. Conventional T1 estimates were obtained with an inversion recovery spin
echo (IR-SE) with TR = 10,000 ms, TE = 8.5 ms, and nine inversion times (TI= 50, 250,
500, 750, 1000, 1500, 2500, 5000, 10,000 ms). Inversion was accomplished using an
adiabatic inversion pulse with a slab thickness of 5x the imaging slice. A single-echo spin
echo (SE) acquisition was used for T2 quantification (TR = 10,000 ms, nine echo times: TE
= 10, 25, 40, 60, 90, 120, 150, 300, 500 ms). Relevant imaging parameters include FOV=2
× 2 cm, matrix=128 × 128, and slice thickness of 1-mm. Excitation was accomplished with
a 10-lobe sinc pulse to reduce any slice profile effects. For both the conventional IR-SE
and SE acquisitions, the refocusing pulse was a 3-lobe sinc pulse with a slice thickness 3x
that of the imaging slice to ensure uniform refocusing of the excited spins across the
imaging slice. Acquired data were used to create relaxation curves which were fit to the
appropriate exponential model using non-linear least squares fitting. Studies were repeated
on five separate days and mean values used as the true value for the T1 and T2 relaxation
times of the phantoms.
64
3.7 Supporting Figures
Supporting Figure S3.1: Workflow of in vivo MRF experiments. Shown is the experimental work flow for two animals (out of 5 total) representing the two possible experimental designs. Three mice were imaged using the protocol demonstrated in Mouse 1 and two mice were imaged using the protocol for Mouse 2. The high anesthesia session was performed first to evaluate motion suppression in the presence of lower respiration and heart rates. Within this session the scan order was alternated to average out any physiological changes over the course of the experiment. Then 3-4 weeks later the low anesthesia state was imaged to analyze higher respiration and heart rates with the same alternation of the scan order. Total imaging time for each MRF acquisition was 45 minutes with the total scanning session being 1.5 hours (RIPE-MRF and SC-MRF).
Supporting Figure S3.2: Shown are representative composite MRF images and MRF-based T1 and T2 maps to illustrate how ROIs were selected. For ANR measurements on the composite images (left column) ROIs were chosen to cover the entire phase encoding direction for the respiration artifact (blue), pulsatility artifact (green), and overall artifact (red) to analyze a similar number of pixels for each animal. T1 and T2 ROIs (middle and right column, respectively) show the presence
65
of ROIs for respiration artifact (blue) and pulsatility artifact (green). Pulsatility was chosen anterior to the aorta and respiration was chosen laterally from the area of pulsatility to get consistently selected ROIs between animals.
Supporting Figure S3.3: Representative M0 maps from in vivo SC-MRF and RIPE-MRF acquisitions. These maps correspond to the T1 and T2 maps seen in Figure 3.6. In this study, the MRF-based M0 maps are estimated as a scale factor and are not matched like T1 and T2. RIPE- MRF maps show distinct reductions in motion artifacts in comparison to SC-MRF similar to the T1 and T2 maps shown in Figure 3.6.
66
Chapter 4: Dual Contrast - Magnetic Resonance Fingerprinting (DC-MRF)
4.1 Introduction
Over the past 3 decades, Magnetic Resonance Imaging (MRI) has become an
essential medical imaging modality due to its exceptional soft-tissue contrast and lack of
ionizing radiation. Along with a wide variety of endogenous tissue contrast mechanisms,
many MRI applications utilize an intravenous injection of an MRI contrast agent (e.g.,
gadolinium chelates or iron oxides) to enable sensitive identification of numerous
pathologies such as tumors188, vascular abnormalities189, and cardiac infarcts190 through
local alterations in the tissue’s magnetic properties (T1 and T2 relaxation times). Clinical
use of these contrast-enhanced MRI scans has further expanded as multiple contrast agents
have been approved for specific clinical imaging applications (e.g., blood pool contrast
agents191, hepatobiliary contrast agents192).
With the emergence of the field of molecular imaging, there has been a dramatic increase in the number of MRI contrast agents targeted to proteins78,79,89, cell receptors88,91,
and other molecular species90,92. In addition, a number of activatable agents have been
described that have different relaxivities based on the local tissue environment in vivo94,110.
In a typical preclinical molecular MRI study, the longitudinal relaxation time (T1) or
transverse relaxation time (T2) (or MRI signal intensity) is measured dynamically before
and after contrast agent administration allowing tracking of the agent’s distribution. These
studies often incorporate non-targeted contrast agents (e.g., scrambled peptides) as controls
to verify the in vivo molecular specificity of the targeting moiety. Due to the difficulty of previously-developed contrast enhanced MRI strategies to uniquely identify two MRI
contrast agents administered simultaneously, the targeted and untargeted contrast agents
67
must be studied in separate imaging sessions and likely in separate animal cohorts86,108,193.
This significant limitation can result in experimental bias due to phenotypic variation. As
such, the development of a “multi-color” MRI methodology to independently monitor simultaneously-administered targeted and control MRI contrast agents, and potentially multiple targeted MRI contrast agents, would significantly improve preclinical molecular
MRI studies. This “multi-color” MRI capability would also provide a robust pathway for clinical translation of molecular MRI contrast agents by providing the capability to simultaneously compare the biodistribution of a molecular imaging agent with a conventional clinical MRI agent.
The primary limiting factor in the simultaneous assessment of multiple paramagnetic MRI contrast agents is that the unique identification of each agent is challenging. MRI contrast agents directly impact both the T1 and T2 relaxation times
according to well-established concentration-dependent linear relationships to their
63 magnetic relaxivities (r1 and r2) shown in equations 4.1 and 4.2 below :
1 1 = + [ ] (4.1)
1𝐴𝐴 �𝑇𝑇1 �𝑇𝑇10 𝑟𝑟 𝐴𝐴 1 = 1 + [ ] (4.2)
2𝐴𝐴 � 2 � 20 𝑟𝑟 𝐴𝐴 where [A] is the concentration of imaging𝑇𝑇 agent𝑇𝑇 A; T10 and T20 are the pre-contrast T1 and
T2 relaxation times of the tissue; T1 and T2 are the post-contrast T1 and T2 relaxation times;
and r1A and r2A are the magnetic relaxivities of contrast agent A. Therefore, while an individual MRI contrast agent is typically more sensitive to a particular relaxation parameter (i.e., Gd-chelates for enhancement in T1-weighted imaging acquisitions), each paramagnetic MRI contrast agent still impacts both the T1 and T2 relaxation times. This
68
important factor limits the capability of MRI to independently assess simultaneously-
administered contrast agents (e.g., a Gd-based T1 agent and an iron-based T2 agent).
The Magnetic Resonance Fingerprinting (MRF) methodology has recently been
developed to simultaneously generate inherently co-registered T1 and T2 relaxation time
maps in both patients149,159,194 and animal models137,153. MRF uses a unique acquisition and
quantification strategy that combines a priori acquisition parameter variation with a
dictionary-based pattern matching algorithm to obtain quantitative assessments of multiple
imaging parameters simultaneously. Importantly, MRF has been shown to provide
quantitative T1 and T2 maps in 10–50 seconds per imaging slice providing the opportunity
to dynamically generate quantitative maps of these two important MRI parameters
simultaneously136,144,154,173. In this study, we demonstrate that the framework for
simultaneous T1 and T2 assessments provided by MRF can be used to analytically quantify
the local concentration of two different MRI contrast agents present at the same time.
Herein, we describe a straightforward multiple contrast agent relaxation model (Equations
4.3 and 4.4 in Section 4.2) that can be used in combination with the rapid, multi-parametric
MRF strategy to independently calculate inherently co-registered concentration maps for two MRI contrast agents. These initial in vitro results represent a proof-of-concept study to: (1) validate the multiple contrast agent relaxation model using a 60 MHz magnetic relaxometer; and (2) demonstrate the application of the rapid MRF method to enable simultaneous calculation of concentration maps for two different paramagnetic MRI contrast agents on a clinical 3 T MRI scanner. Overall, these in vitro studies suggest an imaging framework for future in vivo MRI studies to simultaneously quantify multiple contrast agents.
69
4.2 Methods
This section provides details on the validation of the multiple contrast agent relaxation model and the MRF acquisition used in combination to simultaneously estimate the concentration of two paramagnetic MRI contrast agents.
4.2.1 Multiple Contrast Agent Relaxation Model
As described in Section 4.1, a single paramagnetic MRI contrast agent exhibits
concentration-dependent T1 and T2 relaxation effects as described by Equations 4.1 and
4.2, respectively. Herein, we are proposing a straightforward linear model to incorporate a
second MRI contrast agent B as shown in Equations 4.3 and 4.4:
1 1 = + [ ] + [ ] (4.3)
1𝐴𝐴 1𝐵𝐵 �𝑇𝑇1 �𝑇𝑇10 𝑟𝑟 𝐴𝐴 𝑟𝑟 𝐵𝐵 1 = 1 + [ ] + [ ] (4.4)
2𝐴𝐴 2𝐵𝐵 � 2 � 20 𝑟𝑟 𝐴𝐴 𝑟𝑟 𝐵𝐵 where [B] is the concentration 𝑇𝑇of agent𝑇𝑇 B, and r1B and r2B are the magnetic relaxivities of
contrast agent B. These equations suggest that if T10, T20, T1, and T2 are measured before
and after simultaneous injection of two MRI contrast agents with known relaxivities, then
these two equations have only two unknowns allowing for the direct analytical calculation
of [A] and [B].
In this initial study, we tested the validity of this model using in vitro phantoms
containing varying concentrations of gadolinium (Multihance®, gadobenate dimeglumine)
and manganese (MnCl2,1 M Stock Solution, Sigma-Aldrich, #M1787) contrast agents either as single agents (i.e., Gd or Mn only) or as mixtures of the two agents (i.e., Gd and
Mn combined). Gadolinium and manganese were chosen due to the wide clinical availability (Gd) and distinct relaxivity properties of the two agents. We first prepared 70
serial dilutions of each contrast agent individually in deionized water to enable assessment
of the individual relaxivities for each agent (r1G, r2G, r1M, r2M). A table explicitly listing the
concentrations of the phantoms is provided as Table 4.1. For the gadolinium agent, the
concentration was varied from 0.05 to 0.5 mM (Table 4.1 Phantoms 1–5, n = 5). For the
manganese agent, the concentration was varied from 0.0125 to 0.2 mM (Table 4.1
Phantoms 6–10, n = 5). For both gadolinium and manganese a phantom of pure solvent
was also analyzed (deionized water, Table 4.1 Phantom 17). We then prepared mixtures of
the two MRI contrast agents in the same solvent (Gd concentration range = 0.025 to 0.355
mM; Mn concentration range = 0.00625 to 0.15 mM; Table 4.1 Phantoms 11–16, n = 6).
50mL of each solution was prepared and served as a source for both the 60 MHz
relaxometry and 3 T experiments.
To test the multiple contrast agent relaxation model in Equations 4.3 and 4.4, each
of the samples were transferred to 5 mm NMR tubes (Norell, 507-HP-7) and scanned on a
Bruker Minispec 60 MHz relaxometer (Bruker Biospin, Billerica, MA). The relaxometer
was used to obtain T1 and T2 relaxation time assessments for each sample using an
inversion recovery spin echo technique (T1: 7 inversion times) and a multi-echo Carr-
Purcell-Meiboom-Gill (CPMG) MRI acquisition (1,000–10,0000 echoes)195. All relaxometry experiments were conducted at 37°C. The relaxometric measurements were repeated twice for T1 measurements and three times for T2 measurements to ensure accurate
assessments. These repeated measures were averaged to obtain single T1 and T2 values for
each in vitro phantom.
71
Table 4.1: In vitro Phantom Concentrations. 5 Gadolinium-Only Phantoms (1-5), 5 Manganese- Only Phantoms (6-10), 6 Phantoms with Both Contrast Agent (11-16), and one Solvent-Only Phantom (17, deionized water).
Gadolinium Manganese Phantom Concentration (mM) Concentration (mM)
1 0.5 0 2 0.3 0 3 0.2 0 4 0.1 0 5 0.05 0 6 0 0.2 7 0 0.1 8 0 0.05 9 0 0.025 10 0 0.0125 11 0.03125 0.15 12 0.0625 0.1 13 0.125 0.05 14 0.25 0.025 15 0.355 0.0125 16 0.025 0.00625 17 0 0
The resulting T1 and T2 values for each phantom were then used to calculate the
agent concentration in each tube in a multi-step process. First, the T1 and T2 values for the
solvent (deionized water) alone were measured and established as “pre-contrast” relaxation times (T10 and T20). Second, the phantoms containing the individual agents were analyzed
to calculate the magnetic relaxivities of both agents (r1G, r2G, r1M, r2M) through a linear least-
-1 -1 squares fit to the plot of R1 (1/T1 in ms ; Equation 4.1) and R2 (1/T2 in ms ; Equation 4.2)
72 as a function of gadolinium or manganese concentration (in mM) using established methods. The slopes of the resulting fits were used as the magnetic relaxivities for each agent. From these relaxivity results, as well as the “pre-contrast” T10 and T20 values, the T1 and T2 relaxation times for each sample were used to analytically calculate the gadolinium and manganese contrast agent concentrations for each sample using Equations 4.5 and 4.6.
These equations were derived directly from the algebraic solution to Equations 4.3 and 4.4
(with Gd = A and Mn = B):
( × ) ( × ) [ ] = (4.5) ( × ) ( × ) ∆𝑅𝑅2 𝑟𝑟1𝑀𝑀 − ∆𝑅𝑅1 𝑟𝑟2𝑀𝑀 𝐺𝐺𝐺𝐺 𝑟𝑟2𝐺𝐺 𝑟𝑟1𝑀𝑀( − ×𝑟𝑟1𝐺𝐺[ ]𝑟𝑟)2𝑀𝑀 [ ] = (4.6) ∆𝑅𝑅2 − 𝑟𝑟2𝐺𝐺 𝐺𝐺𝐺𝐺 𝑀𝑀𝑀𝑀 2𝑀𝑀 where ΔR1 = 1/T1–1/T10, ΔR2 = 1/T2–1/T20. The𝑟𝑟 agent concentrations calculated from
Equations 4.5 and 4.6 were then compared with known concentrations in each phantom using Pearson correlations with a probability of p < 0.05 used as a determination for significance.
4.2.2 In vitro DC-MRF Assessments at 3 T
To determine the capability of the DC-MRF technique to provide quantitative imaging based assessments of contrast agent concentration, we obtained MRF-based T1 and T2 maps of the gadolinium-containing and manganese-containing phantoms evaluated in the relaxometry studies described above. The solutions in these phantoms were taken from the same source solutions as the NMR studies and placed into 15 mL centrifuge tubes
(Fisher Scientific, S50712). All MRF studies were conducted on a Siemens Skyra 3 T MRI scanner (Siemens Healthineers, Erlangen, Germany). The MRF acquisition (Siemens
Work-In-Progress #881v23) utilized a FISP acquisition kernel designed with a priori
73
variation in both flip angle (FA) and repetition time (TR, baseline TR of 12 ms) to generate
144 MRF signal evolution profiles sensitive to both T1 and T2 relaxation times . The MRF method acquired 3000 images with time-varying contrast generated by the FA and TR variation. This FISP-MRF implementation included a non-selective inversion preparation
(inversion time = 21 ms) immediately prior to the FISP-MRF image acquisitions to increase
137 the sensitivity of the MRF signal evolution profiles to T1 relaxation times . The MRF acquisition also incorporated undersampled spiral trajectories as described previously144.
The acquisition time of one slice was 47 seconds with a FOV of 380 × 380 mm, an image
matrix of 352 × 352, and a slice thickness of 5 mm. Measurement of the excitation (B1)
field was incorporated to mitigate the effects of inhomogeneous B1 field on the T1 and T2
relaxation time estimates138.
A fundamental component of the MRF reconstruction is the development of a large
dictionary of signal evolution profiles that are subsequently “matched” to the acquired
MRF signal evolution profile for each imaging voxel using vector-based inner product comparisons. The MRF dictionary was created as described previously136 assuming a
mono-exponential relaxation model. MRF image reconstruction as well as subsequent
dictionary matching were performed on the Siemens Skyra 3 T MRI scanner. Generation
of the quantitative MRF-based T1 and T2 relaxation time maps was attained by matching
the acquired MRF profiles on a pixel-by-pixel basis to the MRF dictionary of simulated
profiles from all logical combinations of T1 (10-100 ms, increment = 10 ms; 100-1000 ms, increment = 20 ms; 1000-2000 ms, increment = 40 ms; 2000-4500 ms, increment = 100 ms) and T2 (2-100 ms, increment = 2 ms; 100-150 ms, increment = 5 ms; 160-300 ms,
increment = 10 ms; 300-800 ms, increment = 50 ms; 800-1600 ms, increment = 100 ms;
74
1600-3000 ms, increment = 200 ms). The T1 and T2 maps were exported for further offline
processing in MATLAB (MathWorks, Natick, MA).
Mean T1 and T2 values for each phantom were obtained from the MRF maps using
a region of interest (ROI) analysis. Similar to the relaxometric studies above, mean MRF-
based T1 and T2 values for the sample with no contrast agent (deionized water only) was used as a measure of T10 and T20, respectively. The mean T1 and T2 relaxation times from
the phantoms containing only a single agent were used to estimate the magnetic relaxivities
for the gadolinium and manganese contrast agents. The T1 and T2 maps from the MRF acquisition were then used to calculate gadolinium and manganese concentration maps using the calculated 3 T relaxivities and Equations 4.5 and 4.6, respectively. An ROI analysis of the maps was then used to calculate a mean gadolinium and manganese concentration value for each phantom. The mean DC-MRF concentration estimates were
compared with known values using Pearson correlations. The MRF acquisition was
repeated 12 times allowing for sample repositioning as well as scanner adjustments to test
the capability of the DC-MRF method to statistically differentiate samples with both
contrast agents.
Magnetic relaxivities were also obtained using conventional MRI assessments for
comparison with the MRF-based relaxivity assessments. T1 relaxation time assessments
were obtained with an inversion recovery spin echo acquisition (8 inversion times), and T2
relaxation times were obtained with a single-echo spin echo acquisition (8 echo times).
Fitting was performed with the appropriate mono-exponential model generating T1 and T2
maps. Equations 4.1 and 4.2 were then used to calculate r1 and r2 for the gadolinium and
manganese contrast agents, respectively.
75
4.3 Results
4.3.1 60 MHz Relaxometry: Multiple Contrast Agent Relaxation Model Validation
Figure 4.1 shows the individual magnetic relaxivity plots (R1 and R2 vs
concentration) for the gadolinium and manganese contrast agent phantoms obtained from
the 60 MHz relaxometer using inversion recovery spin-echo acquisitions for T1 relaxation
time measurements followed sequentially by a Carr-Purcell-Meiboom-Gill (CPMG) acquisition for T2 relaxation time measurements. A list of in vitro phantoms and their
respective concentrations of contrast agents is shown in Table 4.1. The T1 measurements
were repeated twice and the T2 assessments were repeated three times to ensure
consistency. The relaxivity data resulted in significant linear correlations as expected from
Equations 4.1 and 4.2 (R2 > 0.993, p < 0.0001). The 60 MHz magnetic relaxivities at 37 ºC
-1 -1 for the gadolinium agent (slope of the linear correlation lines) were 0.0040 mM ms (r1)
-1 -1 and 0.0048 mM ms (r2). The corresponding magnetic relaxivities for the manganese
-1 -1 -1 -1 agent were 0.0054 mM ms (r1) and 0.0652 mM ms (r2). The magnetic relaxivity values
for the 60 MHz relaxometer, as well as the relaxivities measured at 3 T, are shown in Table
4.2. The mean T1 and T2 relaxation times for the deionized water phantom with no contrast agent (deionized water alone) were 4250 ms (T10) and 2760 ms (T20), respectively. NMR
relaxivity measurements are in reasonable agreement with previously reported results for
both the Gd64 and Mn61,66 contrast agents used here.
76
Figure 4.1: Relaxivity assessments for (a) gadolinium (Gd) and (b) manganese (Mn) contrast agents from a 60 MHz relaxometer using phantoms containing varying concentrations of a single contrast agent. Slopes of the fitted lines of R1, R2 vs. agent concentration (n=6 for each agent) were used to determine the relaxivities (r1 and r2) of the two agents. Pearson correlations resulted in 2 significant correlations of concentration vs. R1 and R2 for both contrast agents (R > 0.993, two- tailed probability p < 0.0001). Equations 4.5 and 4.6 were then used to calculate the gadolinium and manganese
concentrations for all of the phantoms (n = 17) using the measured T1 and T2 values for
each in vitro sample, the calculated magnetic relaxivities for each contrast agent, and the
“non-contrast” T10 and T20 values for the deionized water samples. Plots of the
relaxometry-based concentration estimates against the known concentration in each
phantom for both gadolinium and manganese chloride are shown in Fig. 4.2. Concentration
estimates obtained from the 60 MHz relaxometry data shown in Fig. 4.2 were calculated
for both the six samples containing both gadolinium and manganese contrast agents (Table
4.1 Phantoms 11–16) as well as the ten samples containing either gadolinium or manganese
used in the relaxometry analysis (Fig. 4.1, Table 4.1 Phantoms 1–10). One additional
sample containing only deionized water was also scanned (Table 4.1 Phantom 17). Pure
samples with a single contrast agent (either Gd or Mn) were analyzed to increase the
number of concentrations calculated and to verify that the method returned a value of 0 if
the agent was not present in solution. This comparison resulted in a significant linear
correlation (Pearson Correlation: R2 > 0.998, two-tailed probability p < 0.0001) for both
77
the gadolinium and manganese contrast agents. Importantly, the slopes of these correlations
are both near unity (1.003 (Gd) and 0.975 (Mn), respectively) suggesting that the
concentration estimates obtained from the multiple contrast agent model results in good
agreement with the known concentrations. The results from only the six samples with both
agents (Table 4.1 Phantoms 11–16) are shown as a subset of the data separately (Fig. 4.3).
These initial relaxometric results demonstrate that the multiple contrast agent relaxation
model shown in Equations 4.3 and 4.4 is capable of providing accurate estimates for the
concentration of gadolinium and manganese-based contrast agents. Importantly, these
results appear to be consistent whether there is 0, 1, or 2 contrast agents in the phantom.
Figure 4.2: Pearson correlation plots of estimated (a) gadolinium (Gd), and (b) manganese (Mn) concentration versus known phantom concentrations (n=17). Estimated concentrations were obtained from Equations 4.5 and 4.6 for data obtained from a 60 MHz relaxometer. Note the significant correlation between the estimated and actual agent concentrations over all phantoms (Pearson Correlation: R2 > 0.998, two-tailed probability p < 0.0001).
Figure 4.3: Pearson correlation plots of mean DC-MRF estimates for Gd (left) and Mn (right) concentrations against known concentrations at 60 MHz from phantoms containing mixtures of
78
both MRI contrast agents (Table 4.1 Phantoms 11-16; n=6). Concentrations were estimated from the 60 MHz measurements of T1 and T2 and using Equations 4.5 and 4.6. These data are a subset of the data shown in Figure 4.2. 4.3.2 3 T DC-MRF: Simultaneous Assessment of Two Paramagnetic MRI Contrast Agents
Similar to the relaxometry results above, T1 and T2 relaxation time assessments for
gadolinium and manganese containing phantoms (n = 17) were obtained using the MRF
method on a clinical 3 T MRI scanner at room temperature. Phantoms scanned contained
the same contrast agent concentrations as for the NMR relaxometer experiments described
above. This data was acquired using a FISP-MRF acquisition repeated 12 times following repositioning to measure average T1 and T2 values for each phantom. In contrast to the
relaxometer studies above, the MRF-based T1 and T2 measurements were obtained
simultaneously for all phantoms (n = 17). Figure 4.4 shows the individual magnetic
relaxivity plots (R1 and R2 vs concentration) for the gadolinium and manganese contrast
agents obtained from the MRF data averaged over the 12 repeats. The relaxivity data
resulted in a significant linear correlation for both the gadolinium agent (Pearson
2 -1 -1 Correlation: R ≥ 0.997, two-tailed probability p < 0.0001; r1 = 0.0056 mM ms ; r2 =
0.0076 mM-1ms-1) and the manganese agent (R2 ≥ 0.999, two-tailed probability p < 0.0001;
-1 -1 -1 -1 r1 = 0.0067 mM ms ; r2 = 0.1144 mM ms ). These MRF-based relaxivity values compared favorably to relaxivity values obtained from conventional inversion recovery
-1 -1 and single-echo spin echo MRI experiments for both gadolinium (r1 = 5.1 mM ms ; r2 =
-1 -1 -1 -1 -1 -1 6.0 mM ms ) and manganese (r1 = 6.8 mM ms ; r2 = 107.9 mM ms ). Relaxivity results
at 60 MHz and 3 T are summarized in Table 4.2. These relaxivity results also were in
64,66,196 reasonable agreement with literature values . DC-MRF was also used to obtain T1
79
and T2 relaxation times for the deionized water sample with no contrast agent (T10 = 2897
ms; T20 = 946 ms).
−1 −1 Table 4.2: Comparison of relaxivity measurements (r1, r2) in mM ms for each contrast agent in deionized water at room temperature between 60 MHz*, 3.0 T spin echo (SE), and 3.0 T MRF. *60 MHz measurements made at 37 °C. Field Strength 60 MHz 3.0 T
Contrast Agent r1 r2 r1 SE r1 MRF r2 SE r2 MRF
Gd 0.0040 0.0048 0.0051 0.0056 0.0060 0.0076 Mn 0.0054 0.0652 0.0068 0.0067 0.1079 0.1144
Figure 4.4: MRF-based relaxivity assessments for (a) gadolinium (Gd) and (b) manganese (Mn) contrast agents obtained on a 3 T MRI scanner using phantoms containing varying concentrations of a single contrast agent. Slopes of the fitted lines of R1, R2 vs. agent concentration (n=6 for each agent) were used to determine the relaxivities (r1 and r2) of the two agents. Pearson correlations 2 resulted in significant correlations of concentration vs. R1 and R2 for the two contrast agents (R ≥ 0.997, two-tailed probability p < 0.0001). Representative DC-MRF maps of estimated gadolinium and manganese
concentration calculated on a pixel-by-pixel basis from MRF-based T1 and T2 maps using
Equations 4.5 and 4.6 are shown in Fig. 4.5a for both the phantoms containing only
gadolinium contrast agent (Table 4.1 Phantoms 1-5; n = 5), only manganese contrast agent
(Table 4.1 Phantoms 6-10; n = 5), both gadolinium and manganese agents (Table 4.1
Phantoms 11-16; n = 6), or solvent alone (Table 4.1 Phantom 17; deionized water, n = 1).
Theoretical maps of the known phantom concentrations are shown for comparison in Fig.
4.5b. The estimated gadolinium and manganese concentration maps are visually consistent
80
with the known concentrations. A quantitative comparison of the mean DC-MRF concentration estimates from the region of interest analysis with the known concentrations are shown in Figure 4.6 and resulted in significant correlations (Pearson Correlations: Gd:
R2 = 0.9987, two-tailed probability p < 0.0001; Mn: R2 = 0.998, p < 0.0001). The mean
DC-MRF concentrations were also in reasonable agreement with actual values as evidenced by the slopes of the correlation lines equal to 0.988 (Gd) and 0.980 (Mn), respectively. Results from the phantoms containing mixtures of both contrast agents are presented separately (Fig. 4.7).
Figure 4.5: Maps of estimated gadolinium (Gd) and manganese (Mn) concentration from DC-MRF method (a). Simulated maps of known concentrations are shown for comparison (b). Note the general agreement between DC-MRF estimates and actual concentrations over a wide range of concentrations (n=17). Note also the absence of signal from the vials containing only a single agent (bottom row of maps marked by green arrow contain only Gd, 3rd row of maps marked by white arrow contain only Mn) indicating that the multiple contrast agent relaxation model and acquisition appears to be valid when the agents are used alone or in tandem (top two rows marked by blue arrows contain mixtures of both Gd and Mn contrast agents).
81
Figure 4.6: Pearson correlation plots of mean DC-MRF estimates for (a) gadolinium (Gd), and (b) manganese (Mn) concentration versus known phantom concentrations (n=17). Mean DC-MRF concentration estimates were obtained from an ROI analysis of the MRF-based T1 and T2 relaxation time maps. The gadolinium and manganese concentrations were calculated for each of the MRF scans (n=12) and averaged to calculate the mean DC-MRF concentration estimates shown in the plots. The mean DC-MRF concentration estimates resulted in a significant correlation over all phantoms (Pearson Correlations: R2 > 0.998, p < 0.0001). Note also that the slopes of the correlations are nearly equal to 1 (0.988 and 0.980 for Gd and Mn, respectively) indicative of limited bias in the DC-MRF results. Mean and standard deviations of the gadolinium and manganese concentrations
were calculated from the MRF-based concentration maps for the 6 phantoms containing
both agents using an ROI analysis. The DC-MRF method resulted in significant differences
among all samples for both the gadolinium concentration estimates (two-tailed unpaired
Student’s t-test, p < 0.01) and the manganese concentration estimates (two-tailed unpaired
Student’s t-test, p < 0.001). Overall, these results suggest that the DC-MRF methodology provides accurate and precise assessments of two paramagnetic MRI contrast agents at the same time.
82
Figure 4.7: Pearson correlation plots of mean DC-MRF estimates for Gd (left) and Mn (right) concentrations against known concentrations at 3 T from phantoms containing mixtures of both contrast agents (Table 4.1 Phantoms 11-16; n=6). Concentrations were estimated using Equations 4.5 and 4.6 and mean T1 and T2 values obtained from MRF-based maps. These data are a subset of the data shown in Figure 4.6.
4.4 Discussion
In this initial report, we have demonstrated the capability of the Dual Contrast
(DC)-MRF method to simultaneously measure the concentration of two paramagnetic MRI
contrast agents. Herein, we present initial in vitro relaxometry measurements at 60 MHz to
evaluate the proposed multiple contrast agent relaxation model for two paramagnetic MRI
contrast agents (Equations 4.3 and 4.4). We also show initial in vitro DC-MRF results on
a clinical 3 T MRI scanner demonstrating the capability of DC-MRF to independently quantify the local concentration of two paramagnetic MRI contrast agents using simultaneously-measured T1 and T2 relaxation times. Overall, DC-MRF provides an
imaging platform that can be used to independently monitor multiple paramagnetic MRI
contrast agents with numerous clinical and preclinical molecular imaging applications.
We first validated the multiple contrast agent relaxation model described in
Equations 4.3 and 4.4. The significant correlations between the known phantom
concentrations and the 60 MHz relaxometric estimates (R2 > 0.998, Fig. 4.2, 4.3) as well
as the 3 T DC-MRF estimates (R2 > 0.998, Figs. 4.5, 4.6) suggests that this model is
83
accurate for the concentration ranges of the two specific MRI contrast agents
(Multihance®, gadobenate dimeglumine and MnCl2) used in this study. These results also
demonstrate the ability of the model to accurately measure contrast agent concentrations at
multiple field strengths. This straightforward linear relaxation model assumes that the two
contrast agents impact the overall T1spatial and T2 relaxation times independently with
minimal interactions between the two contrast agents. While this model may be expected
to be valid for moderate agent concentrations, high local concentrations of one or both of
the contrast agents could result in deviations from the linear model as the agents compete
for interactions with the surrounding water molecules. Implicit in this model also is the
assumption that the relaxivities (r1 and r2) are distinctly different for the two MRI contrast
agents. Therefore, this relaxation model may have limitations if used to detect two MRI
contrast agents with similar r1 and r2 values (e.g., gadopentetate dimeglumine and
gadolinium-diethylenetriamine pentaacetic acid (Gd-DTPA)). Furthermore, multiple
follow-on studies (in vitro and in vivo) will be needed to thoroughly explore the limitations of the proposed relaxation model. Regardless, these initial results demonstrate that the multiple contrast agent relaxation model can provide an analytical basis to determine the concentration of two different paramagnetic MRI contrast agents.
A key advantage of DC-MRF is that it provides the opportunity to simultaneously
and dynamically detect multiple MRI contrast agents. Prior studies have attempted to detect
multiple MRI contrast agents in a single scanning session using sequential administration
of contrast agents197; ratiometric methods to detect the presence of activatable MRI agents110, and machine learning198. Other groups have utilized chemical exchange
saturation transfer (CEST) MRI techniques which can have reduced sensitivity on low-
84
field (≤ 3 T) MRI scanners109,199,200. It is important to note that these prior studies primarily
incorporated sequential MRI assessments to separately assess the different MRI contrast
agents. Sequential T1 and T2 measurements (or any other method that results in both T1 and
T2 maps) could be used with equations 4.5 and 4.6 in place of MRF-based T1 and T2
assessments under the assumption that the concentration of the agent is not appreciably
changing during the measurement time. Difficulty with this strategy arises because this
assumption may or may not be valid based on the disease state and agents used. The ability
of the MRF data to be acquired in as little as 10 seconds provides the opportunity to
dynamically assess a wide variety of contrast agents regardless of the pharmacokinetics
making it a more general solution with fewer required assumptions. Further, prior studies
have shown that MRF is more temporally efficient than other rapid MRI techniques49,52,
and has been implemented on both clinical and preclinical MRI scanners. An additional
benefit of the simultaneous measurement of T1 and T2 provided by MRF is these two
relaxation time maps always being co-registered regardless of subject motion. This
important feature allows the pixel-wise concentration maps shown in Fig. 4.5a to be
calculated without any mismatch errors or utilization of additional co-registration
methodology. Therefore, DC-MRF may provide an adaptable, quantitative imaging
framework to assess two MRI contrast agents simultaneously for a wide variety of imaging
applications.
While these initial in vitro results suggest that DC-MRF can provide accurate
assessments of two paramagnetic MRI contrast agents, in vivo imaging studies will be
required to more fully evaluate the DC-MRF methodology. A schematic for an eventual in vivo DC-MRF implementation is shown in Fig. 4.8. These in vivo DC-MRF experiments
85
could be conducted similarly to conventional dynamic contrast enhanced MRI studies
where images are collected before, during, and after administration of a contrast
agent105,201–203. For in vivo DC-MRF, a pre-contrast MRF scan is performed to obtain baseline T1 and T2 maps (T10 and T20 in Equations 4.3 and 4.4). Both MRI contrast agents
would then be administered simultaneously as a mixture (pseudo-color orange in syringe
and tumor in Fig. 4.8) during dynamic acquisition of MRF-based T1 and T2 maps. These
inherently co-registered T1 and T2 maps would be then used with Equations 4.5 and 4.6 to
calculate in vivo concentration maps for each contrast agent (Agent A, red; Agent B, yellow; Fig. 4.8).
Figure 4.8: Proposed workflow for in vivo DC-MRF application. This workflow describes how two agents (Agent A, red; Agent B, yellow) could be applied and quantified in vivo. A baseline MRF scan is first performed to provide pre-contrast T1 and T2 relaxation time maps (T10 and T20). After the baseline scans, a solution containing a mixture of two contrast agents is injected (pseudo-color orange in syringe and tumor) while MRF-based T1 and T2 maps are dynamically collected. These T1 and T2 maps can be used in the DC-MRF framework to yield pixel-wise calculations of ion concentration resulting in maps of each individual agent using Equations 4.5 and 4.6. This process assumes that the in vivo relaxivities (r1, r2) of the two MRI contrast agents have already been determined. Importantly, there are several challenges that must be overcome to utilize the DC-
MRF methodology for in vivo experiments. First, MRF has been previously shown to be
sensitive to inhomogeneities in both the B1 and B0 fields. As variation in B1 and B0 are
86
expected to significantly increase for in vivo experiments, accurate B0 shimming and/or B1
138 corrections will be needed to avoid significant errors in T1 and T2 measurements as well
as the calculated contrast agent concentrations. Additionally, in vivo relaxivities will likely be different from the in vitro relaxivities shown here due to complex molecular interactions experienced by the contrast agents in vivo. This may be particularly complicated as these
interactions can vary considerably between normal tissues and pathologies. Addressing this
important limitation will require careful in vivo relaxivity assessments and potentially
validation using elemental analysis of excised tissues. The pre-contrast T1 and T2 relaxation
times of the tissues and pathologies of interest may also pose specific challenges. For
example, if the tissue of interest has a low T2 relaxation time before contrast administration
(e.g., liver), resolving the concentration of the two agents may be problematic due to
excessive T2 decay. In addition, in vivo studies may be impacted by multiple factors including partial volume effects, flow effect, and magnetization transfer and/or chemical exchange that may benefit from multi-exponential relaxation models instead of the simple mono-exponential models used here73,204. Despite these challenges, DC-MRF offers a
unique opportunity to expand the portfolio of in vivo contrast-enhanced MRI applications.
In addition to the need for follow-on in vivo studies, the DC-MRF results presented in this initial report have multiple areas for future exploration. As described above, the multiple contrast agent relaxation model described herein provides an analytical solution to independently quantify two MRI contrast agents. It is conceivable that DC-MRF could
be applied for three or more contrast agents. However, in that case, Equations 4.3 and 4.4
become underdetermined. As such, using the DC-MRF approach to differentiate three or
more agents would require additional measurements and/or alternative numerical or
87
machine learning198 approaches to resolve these agent concentrations accurately. As described above, another limitation of the DC-MRF methodology is the requirement for two agents with different relaxivities. In this initial proof-of-concept study, we used a clinical gadolinium agent and a manganese chloride based agent with substantial differences in r1 and r2 (Figs. 4.1 and 4.4). The minimum differences in relaxivity needed
to reliably differentiate two contrast agents remains undetermined. Reduced relaxivity
differences would also likely limit the ability of the DC-MRF methodology to detect small concentration changes. Additionally, these relaxivities are also known to change as a
function of magnetic field strength. Therefore, in vivo validation studies would likely be
required for each MRI field strength and for a variety of MRI contrast agents to explore
the limitations of DC-MRF.
Although DC-MRF may have numerous applications, this new methodology is
particularly well suited to support the development and eventual clinical translation of
molecular MRI contrast agents. As described above, DC-MRF would provide the
opportunity to detect both a molecularly-targeted contrast agent and a control non-targeted contrast agent at the same time in the same subject. The only constraint in achieving these simultaneous assessments using DC-MRF is that the two contrast agents must have different relaxivities. This constraint may require that the targeted and control contrast agents incorporate different lanthanides (e.g., gadolinium and dysprosium) in order to produce differential relaxivities while also retaining similar pharmacokinetic properties. A similar approach could also be used to allow multiple molecular imaging targets to be assessed simultaneously. For example, two targeted MRI contrast agents could be used to simultaneously track drug delivery and therapeutic impact in cancer treatment (e.g., a
88
therapeutic agent205 labeled with gadolinium, and a second agent targeted to apoptosis206
labeled with dysprosium). Importantly, the DC-MRF methodology may also directly aid in the clinical translation of molecular imaging agents by allowing for the direct comparison between a conventional, non-targeted clinical MRI contrast agent (e.g., Multihance®) with a molecular imaging agent in a single patient. In this way, DC-MRF would provide the opportunity to efficiently establish the molecular specificity of the molecular contrast agent in heterogeneous human diseases.
4.5 Conclusions In conclusion, we describe a new Dual Contrast - Magnetic Resonance
Fingerprinting technique that can be used to independently quantify the local concentration of two paramagnetic MRI contrast agents administered simultaneously. These initial in vitro results validate the proposed multiple contrast agent relaxation model and demonstrate a new quantitative imaging methodology that can be used to simultaneously generate concentration maps for two different MRI contrast agents. Overall, these results suggest a new application for the MRF technology to provide quantitative assessments that lays the foundation for numerous clinical and preclinical multi-agent imaging applications.
89
Chapter 5: Summary and Future Directions
The prior chapters describe how the MRF method could be improved and expanded
for preclinical and molecular imaging applications. In summary, the RIPE-MRF method
(Chapter 3) introduced temporal incoherence to preclinical motion artifacts effectively suppressing them during quantification and improving MRF-based preclinical
quantification. The DC-MRF method (Chapter 4) introduced and validated an extended
relaxivity model that allows direct calculation of contrast agent concentrations following
simultaneous T1 and T2 quantification. Performing simultaneous multi-property
quantification with a high temporal resolution in DC-MRF makes dynamic and quantitative
multi-agent studies possible for the first time and expands the application of MRF to
contrast-enhanced molecular MRI studies.
5.1 Future Directions for MRF in Molecular MRI
5.1.1 In vivo Application of DC-MRF
Despite compelling in vitro results, the DC-MRF method must still be validated
during in vivo studies. This will require understanding how relaxivity values change in vivo
for different MRI contrast agents. Initial in vivo experiments using the DC-MRF method
with injections of single agents at varying doses in tumor bearing mice have shown that
reasonable relaxivities for both r1 and r2 can be simultaneously estimated from in vivo MRF data (Fig. 5.1). As an example, studies with a commercially available Gd contrast agent
(gadobenate dimeglumine, MultiHance) revealed different MRF-based r2/r1 ratios in vivo
as compared to in vitro (1.5 in vitro vs. 6.8 in vivo). While promising, these preliminary
results confirm the need to obtain accurate in vivo relaxivity values74 in order to generate
90
accurate concentrations estimates for multiple MRI contrast agents using the DC-MRF technique.
In vivo MRF-based Gd Relaxivity Estimation
6 Gd-R1 5 y = 39.70x - 0.42 Gd-R2 4 R² = 0.86 (1/s) 3 1,2 y = 5.85x - 0.02 Δ R 2 R² = 0.92 1 0 0 0.025 0.05 0.075 0.1 0.125 0.15 Concentration Gd (mM)
Figure 5.1: In vivo Gd contrast agent relaxation enhancement curves. Dynamic MRF acquisitions (2-minute temporal resolution) were used to serially measure T1 and T2 after injection of gadobenate dimeglumine (MultiHance). MRI-specific tissue property enhancement (ΔR1,2=1/T1,2- 1/T10,20) in the MRF-based map immediately preceding tumor excision was compared to the tissue Gd concentration (measured via inductively coupled plasma mass spectrometry) to generate curves for estimating relaxivity. Only 7 samples were used for r2 relaxivity (compared to 16 for r1) due to the limited number of samples that resulted in reliable T2 enhancement measurements. These curves represent a preliminary assessment of in vivo relaxivity and highlight the difference between in vivo and in vitro relaxivity. Continuing the discussion from the previous paragraph, sensitivity of the MRF
experiment to T2 is another potential barrier to successful in vivo translation. DC-MRF
based concentration estimates are calculated from the relative change in both T1 and T2
after injection of the contrast agents. Since in vivo T2 values are often short (typically < 50 ms), the observable dynamic range in T2 is limited. As such, MRF measurements with high
T2 precision will be necessary to avoid T2 saturation (T2 < 10 ms) at high agent
concentrations as well as ensuring accurate concentration estimates in short T2 tissues. In
91
these situations, ultra-short echo time imaging techniques that enable detection of short T2
species may be useful55.
5.1.2 Opportunities for DC-MRF Technical Development
While the DC-MRF methodology has been shown to accurately determine the
concentration of two different MRI contrast agents, DC-MRF could potentially be used to
simultaneously estimate the concentration of three or more agents. One approach to this
would be to quantify additional MRI-specific properties that are sensitive to agent
concentration (e.g. chemical exchange97), with each additional property allowing quantification of another agent. This method would allow an unambiguous analytical calculation of agent concentration but relies on having accurate in vivo estimates of the relevant magnetic properties. Furthermore, quantification of additional properties may require more image samples or the addition of preparation blocks ultimately reducing the temporal resolution. A theoretically possible alternative approach is incorporation of the contrast agents into the dictionary simulation and considering their concentrations as properties to be estimated during quantification. Dictionary-based quantification has potentially greater flexibility in agent number but may require assumptions about pharmacokinetics and agent relaxivities for accurate estimation. Alternatively, numerical methods may be employed to estimate the properties of interest but these may introduce errors into the property estimates. Expansion to three or more agents is likely to require substantial sequence and property estimation development in order to generate accurate concentration estimates. Despite these potential challenges, the potential applications of three or more agents may make this a valuable area of investigation.
92
A potential concern for the DC-MRF method, and one this is likely to be even more concerning as more agents are added, is the possibility for magnetic interactions between
agents. The current model assumes that each agent impacts the relaxation rates
independently but this may not be the case when multiple agents are present in high
concentrations. Magnetic interactions could possibly be accounted for by adding a
multiplicative term and an interaction coefficient into the proposed model (Equations 4.3
and 4.4), but this may complicate the quantification process and require alternative
methods of estimating the agent concentrations (i.e. machine learning). This is potentially
complicated by the potential for these interactions to be agent-specific. Regardless,
understanding agent interactions will likely be important for in vivo applications where
there may be substantial agent accumulation.
5.1.3 Potential Applications of MRF for Molecular MRI
Of particular interest for future development are the opportunities for novel
molecular MRI applications using MRF. Using the DC-MRF method, fully quantitative
studies can be performed using two co-administered contrast agents potentially generating
complementary information and providing a more complete picture of the disease state.
One example of a novel, two agent DC-MRF study is co-administering two agents targeted to relevant cell and tissue markers for sub-voxel analysis of tissue fractions.
Voxel-wise DC-MRF measurements of agent concentration would allow direct comparison
of the amount of each agent present within the voxel. Comparing the concentration of an
agent targeted to tumor tissue92 to an agent targeted to healthy tissue would indicate how
much of that voxel is occupied by tumor versus healthy tissue. Alternatively, tissue
vascularization and vessel permeability could be assessed with a blood pool agent191
93
combined with a vessel permeable small molecule agent (e.g. clinical Gd agents). This type of experiment could be done with any pair of agents sensitive to the sub-voxel components of interest (e.g. myelin and nerve tissue) giving a more complete picture of the biology in each voxel.
A similar targeted agent approach could be used in the validation and imaging of genetic reporters. Modification of the genome using technology such as CRISPR/Cas9 is becoming a common research technique as well as a potential target for therapeutic intervention. Imaging a reporter gene is a non-invasive way to localize the genetic modification, track gene expression over time, and identify response to therapy. An MRI- sensitive genetic reporter (e.g., urea transporter207) will generate contrast enhancement in
MRI images related to the expression level of the chosen gene. Combining this genetic reporter with an agent targeted to the gene product would identify both gene transcription
and translation separately tracking gene expression and protein production. This strategy
would create a more comprehensive MRI-based system for genetic imaging that would be
applicable for both animal and, in the future, human studies.
Development of activatable MRI contrast agents has created a new class of agents
that, while administered alone, act as two agents in vivo and could benefit from the DC-
MRF method208. These agents provide functional information by converting between two
different relaxivity states in response to enzyme activity or the tissue microenvironment. If
the relaxivity values of the two states are known, DC-MRF could be used to measure the amounts of activated versus inactivated agent within a voxel. This ratio could then be used to measure enzyme activity209,210, pH211, or redox state94,212 within a voxel providing a
functional assessment of cellular/metabolic activity or the local tissue microenvironment.
94
While DC-MRF has numerous applications for multi-agent studies, it can also be used in single agent studies. Design and development of molecular MRI contrast agents is currently non-optimal due to the agents being tuned based on in vitro relaxivity. In vitro testing uses blood and plasma64 to simulate in vivo application but these may not completely reflect the tissue environment of interest and may not give optimal performance in vivo. It would be useful to measure the r2/r1 ratio in situ so agents could be designed and optimized based on the target application. The simultaneous multi-property estimation provided by MRF allows relaxation enhancement to serve as a proxy for the relaxivity ratio r2/r1:
= + [ ] (5.1)
𝑅𝑅1 = 𝑅𝑅10 + 𝑟𝑟1𝐴𝐴[𝐴𝐴] (5.2) Rearranging terms: 𝑅𝑅2 𝑅𝑅20 𝑟𝑟2𝐴𝐴 𝐴𝐴 = [ ] (5.3)
𝑅𝑅1 − 𝑅𝑅10 = 𝑟𝑟1𝐴𝐴[𝐴𝐴] (5.4) 2 20 2𝐴𝐴 𝑅𝑅 − 𝑅𝑅 =𝑟𝑟 𝐴𝐴 (5.5) 𝑅𝑅2 − 𝑅𝑅20 𝑟𝑟2𝐴𝐴 where R1,2=1/T1,2. While Equation 5.5𝑅𝑅 1doesn’t− 𝑅𝑅10 result𝑟𝑟1𝐴𝐴 in absolute quantification, it provides a simple method for estimating the in vivo relaxivity ratios which can be used for agent design and development.
5.2 Future Directions for High-Field Magnetic Resonance Fingerprinting
5.2.1 Quantitative Opportunities in High-Field Preclinical MRF
Gao et. al. successfully implemented MRF on preclinical scanners137, and it has since been accelerated using undersampled spiral acquisitions led by Dr. Xin Yu’s group.
Currently, there are several opportunities that should be explored to improve the
95
quantitative accuracy and precision of preclinical MRF. This work, similar to other
preclinical MRF applications137, naïvely implemented the MRF acquisition using the TR
and FA patterns proposed in the original MRF methods on human scanners136,144. It is likely
that the different imaging conditions on high-field scanners, including longer T1 values,
shorter T2*, will require different acquisition sequence parameters to optimize the measurement of the MRI-specific tissue properties of interest (e.g., T1, T2). Modifications
may include different FA and TR ranges/patterns in addition to fixed, short echo times.
These optimized acquisitions will also likely have a positive effect on any DC-MRF concentration estimates through improved quantification of T1 and T2.
While this work exclusively quantified T1 and T2, there are opportunities to expand
the quantitative capabilities of preclinical MRF. A property that would likely not require
extensive sequence modification or have a deleterious impact on T1 and T2 quantification
is T1ρ. This property describes the relaxation of magnetization in the rotating reference
frame during spin-lock and is frequently used to analyze the collagen content and structure
in cartilage and other fibrous tissues115. The adiabatic inversion pulses (spin-lock pulses)
frequently used during MRF to enhance T1 sensitivity could also be used as T1ρ preparation
pulses. Adding additional inversion pulses with varying inversion times would introduce
sensitivity to T1ρ that could evolve over the course of the MRF experiment. The unique
relaxation mechanism of T1ρ makes it a natural addition to the MRF experiment with
theoretically little negative impact on T1 and T2 quantification.
Diffusion during high-field preclinical MRF experiments presents both a challenge
to accurate quantification of T2 and an opportunity to quantify another valuable MRI-
specific tissue property. The large unbalanced gradient moments used in preclinical
96
imaging introduce diffusion sensitivity to the sequence which may result in substantial
signal attenuation146. If is not included in the dictionary simulation, diffusion effects will
appear similar to T2 relaxation and MRF will report a shorter “apparent” T2 value.
Alternatively, diffusion effects can be included in the dictionary and matched during
quantification. The challenge with this strategy is determining the amount of diffusion
weighting in the sequence and then accurately simulating the diffusion effects. Despite the
potential challenges, diffusion serves as an important disease marker115,213–215 and it would be a valuable addition to preclinical MRF studies.
5.2.2 Technical Developments for High-Field Preclinical MRF
A challenge unique to preclinical MRF is the rapid and periodic motion of animals under anesthesia during imaging. While handling the artifacts has already been described, this motion may also spoil the transverse magnetization by causing additional spin dephasing effectively serving as an additional T2 decay mechanism. Motion also
complicates MRF because it interrupts the coherent buildup of signal over time causing the
profile to deviate from the simulated dictionary profiles. A gated/triggered acquisition that
only acquires data during the quiescent period of motion or simulation of motion in the
dictionary could be used to minimize the effects of motion on the MRF scan. While gating
would increase scan time and both methods would require simulation of a new dictionary
for every acquisition149, the potential benefit of improved MRI property quantification
makes this a valuable area of exploration.
An opportunity to improve upon the work presented here is acceleration of the
Cartesian-based sampling of the RIPE-MRF method. Compressed sensing reconstructions
are a promising way for generating alias-free maps from undersampled Cartesian images
97
allowing acceleration factors of R=8 and greater to be achieved187. This would reduce the
45-minute scan time of RIPE-MRF to roughly 6 minutes or less making it possible to use
this method with dynamic contrast enhanced scans. While compressed sensing would
introduce additional computational complexity, it would allow for accelerated MRF
acquisitions in situations where non-Cartesian trajectories are challenging and/or not
available. Additionally, compressed sensing methods typically use variable k-space
acquisitions which complements the RIPE-MRF acquisition framework and would likely provide further motion artifact suppression.
5.3 Conclusion
Magnetic resonance imaging is an incredibly valuable imaging modality for both clinical and preclinical applications. The Magnetic Resonance Fingerprinting method is a wholly new method for quantitative MRI enabling rapid quantification of multiple MRI- specific tissue properties. Its application to high-field preclinical imaging is complicated by physiological motion but this can be overcome by using the RIPE-MRF method for motion artifact suppression. In addition, the multi-property quantification of MRF can be exploited for molecular imaging studies. The DC-MRF methods allows unambiguous quantification of two paramagnetic MRI contrast agents creating novel opportunities for fully quantitative multi-agent studies. Combined, these methods create the opportunity for
MRF to expand the capabilities and applications of molecular MRI for both clinical and preclinical applications.
98
Bibliography 1. Bloch, F. Nuclear induction. Phys. Rev. 70, 460–474 (1946). 2. Torrey, H. C. Bloch equations with diffusion terms. Phys. Rev. 104, 563–565 (1956). 3. Ogawa, S. & Lee, T. Brain magnetic resonance imaging with contrast dependent on blood oxygenation. Proc. Natl. Acad. Sci. 87, 9868–9872 (1990). 4. Haacke, E. M., Brown, R. W., Thompson, M. R. & Venkatesan, R. Magnetic Resonance Imaging Physical Principles and Sequence Design. (John Wiley & Sons, Inc., 1999). 5. Twieg, D. B. The k-trajectory Formulation of the NMR Imaging Process with Applications in Analysis and Synthesis of Imaging Methods. Medical Physics 10, 610–621 (1983). 6. Stehling, M. K., Turner, R. & Mansfield, P. Echo-planar imaging: Magnetic resonance imaging in a fraction of a second. Science (80-. ). 254, 43–50 (1991). 7. Noll, D. C. Multishot rosette trajectories for spectrally selective mr imaging. IEEE
99
Trans. Med. Imaging 16, 372–377 (1997). 8. Pipe, J. P. Motion correction with PROPELLER MRI: Application to head motion and free-breating cardiac imaging. Magn. Reson. Med. 42, 963–969 (1999). 9. Ahn, C. B., Kim, J. H. & Cho, Z. H. High-Speed Spiral-Scan Echo Planar NMR Imaging-I. IEEE Trans. Med. Imaging 5, 2–7 (1986). 10. Glover, G. H. & Pauly, J. M. Projection Reconstruction Techniques for Reduction of Motion Effects in MRI. Magn. Reson. Med. 28, 275–289 (1992). 11. Jackson, J. I., Meyer, C. H. & Nishimura, D. G. Selection of a Convolution Function for Fourier Inversion Using Gridding. IEEE Trans. Med. Imaging 10, 473–478 (1991). 12. Pipe, J. G. & Menon, P. Sampling density compensation in MRI: Rationale and an iterative numerical solution. Magn. Reson. Med. 41, 179–186 (1999). 13. Fessler, J. A. & Sutton, B. P. Nonuniform Fast Fourier Transforms Using Min- Max Interpolation. IEEE Trans. Signal Process. 51, 560–574 (2003). 14. Wright, K. L., Hamilton, J. I., Griswold, M. a., Gulani, V. & Seiberlich, N. Non- Cartesian parallel imaging reconstruction. J. Magn. Reson. Imaging 1040, 1022– 1040 (2014). 15. Castets, C. R. et al. Fast 3D ultrashort echo-time spiral projection imaging using golden-angle: A flexible protocol for in vivo mouse imaging at high magnetic field. Magn. Reson. Med. 77, 1831–1840 (2016). 16. Gai, N. D., Malayeri, A. A. & Bluemke, D. A. Three-dimensional T1 and T2* mapping of human lung parenchyma using interleaved saturation recovery with dual echo ultrashort echo time imaging (ITSR-DUTE). J. Magn. Reson. Imaging 45, 1097–1104 (2017). 17. Castets, C. R. et al. Fast and robust 3D T1 mapping using spiral encoding and steady RF excitation at 7T: Application to cardiac manganese enhanced MRI (MEMRI) in mice. NMR Biomed. 28, 881–889 (2015). 18. Griswold, M. A. et al. Generalized Autocalibrating Partially Parallel Acquisitions (GRAPPA). Magn. Reson. Med. 47, 1202–1210 (2002). 19. Pruessmann, K. P., Weiger, M., Scheidegger, M. B. & Boesiger, P. SENSE: Sensitivity Encoding for Fast MRI. Magn. Reson. Med. 42, 952–962 (1999). 20. Hargreaves, B. Rapid gradient-echo imaging. J. Magn. Reson. Imaging 36, 1300– 1313 (2012). 21. Hahn, E. L. Spin echoes. Physical Review 80, 580–594 (1950). 22. Jung, B. A. & Weigel, M. Spin echo magnetic resonance imaging. J. Magn. Reson. Imaging 37, 805–17 (2013). 23. Haase, A., Frahm, J., Matthaei, D., Hanicke, W. & Merboldt, K. D. FLASH imaging. Rapid NMR imaging using low flip-angle pulses. J. Magn. Reson. 67, 258–266 (1986). 24. Carr, H. Y. Steady-state free precession in nuclear magnetic resonance. Physical Review 112, 1693–1701 (1958). 25. Bieri, O. & Scheffler, K. Fundamentals of balanced steady state free precession MRI. J. Magn. Reson. Imaging 38, 2–11 (2013). 26. Scheffler, K. & Lehnhardt, S. Principles and applications of balanced SSFP techniques. Eur. Radiol. 13, 2409–2418 (2003). 27. Oppelt, A. et al. FISP — a new fast MRI sequence. Electromedica 54, 15–18
100
(1986). 28. Kraff, O. & Quick, H. H. 7T: Physics, safety, and potential clinical applications. J. Magn. Reson. Imaging 46, 1573–1589 (2017). 29. Wei, Z. et al. Quantitative assessment of cerebral venous blood T 2 in mouse at 11.7T: Implementation, optimization, and age effect. Magn. Reson. Med. 0, 2–9 (2017). 30. Marzola, P., Osculati, F. & Sbarbati, A. High field MRI in preclinical research. Eur. J. Radiol. 48, 165–170 (2003). 31. Zhang, H. et al. Improve myocardial T1 measurement in rats with a new regression model: Application to myocardial infarction and beyond. Magn. Reson. Med. 72, 737–748 (2013). 32. Does, M. D. Inferring brain tissue composition and microstructure via MR relaxometry. Neuroimage 1–13 (2018). doi:10.1016/j.neuroimage.2017.12.087 33. Bottomley, P. A., Hardy, C. J., Ardersinger, R. E. & Allen-Moore, G. A Review of 1H Nuclear Magentic Resonance Relaxation in Pathology: Are T1 and T2 Diagnostic? Med. Phys. 14, 1–37 (1987). 34. Carr, H. Y. & Purcell, E. M. Effects of Diffusion on Free Precession in Nuclear Magnetic Resonance Experiments. Phys. Rev. 94, 630–638 (1954). 35. Meiboom, S. & Gill, D. Modified spin-echo method for measuring nuclear relaxation times. Rev. Sci. Instrum. 29, 688–691 (1958). 36. Deichmann, R. & Haase, a. Quantification of T1 values by SNAPSHOT-FLASH NMR imaging. Journal of Magnetic Resonance (1969) 96, 608–612 (1992). 37. Drain, L. E. A Direct Method of Measuring Nuclear Spin-Lattice Relaxation Times. Proc. Phys. Soc. Sect. A 62, 301–306 (1949). 38. Li, W., Griswold, M. & Yu, X. Rapid T1 mapping of mouse myocardium with saturation recovery look-locker method. Magn. Reson. Med. 64, 1296–1303 (2010). 39. Look, D. C. & Locker, D. R. Time saving in measurement of NMR and EPR relaxation times. Rev. Sci. Instrum. 41, 250–251 (1970). 40. Kaptein, R., Dijkstra, K. & Tarr, C. . A single-scan fourier transform method for measuring spin-lattice relaxation times. J. Magn. Reson. 24, 295–300 (1976). 41. Zurek, M. et al. Accurate T1 mapping for oxygen-enhanced MRI in the mouse lung using a segmented inversion-recovery ultrashort echo-time sequence. Magn. Reson. Med. 71, 2180–2185 (2014). 42. Christensen, K. A., Grant, D. M., Schulman, E. M. & Walling, C. Optimal determination of relaxation times of fourier transform nuclear magnetic resonance. Determination of spin-lattice relaxation times in chemically polarized species. J. Phys. Chem. 78, 1971–1977 (1974). 43. Deoni, S. C. L., Rutt, B. K. & Peters, T. M. Rapid combined T1 and T2 mapping using gradient recalled acquisition in the steady state. Magn. Reson. Med. 49, 515– 526 (2003). 44. Giri, S. et al. T2 quantification for improved detection of myocardial edema. J. Cardiovasc. Magn. Reson. 11, 56 (2009). 45. Deichmann, R. et al. Fast T2-Mapping With SNAPSHOT FLASH Imaging. Magn. Reson. Imaging 13, 633–639 (1995). 46. Yu, H. et al. Application of texture analysis on parametric T1 and T2 maps for
101
detection of hepatic fibrosis. J. Magn. Reson. Imaging (2016). doi:10.1002/jmri.25328 47. Gao, Y. et al. Initial evaluation of hepatic T1 relaxation time as an imaging marker of liver disease associated with autosomal recessive polycystic kidney disease (ARPKD). NMR Biomed. 29, 84–89 (2016). 48. Akçakaya, M. et al. Joint myocardial T1 and T2 mapping using a combination of saturation recovery and T2-preparation. Magn. Reson. Med. 76, 888–896 (2016). 49. Deoni, S. C. L., Peters, T. M. & Rutt, B. K. High-resolution T1 and T2 mapping of the brain in a clinically acceptable time with DESPOT1 and DESPOT2. Magn. Reson. Med. 53, 237–241 (2005). 50. Hueper, K. et al. T1-mapping for assessment of ischemia-induced acute kidney injury and prediction of chronic kidney disease in mice. Eur. Radiol. 24, 2252– 2260 (2014). 51. Franke, M. et al. Magnetic resonance T2 mapping and diffusion-weighted imaging for early detection of cystogenesis and response to therapy in a mouse model of polycystic kidney disease. Kidney Int. 92, 1544–1554 (2017). 52. Schmitt, P. et al. Inversion Recovery TrueFISP: Quantification of T1, T2, and Spin Density. Magn. Reson. Med. 51, 661–667 (2004). 53. Triphan, S. M. F., Breuer, F. A., Gensler, D., Kauczor, H. U. & Jakob, P. M. Oxygen enhanced lung MRI by simultaneous measurement of T1 and T2* during free breathing using ultrashort TE. J. Magn. Reson. Imaging 41, 1708–1714 (2014). 54. Nataraj, G., Nielsen, J. F. & Fessler, J. A. Optimizing MR Scan Design for Model- Based T1, T2 Estimation from Steady-State Sequences. IEEE Trans. Med. Imaging 36, 467–477 (2017). 55. Guo, J., Cao, X., Cleveland, Z. I. & Woods, J. C. Murine pulmonary imaging at 7T: T2* and T 1 with anisotropic UTE. Magn. Reson. Med. 0, (2017). 56. Heule, R., Celicanin, Z., Kozerke, S. & Bieri, O. Simultaneous multislice triple- echo steady-state (SMS-TESS) T1 , T2 , PD, and off-resonance mapping in the human brain. Magn. Reson. Med. (2018). doi:10.1002/mrm.27126 57. Shcherbakova, Y., van den Berg, C. A. T., Moonen, C. T. W. & Bartels, L. W. PLANET: An Ellipse Fitting Approach for simultaneous T1 and T2 mapping using phase-cycled balanced steady-state free precession. Magn. Reson. Med. 79, 711– 722 (2018). 58. Lehman, C. D. et al. MRI Evaluation of the Contralateral Breast in Women with Recently Diagnosed Breast Cancer. N. Engl. J. Med. 356, 1295–1303 (2007). 59. Liu, X., Zou, L., Liu, F., Zhou, Y. & Song, B. Gadoxetic Acid Disodium- Enhanced Magnetic Resonance Imaging for the Detection of Hepatocellular Carcinoma: A Meta-Analysis. PLoS One 8, (2013). 60. Kushnirsky, M. et al. Time-delayed contrast-enhanced MRI improves detection of brain metastases and apparent treatment volumes. J. Neurosurg. 124, 489–95 (2016). 61. Lauffer, R. B. Paramagnetic metal complexes as water proton relaxation agents for NMR imaging: theory and design. Chem. Rev. 87, 901–927 (1987). 62. Caravan, P., Ellison, J. J., McMurry, T. J. & Lauffer, R. B. Gadolinium(III) chelates as MRI contrast agents: Structure, dynamics, and applications. Chem. Rev.
102
99, 2293–2352 (1999). 63. Wood, M. L. & Hardy, P. A. Proton relaxation enhancement. J. Magn. Reson. Imaging 3, 149–156 (1993). 64. Rohrer, M., Bauer, H., Mintorovitch, J., Requardt, M. & Weinmann, H.-J. Comparison of Magnetic Properties of MRI Contrast Media Solutions at Different Magnetic Field Strengths. Invest. Radiol. 40, 715–724 (2005). 65. Fringuello Mingo, A. et al. Macrocyclic paramagnetic agents for MRI: Determinants of relaxivity and strategies for their improvement. Magn. Reson. Med. 78, 1523–1532 (2017). 66. Caravan, P., Farrar, C., Frullano, L. & Uppal, R. Influence of molecular parameters and increasing magnetic field strength on relaxivity of gadolinium- and manganese-based T1 contrast agents. Contrast Media Mol. Imaging 4, 89–100 (2009). 67. Marangoni, V. S. et al. Enhancing T 1 magnetic resonance imaging contrast with internalized gadolinium(III) in a multilayer nanoparticle. Proc. Natl. Acad. Sci. 114, 6960–6965 (2017). 68. Soesbe, T. C. et al. Maximizing T2-exchange in Dy3+DOTA-(amide) X chelates: Fine-tuning the water molecule exchange rate for enhanced T2 contrast in MRI. Magn. Reson. Med. 71, 1179–1185 (2014). 69. Caravan, P., Greenfield, M. T. & Bulte, J. W. M. Molecular Factors That Determine Curie Spin Relaxation in Dysprosium Complexes. Magn. Reson. Med. 46, 917–922 (2001). 70. Vander Elst, L. et al. Dy-Complexes as High Field T2 Contrast Agents. Acad. Radiol. 9, S297–S299 (2002). 71. Henrotte, V., Muller, R. N., Bartholet, A. & Vander Elst, L. The presence of halide salts influences the non-covalent interaction of MRI contrast agents and human serum albumin. Contrast Media Mol. Imaging 2, 258–261 (2007). 72. Henoumont, C., Vander Elst, L., Laurent, S. & Muller, R. N. Study of non- covalent interactions between MRI contrast agents and human serum albumin by NMR diffusometry. JBIC J. Biol. Inorg. Chem. 14, 683–691 (2009). 73. Stanisz, G. J. & Henkelman, R. M. Gd-DTPA relaxivity depends on macromolecular content. Magn. Reson. Med. 44, 665–667 (2000). 74. Jiang, K., Tang, H., Mishra, P. K., Macura, S. I. & Lerman, L. O. A rapid T 1 mapping method for assessment of murine kidney viability using dynamic manganese-enhanced magnetic resonance imaging. Magn. Reson. Med. 80, 190– 199 (2018). 75. Schmalbrock, P., Hines, J. V., Lee, S. M., Ammar, G. M. & Kwok, E. W. C. T1 measurements in cell cultures: A new tool for characterizing contrast agents at 1.5T. J. Magn. Reson. Imaging 14, 636–648 (2001). 76. Zhang, Z. et al. Multilocus binding increases the relaxivity of protein-bound MRI contrast agents. Angew. Chemie - Int. Ed. 44, 6766–6769 (2005). 77. Aime, S., Barge, A., Botta, M., Parker, D. & De Sousa, A. S. Prototropic vs whole water exchange contributions to the solvent relaxation enhancement in the aqueous solution of a cationic Gd3+ macrocyclic complex. J. Am. Chem. Soc. 119, 4767– 4768 (1997). 78. Zhou, Z., Han, Z. & Lu, Z. R. A targeted nanoglobular contrast agent from host-
103
guest self-assembly for MR cancer molecular imaging. Biomaterials 85, 168–179 (2016). 79. Pu, F. et al. Prostate-specific membrane antigen targeted protein contrast agents for molecular imaging of prostate cancer by MRI. Nanoscale 8, 12668 (2016). 80. Li, B., Gu, Z., Kurniawan, N., Chen, W. & Xu, Z. P. Manganese-Based Layered Double Hydroxide Nanoparticles as a T1-MRI Contrast Agent with Ultrasensitive pH Response and High Relaxivity. Adv. Mater. 29, 1–8 (2017). 81. Hu, H. et al. Dysprosium-Modified Tobacco Mosaic Virus Nanoparticles for Ultra-High-Field Magnetic Resonance and Near-Infrared Fluorescence Imaging of Prostate Cancer. ACS Nano 11, 9249–9258 (2017). 82. Hermann, P., Kotek, J., Kubíček, V. & Lukeš, I. Gadolinium(III) Complexes as MRI Contrast Agents: Ligand Design and Properties of the Complexes. Dalt. Trans. 23, 3027–3047 (2008). 83. Hao, D. et al. MRI contrast agents: Basic chemistry and safety. J. Magn. Reson. Imaging 36, 1060–1071 (2012). 84. Pintaske, J. et al. Relaxivity of gadopentetate dimeglumine (Magnevist), gadobutrol (Gadovist), and gadobenate dimeglumine (MultiHance) in human blood plasma at 0.2, 1.5, and 3 Tesla. Invest. Radiol. 41, 213–221 (2006). 85. Shen, Y. et al. T1 relaxivities of gadolinium-based magnetic resonance contrast agents in human whole blood at 1.5, 3, and 7 T. Invest. Radiol. 50, 330–338 (2015). 86. Kircher, M. F. et al. A brain tumor molecular imaging strategy using a new triple- modality MRI-photoacoustic-Raman nanoparticle. Nat. Med. 18, 829–834 (2012). 87. Johansen, M. L. et al. Quantitative Molecular Imaging with a Single Gd-Based Contrast Agent Reveals Specific Tumor Binding and Retention in vivo. Anal. Chem. 89, 5932–5939 (2017). 88. Liu, X. et al. MRI contrast agent for targeting glioma: Interleukin-13 labeled liposome encapsulating gadolinium-DTPA. Neuro. Oncol. 18, 691–699 (2016). 89. Kitagawa, T. et al. RGD targeting of human ferritin iron-oxide nanoparticles enhances in vivo molecular MRI of experimental aortic aneurysms. J. Magn. Reson. Imaging 45, 1144–1153 (2016). 90. Zhou, Z. et al. MRI detection of breast cancer micrometastases with a fibronectin- targeting contrast agent. Nat. Commun. 6, (2015). 91. Li, T. et al. A New Interleukin-13 Amino-Coated Gadolinium Metallofullerene Nanoparticle for Targeted MRI Detection of Glioblastoma Tumor Cells. J. Am. Chem. Soc. 137, 7881–7888 (2015). 92. Abakumova, T. et al. Connexin 43-targeted T1 contrast agent for MRI diagnosis of glioma. Contrast Media Mol. Imaging 11, 15–23 (2016). 93. Hingorani, D. V., Bernstein, A. S. & Pagel, M. D. A review of responsive MRI contrast agents: 2005-2014. Contrast Media Mol. Imaging 10, 245–265 (2015). 94. Gale, E. M., Jones, C. M., Ramsay, I., Farrar, C. T. & Caravan, P. A Janus Chelator Enables Biochemically Responsive MRI Contrast with Exceptional Dynamic Range. J. Am. Chem. Soc. 138, 15861–15864 (2016). 95. Ni, D. et al. PEGylated NaHoF4nanoparticles as contrast agents for both X-ray computed tomography and ultra-high field magnetic resonance imaging. Biomaterials 76, 218–225 (2016).
104
96. Gordon, J. W., Fain, S. B. & Rowland, I. J. Effect of lanthanide ions on dynamic nuclear polarization enhancement and liquid-state T1 relaxation. Magn. Reson. Med. 68, 1949–1954 (2012). 97. Cakić, N. et al. Paramagnetic lanthanide chelates for multicontrast MRI. Chem. Commun. 52, 9224–9227 (2016). 98. Funk, A. M., Clavijo Jordan, V., Sherry, A. D., Ratnakar, S. J. & Kovacs, Z. Oxidative Conversion of a Europium(II)-Based T1 Agent into a Europium(III)- Based paraCEST Agent that can be Detected in vivo by Magnetic Resonance Imaging. Angew. Chemie - Int. Ed. 55, 5024–5027 (2016). 99. Green, K. N., Viswanathan, S., Rojas-Quijano, F. A., Kovacs, Z. & Sherry, A. D. Europium(III) DOTA-derivatives having ketone donor pendant arms display dramatically slower water exchange. Inorg. Chem. 50, 1648–1655 (2011). 100. Zhang, L. et al. Lanthanide-Based T2ex and CEST Complexes Provide Insights into the Design of pH Sensitive MRI Agents. Angew. Chemie - Int. Ed. 56, 16626– 16630 (2017). 101. Kattel, K. et al. A facile synthesis, in vitro and in vivo MR studies of d-glucuronic acid-coated ultrasmall Ln₂O₃ (Ln = Eu, Gd, Dy, Ho, and Er) nanoparticles as a new potential MRI contrast agent. ACS Appl. Mater. Interfaces 3, 3325–34 (2011). 102. Terreno, E., Castelli, D., Viale, A. & Aime, S. Challenges for molecular magnetic resonance imaging. 110, 3019–3042 (2010). 103. Smolensky, E. D. et al. Scaling laws at the nanosize: the effect of particle size and shape on the magnetism and relaxivity of iron oxide nanoparticle contrast agents. J. Mater. Chem. B 1, 2818 (2013). 104. Wei, H. et al. Exceedingly small iron oxide nanoparticles as positive MRI contrast agents. Proc. Natl. Acad. Sci. 114, 2325–2330 (2017). 105. Herrmann, K. et al. Dynamic quantitative T1 mapping in orthotopic brain tumor xenografts. Transl. Oncol. 9, 147–154 (2016). 106. Raghunand, N., Howison, C., Sherry, A. D., Zhang, S. & Gillies, R. J. Renal and systemic pH imaging by contrast-enhanced MRI. Magn. Reson. Med. 49, 249–257 (2003). 107. Garcia-Martin, M. L. et al. High resolution pHe imaging of rat glioma using pH- dependent relaxivity. Magn. Reson. Med. 55, 309–315 (2006). 108. Kim, K. S., Park, W., Hu, J., Bae, Y. H. & Na, K. A cancer-recognizable MRI contrast agents using pH-responsive polymeric micelle. Biomaterials 35, 337–343 (2014). 109. Hingorani, D. V. et al. A single diamagnetic catalyCEST MRI contrast agent that detects cathepsin B enzyme activity by using a ratio of two CEST signals. Contrast Media Mol. Imaging 11, 130–138 (2016). 110. Catanzaro, V. et al. A R2p/R1p ratiometric procedure to assess matrix metalloproteinase-2 activity by magnetic resonance imaging. Angew. Chemie - Int. Ed. 52, 3926–3930 (2013). 111. Daryaei, I., Jones, K. M. & Pagel, M. D. Detection of DT-diaphorase Enzyme with a ParaCEST MRI Contrast Agent. Chem. - A Eur. J. 23, 6514–6517 (2017). 112. Chouhan, M. et al. Use of Caval Subtraction 2D Phase-Contrast MR Imaging to Measure Total Liver and Hepatic Arterial Blood Flow: Preclinical Validation and Initial Clinical Translation. Radiology 280, 916–923 (2016).
105
113. Chen, W. et al. T2 mapping at 7T MRI can quantitatively assess intramyocardial hemorrhage in rats with acute reperfused myocardial infarction in vivo. J. Magn. Reson. Imaging 44, 194–203 (2016). 114. Yin, T. et al. Characterization of a rat orthotopic pancreatic head tumor model using three-dimensional and quantitative multi-parametric MRI. NMR Biomed. 30, e3676 (2017). 115. Zhang, H. et al. Comparison of T2, T1rho, and diffusion metrics in assessment of liver fibrosis in rats. J. Magn. Reson. Imaging 45, 741–750 (2017). 116. Hammer-Hansen, S. et al. Early Gadolinium Enhancement for Determination of Area at Risk: A Preclinical Validation Study. JACC Cardiovasc. Imaging 10, 130– 139 (2017). 117. van de Looij, Y. et al. Diffusion tensor imaging of diffuse axonal injury in a rat brain trauma model. NMR Biomed. 25, 93–103 (2012). 118. Pan, F. et al. The Therapeutic Response of Gastrointestinal Stromal Tumors to Imatinib Treatment Assessed by Intravoxel Incoherent Motion Diffusion-Weighted Magnetic Resonance Imaging with Histopathological Correlation. PLoS One 11, (2016). 119. Cutillo, A. G. et al. Characterization of bleomycin lung injury by nuclear magnetic resonance: Correlation between NMR relaxation times and lung water and collagen content. Magn. Reson. Med. 47, 246–256 (2002). 120. Jackson, L. H. et al. Non-invasive MRI biomarkers for the early assessment of iron overload in a humanized mouse model of β-thalassemia. Sci. Rep. 7, 1–10 (2017). 121. Fries, P. et al. Comparison of self-gated and prospectively triggered fast low angle shot (FLASH) sequences for contrast-enhanced magnetic resonance imaging of the liver at 9.4 T in a rat model of colorectal cancer metastases. Invest. Radiol. 48, 738–744 (2013). 122. Winter, P. et al. Self-navigation under non-steady-state conditions: cardiac and respiratory self-gating of inversion recovery snapshot FLASH acquisitions in mice. Magn. Reson. Med. 76, 1887–1894 (2016). 123. Yang, H. J. et al. Free-breathing, motion-corrected, highly efficient whole heart T2 mapping at 3T with hybrid radial-cartesian trajectory. Magn. Reson. Med. 75, 126– 136 (2016). 124. Bieri, O. Ultra-Fast Steady State Free Precession and Its Application to In vivo 1H Morphological and Functional Lung Imaging at 1.5 Tesla. Magn. Reson. Med. 70, 657–663 (2013). 125. Reeder, S. B., Herzka, D. A. & McVeigh, E. R. Signal-to-Noise Ratio Behavior of Steady-State Free Precession. Magn. Reson. Med. 52, 123–130 (2004). 126. Duerk, J. L., Lewin, J. S., Wendt, M. & Petersilge, C. Remember True FISP? A high SNR, Near 1-Second Imaging Method for T2-Like Contrast in Interventional MRI at .2 T. J. Magn. Reson. Imaging 8, 203–208 (1998). 127. Vasanawala, S. S., Pauly, J. M. & Nishimura, D. G. Linear Combination Steady- State Free Precession MRI. Magn. Reson. Med. 43, 82–90 (2000). 128. Buonocore, M. H. & Gao, L. Ghost artifact reduction for echo planar imaging using image phase correction. Magn. Reson. Med. 38, 89–100 (1997). 129. Brodsky, E. K., Klaers, J. L., Samsonov, A. A., Kijowski, R. & Block, W. F. Rapid Measurement and Correction of Phase Errors from B0 Eddy Currents: Impact on
106
Image Quality for Non-Cartesian Imaging. Magn. Reson. Med. 69, 509–515 (2013). 130. Collins, C. M., Liu, W., Schreiber, W., Yang, Q. X. & Smith, M. B. Central Bightening Due to Constructive Interference With, Without, and Despite Dielectric Resonance. J. Magn. Reson. Imaging 21, 192–196 (2005). 131. Fleysher, R., Fleysher, L., Liu, S. & Gonen, O. TriTone: a Radiofrequency Field (B1)-Insensitive T1 Estimator for MRI at High Magnetic Fields. Magn. Reson. Imaging 26, 781–789 (2008). 132. Sled, J. G. & Pike, G. B. Correction for B1 and B0 Variations in Quantitative T2 Measurements Using MRI. Magn. Reson. Med. 43, 589–593 (2000). 133. Sacolick, L. I., Wiesinger, F., Hancu, I. & Vogel, M. W. B1 Mapping by Bloch- Siegert Shift. Magn. Reson. Med. 63, 1315–1322 (2010). 134. Boudreau, M. et al. B1 Mapping for Bias-Correction in Quantitative T1 Imaging of the Brain at 3T Using Standard Pulse Sequences. J. Magn. Reson. Imaging 46, 1673–1682 (2017). 135. Katscher, U. & Bornert, P. Parallel RF transmission in MRI. NMR Biomed. 19, 393–400 (2006). 136. Ma, D. et al. Magnetic resonance fingerprinting. Nature 495, 187–192 (2013). 137. Gao, Y. et al. Preclinical MR fingerprinting (MRF) at 7 T: effective quantitative imaging for rodent disease models. NMR Biomed. 28, 384–394 (2015). 138. Ma, D. et al. Slice profile and B1 corrections in 2D magnetic resonance fingerprinting. Magn. Reson. Med. (2017). doi:10.1002/mrm.26580 139. Sommer, K. et al. Towards predicting the encoding capability of MR fingerprinting sequences. Magn. Reson. Imaging 41, 7–14 (2017). 140. Ostenson, J., Robison, R. K., Zwart, N. R. & Welch, E. B. Multi-frequency interpolation in spiral magnetic resonance fingerprinting for correction of off- resonance blurring. Magn. Reson. Imaging 41, 63–72 (2017). 141. McNab, J. A. & Miller, K. L. Sensitivity of diffusion weighted steady state free precession to anisotropic diffusion. Magn. Reson. Med. 60, 405–413 (2008). 142. Scheffler, K. & Hennig, J. Is TrueFISP a gradient-echo or a spin-echo sequence? Magn. Reson. Med. 49, 395–397 (2003). 143. Kim, D., Cernicanu, A. & Axel, L. B0 and B1-Insensitive Uniform T1-Weighting for Quantitative, First-Pass Myocardial Perfusion Magnetic Resonance Imaging. Magn. Reson. Med. 54, 1423–1429 (2005). 144. Jiang, Y., Ma, D., Seiberlich, N., Gulani, V. & Griswold, M. a. MR fingerprinting using fast imaging with steady state precession (FISP) with spiral readout. Magn. Reson. Med. 74, 1621–1631 (2015). 145. Rieger, B., Zimmer, F., Zapp, J., Weingärtner, S. & Schad, L. R. Magnetic resonance fingerprinting using echo-planar imaging: Joint quantification of T1 and T2* relaxation times. Magn. Reson. Med. 75, 1724–1733 (2017). 146. McNab, J. A. & Miller, K. L. Steady-state diffusion-weighted imaging: Theory, acquisition and analysis. NMR Biomed. 23, 781–793 (2010). 147. Wu, E. X. & Buxton, R. B. Effect of diffusion on the steady-state magnetization with pulsed field gradients. J. Magn. Reson. 90, 243–253 (1990). 148. Su, P. et al. Multiparametric estimation of brain hemodynamics with MR fingerprinting ASL. Magn. Reson. Med. 78, 1812–1823 (2017).
107
149. Hamilton, J. I. et al. MR fingerprinting for rapid quantification of myocardial T1, T2, and proton spin density. Magn. Reson. Med. 77, 1446–1459 (2016). 150. Chen, Y. et al. MR fingerprinting for rapid quantitative abdominal imaging. Radiology 276, 278–286 (2016). 151. Kellman, P., Aletras, A. H., Mancini, C., McVeigh, E. R. & Arai, A. E. T2- Prepared SSFP Improves Diagnostic Confidence in Edema Imaging in Acute Myocardial Infarction Compared to Turbo Spin Echo. Magn. Reson. Med. 57, 891–897 (2007). 152. Stejskal, E. O. & Tanner, J. E. Spin Diffusion Measurements: Spin Echoes in the Presence of a Time-Dependent Field Gradient. J. Chem. Phys. 42, 288–292 (1965). 153. Buonincontri, G. & Sawiak, S. J. MR fingerprinting with simultaneous B1 estimation. Magn. Reson. Med. 76, 1127–1135 (2016). 154. Cloos, M. A. et al. Multiparametric imaging with heterogeneous radiofrequency fields. Nat. Commun. 7, 12445 (2016). 155. Bammer, R. et al. Analysis and generalized correction of the effect of spatial gradient field distortions in diffusion-weighted imaging. Magn. Reson. Med. 50, 560–569 (2003). 156. Addy, N. O., Wu, H. H. & Nishimura, D. G. A Simple Method for MR Gradient System Characterization and k-Space Trajectory Estimation. Magn. Reson. Med. 68, 120:129 (2013). 157. Le Bihan, D., Turner, R. & Macfall, J. R. Effects of Intravoxel Incoherent Motions ( IVIM ) in Steady-State Free Precession ( SSFP ) Imaging : Application to Molecular Diffusion Imaging. Magn. Reson. Med. 10, 324–337 (1989). 158. McGivney, D. F. et al. SVD compression for magnetic resonance fingerprinting in the time domain. IEEE Trans. Med. Imaging 33, 2311–2322 (2014). 159. Cauley, S. F. et al. Fast group matching for MR fingerprinting reconstruction. Magn. Reson. Med. 74, 523–528 (2014). 160. Yang, M. et al. Low rank approximation methods for MR fingerprinting with large scale dictionaries. Magn. Reson. Med. 79, 2392–2400 (2018). 161. Lattanzi, R., Zhang, B., Knoll, F., Assländer, J. & Cloos, M. A. Phase unwinding for dictionary compression with multiple channel transmission in magnetic resonance fingerprinting. Magn. Reson. Imaging 49, 32–38 (2018). 162. Ma, D. et al. Music-based magnetic resonance fingerprinting to improve patient comfort during MRI examinations. Magn. Reson. Med. 75, 2303–2314 (2016). 163. Zhao, B. et al. Improved magnetic resonance fingerprinting reconstruction with low-rank and subspace modeling. Magn. Reson. Med. 79, 933–942 (2018). 164. Ye, H. et al. Simultaneous multislice magnetic resonance fingerprinting (SMS- MRF) with direct-spiral slice-GRAPPA (ds-SG) reconstruction. Magn. Reson. Med. 77, 1966–1974 (2017). 165. Yamaguchi, M. et al. Artifact-reduced simultaneous MRI of multiple rats with liver cancer using PROPELLER. J. Magn. Reson. Imaging 38, 225–230 (2013). 166. Lemasson, B. et al. MR Vascular Fingerprinting in Stroke and Brain Tumors Models. Sci. Rep. 6, 37071 (2016). 167. Ramasawmy, R. et al. Hepatic arterial spin labelling MRI: an initial evaluation in mice. NMR Biomed. 28, 272–280 (2015). 168. Naresh, N. K. et al. Accelerated dual-contrast first-pass perfusion MRI of the
108
mouse heart: Development and application to diet-induced obese mice. Magn. Reson. Med. 73, 1237–1245 (2015). 169. Parzy, E., Miraux, S., Franconi, J. M. & Thiaudière, E. In vivo quantification of blood velocity in mouse carotid and pulmonary arteries by ECG-triggered 3D time-resolved magnetic resonance angiography. NMR Biomed. 22, 532–537 (2009). 170. Herrmann, K. H., Pfeiffer, N., Krumbein, I., Herrmann, L. & Reichenbach, J. R. MRI compatible small animal monitoring and trigger system for whole body scanners. Z. Med. Phys. 24, 55–64 (2014). 171. Chow, A. M. et al. Measurement of liver T₁ and T₂ relaxation times in an experimental mouse model of liver fibrosis. J. Magn. Reson. Imaging 36, 152–158 (2012). 172. Anderson, C. et al. Regularly incremented phase encoding magnetic resonance fingerprinting ( RIPE-MRF ) for enhanced motion suppression in cartesian MRF. Proc. 25th Annu. Meet. ISMRM, Honolulu, Hawaii, USA 5307 (2017). 173. Ye, H. et al. Accelerating magnetic resonance fingerprinting (MRF) using t- blipped simultaneous multislice (SMS) acquisition. Magn. Reson. Med. 75, 2078– 2085 (2016). 174. Epstein, F. H. & Arai, A. E. Optimization of fast cardiac imaging using an echo- train readout. J. Magn. Reson. Imaging 11, 75–80 (2000). 175. Nguyen, T. D., Ding, G., Watts, R. & Wang, Y. Optimization of view ordering for motion artifact suppression. Magn. Reson. Imaging 19, 951–957 (2001). 176. Maclaren, J. R., Bones, P. J., Millane, R. P. & Watts, R. A modified view ordering for artifact reduction in MRI. Annu. Int. Conf. IEEE Eng. Med. Biol. - Proc. 2053– 2056 (2007). doi:10.1109/IEMBS.2007.4352724 177. Wilman, A. H. & Riederer, S. J. Performance of an elliptical centric view order for signal enhancement and motion artifact suppression in breath-hold three- dimensional gradient echo imaging. Magn. Reson. Med. 38, 793–802 (1997). 178. Haacke, E. M. & Patrick, J. L. Reducing motion artifacts in two-dimensional Fourier transform imaging. Magn. Reson. Imaging 4, 359–376 (1986). 179. Tibiletti, M. et al. Respiratory self-gated 3D UTE for lung imaging in small animal MRI. Magn. Reson. Med. 78, 739–745 (2017). 180. Ribot, E. J. et al. Self-gated bSSFP sequences to detect iron-labeled cancer cells and/or metastases in vivo in mouse liver at 7 Tesla. J. Magn. Reson. Imaging 41, 1413–1421 (2015). 181. Campbell-Washburn, A. E. et al. Cardiac arterial spin labeling using segmented ECG-gated Look-Locker FAIR: Variability and repeatability in preclinical studies. Magn. Reson. Med. 69, 238–247 (2013). 182. Vandsburger, M. H. et al. Improved arterial spin labeling after myocardial infarction in mice using cardiac and respiratory gated look-locker imaging with fuzzy C-means clustering. Magn. Reson. Med. 63, 648–657 (2010). 183. Pandit, P., Qi, Y., Story, J., King, K. F. & Johnson, G. A. Multishot PROPELLER for high-field preclinical MRI. Magn. Reson. Med. 64, 47–53 (2010). 184. Messroghli, D. R. et al. Small Animal Look-Locker Inversion Recovery (SALLI) for Simultaneous Generation of Cardiac T1 Maps and Cine and Inversion Recovery-prepared Images at High Heart Rates: Initial Experience. Radiology 261,
109
258–265 (2011). 185. Vautier, J., Heilmann, M., Walczak, C., Mispelter, J. & Volk, A. 2D and 3D radial multi-gradient-echo DCE MRI in murine tumor models with dynamic R2*- corrected R1 mapping. Magn. Reson. Med. 64, 313–318 (2010). 186. Beaumont, M., Lamalle, L., Segebarth, C. & Barbier, E. L. Improved k-space trajectory measurement with signal shifting. Magn. Reson. Med. 58, 200–205 (2007). 187. Davies, M., Puy, G., Vandergheynst, P. & Wiaux, Y. A compressed sensing framework for magnetic resonance fingerprinting. SIAM J. Imaging Sci. 7, 2623– 2656 (2014). 188. Padhani, A. R. et al. Dynamic contrast enhanced MRI of prostate cancer: Correlation with morphology and tumour stage, histological grade and PSA. Clin. Radiol. 55, 99–109 (2000). 189. Herborn, C. U. et al. Comprehensive Time-Resolved MRI of Peripheral Vascular Malformations. Am. J. Roentgenol. 181, 729–735 (2003). 190. Lima, J. A. C. et al. Regional Heterogeneity of Human Myocardial Infarcts Demonstrated by Contrast-Enhanced MRI. Circulation 92, 1117–1125 (1995). 191. Bock, J. et al. 4D phase contrast MRI at 3 T: Effect of standard and blood-pool contrast agents on SNR, PC-MRA, and blood flow visualization. Magn. Reson. Med. 63, 330–338 (2010). 192. Reimer, P., Schneider, G. & Schima, W. Hepatobiliary contrast agents for contrast- enhanced MRI of the liver: Properties, clinical development and applications. Eur. Radiol. 14, 559–578 (2004). 193. Fuchs, B. C. et al. Molecular MRI of collagen to diagnose and stage liver fibrosis. J. Hepatol. 59, 992–998 (2013). 194. Assländer, J., Glaser, S. J. & Hennig, J. Pseudo Steady-State Free Precession for MR-Fingerprinting. Magn. Reson. Med. 77, 1151–1161 (2016). 195. Majumdar, S., Zoghbi, S., Pope, C. F. & Gore, J. C. Quantitation of MR Relaxation Effects of Iron Oxide Particles in Liver and Spleen. Radiology 169, 653–658 (1988). 196. Nofiele, J. T. & Cheng, H. L. M. Ultrashort Echo Time for Improved Positive- Contrast Manganese-Enhanced MRI of Cancer. PLoS One 8, (2013). 197. Kubaska, S., Sahani, D. V., Saini, S., Hahn, P. F. & Halpern, E. Dual contrast enhanced magnetic resonance imaging of the liver with superparamagnetic iron oxide followed by gadolinium for lesion detection and characterization. Clin. Radiol. 56, 410–415 (2001). 198. Hung, A. H., Lilley, L. M., Hu, F., Harrison, V. S. R. & Meade, T. J. Magnetic barcode imaging for contrast agents. Magn. Reson. Med. 77, 970–978 (2017). 199. Ali, M. M., Liu, G., Shah, T., Flask, C. A. & Pagel, M. D. Using Two Chemical Exchange Saturation Transfer Magnetic Resonance Imaging Contrast Agents for Molecular Imaging Studies. Acc. Chem. Res. 42, 915–924 (2009). 200. Fernández-Cuervo, G., Sinharay, S. & Pagel, M. D. A CatalyCEST MRI Contrast Agent that Can Simultaneously Detect Two Enzyme Activities. ChemBioChem 17, 383–387 (2016). 201. Towner, R. A. et al. In vivo detection of c-Met expression in a rat C6 glioma model. J. Cell. Mol. Med. 12, 174–186 (2008).
110
202. Ingrisch, M. et al. Quantification of Perfusion and Permeability in Multiple Sclerosis. Invest. Radiol. 47, 252–258 (2012). 203. Taheri, S., Gasparovic, C., Shah, N. J. & Rosenberg, G. A. Quantitative measurement of blood-brain barrier permeability in human using dynamic contrast-enhanced MRI with fast T1 mapping. Magn. Reson. Med. 65, 1036–1042 (2011). 204. Dula, A. N., Gochberg, D. F., Valentine, H. L., Valentine, W. M. & Does, M. D. Multiexponential T2, magnetization transfer, and quantitative histology in white matter tracts of rat spinal cord. Magn. Reson. Med. 63, 902–909 (2010). 205. Beg, M. S., Mohapatra, J., Pradhan, L., Patkar, D. & Bahadur, D. Porous Fe3O4- SiO2 core-shell nanorods as high-performance MRI contrast agent and drug delivery vehicle. J. Magn. Magn. Mater. 428, 340–347 (2017). 206. Hiller, K.-H., Waller, C., Nabrendorf, M., Bauer, W. R. & Jakob, P. M. Assessment of cardiovascular apoptosis in the isolated rat heart by magnetic resonance molecular imaging. Mol. Imaging 5, 115–121 (2006). 207. Schilling, F. et al. MRI measurements of reporter-mediated increases in transmembrane water exchange enable detection of a gene reporter. Nat. Biotechnol. 35, 75–80 (2017). 208. Carril, M. Activatable probes for diagnosis and biomarker detection by MRI. J. Mater. Chem. B 5, 4332–4347 (2017). 209. Loving, G. S. & Caravan, P. Activation and retention: A magnetic resonance probe for the detection of acute thrombosis. Angew. Chemie - Int. Ed. 53, 1140–1143 (2014). 210. He, J. et al. Prototypes of Lanthanide(III) Agents Responsive to Enzymatic Activities in Three Complementary Imaging Modalities: Visible/Near-Infrared Luminescence, PARACEST-, and T1-MRI. J. Am. Chem. Soc. 138, 2913–2916 (2016). 211. Bhuiyan, M. P. I. et al. Concentration-independent MRI of pH with a dendrimer- based pH-responsive nanoprobe. Contrast Media Mol. Imaging 10, 481–486 (2015). 212. Ye, D. et al. Redox-triggered self-assembly of gadolinium-based MRI probes for sensing reducing environment. Bioconjug. Chem. 25, 1526–1536 (2014). 213. Barral, M. et al. Diffusion-weighted MR Imaging of the Pancreas : Current Status and Recommendations. Radiology 274, 45–63 (2015). 214. Anderson, S. W. et al. Effect of disease progression on liver apparent diffusion coefficient values and T2 Values in a Murine Model of Hepatic Fibrosis at 11.7 Tesla MRI. J. Magn. Reson. Imaging 33, 140–146 (2012). 215. Dallaudière, B. et al. Diffusion-weighted MR imaging in musculoskeletal diseases: Current concepts. Diagn. Interv. Imaging 96, 327–340 (2015).
111