MRI Analysis to Detect Gray Matter Tissue Loss in Multiple Sclerosis

Home , Grey matter

MRI ANALYSIS TO DETECT GRAY MATTER TISSUE LOSS

IN MULTIPLE SCLEROSIS

KUNIO NAKAMURA

Submitted in partial fulfillment of the requirements

For the degree of Doctor of Philosophy

Dissertation advisor: Elizabeth Fisher, PhD

Department of Biomedical Engineering

CASE WESTERN RESERVE UNIVERSITY

August, 2011 CASE WESTERN RESERVE UNIVERSITY

SCHOOL OF GRADUATE STUDIES

We hereby approve the thesis/dissertation of

Kunio Nakamura ______

Ph.D. candidate for the ______degree *.

Andrew M. Rollins (signed)______(chair of the committee)

Elizabeth Fisher ______

Bruce D. Trapp ______

David L. Wilson ______

______

June 10, 2011 (date) ______

*We also certify that written approval has been obtained for any proprietary material contained therein.

To my mother, my father, and my wife for their love, patience and support.

Table of Contents

Table of Contents ...... 1 List of Tables ...... 4 List of Figures...... 5 Acknowledgements ...... 8 List of Abbreviations ...... 9 Abstract...... 11 Chapter 1 Introduction...... 13 1.1 Overview of Multiple Sclerosis ...... 13 1.2 Clinical Features of MS ...... 14 1.2.1 Symptoms and Diagnosis...... 14 1.2.2 MS Disease Course ...... 15 1.2.3 Clinical Measures of Disability...... 17 1.2.4 MS Treatment...... 18 1.3 Pathology in MS ...... 19 1.3.1 Focal Pathology in White Matter ...... 19 1.3.2 Diffuse Pathology in White Matter ...... 21 1.3.3 Pathology in Gray Matter...... 22 1.4 Imaging Markers in MS...... 25 1.4.1 Conventional MRI in MS...... 26 1.4.2 Advanced MRI to Detect GM Abnormality...... 28 1.5 Brain Atrophy in MS ...... 35 1.6 Gray Matter Atrophy in MS...... 37 1.6.1 Volumetric Gray Matter Atrophy...... 38 1.6.2 Cortical Thickness Measurements in MS...... 42 1.7 Objectives ...... 44 Chapter 2 Gray Matter Segmentation Algorithm...... 46 2.1 Introduction...... 46 2.1.1 MRI Format and Viewing ...... 46 2.1.2 Background ...... 47 2.1.3 Overview of Segmentation Algorithm ...... 49 2.2 Segmentation Algorithm...... 51 2.2.1 Image Correction...... 51 2.2.2 Fuzzy C-means Clustering Method...... 54 2.2.3 Anatomic Probability Map...... 55 2.2.4 Morphologic Probability Map...... 57 2.2.5 Partial Volume Correction ...... 59 2.3 Validation...... 60 2.3.1 Accuracy Test with Simulated MRI...... 60 2.3.2 Accuracy Test with Manually Traced Real MRIs...... 61 2.3.3 Reproducibility Test...... 64 2.3.4 Comparison to Other Segmentation Methods ...... 65 2.3.5 Sensitivity to Lesion Size...... 67 2.4 Discussion...... 70

2.4.1 Segmentation Method ...... 70 2.4.2 Validation Tests ...... 71 2.4.3 Conclusion ...... 74 Chapter 3 Measurement of Gray Matter Atrophy in MS...... 75 3.1 Introduction...... 75 3.2 Materials and Methods...... 75 3.2.1 Subjects ...... 75 3.2.2 MRI ...... 77 3.2.3 Image Analysis...... 77 3.2.4 Statistical Analysis...... 78 3.3 Results...... 79 3.3.1 Baseline Characteristics ...... 79 3.3.2 Group Comparison of Gray Matter Atrophy...... 80 3.3.3 Clinical Correlation of Gray Matter Atrophy...... 81 3.3.4 Predictors of Gray Matter Atrophy ...... 84 3.4 Discussion...... 84 Chapter 4 Cortical Longitudinal Atrophy Detection Algorithm...... 88 4.1 Introduction...... 88 4.1.1 Background ...... 88 4.1.2 Overview of CLADA...... 91 4.1.3 Deformable Model Format and Viewing ...... 92 4.2 Algorithm Description...... 93 4.2.1 Image Acquisition ...... 93 4.2.2 Pre-processing...... 93 4.2.3 Registration and Averaging...... 95 4.2.4 Segmentation...... 97 4.2.5 Deformable Model ...... 98 4.2.6 Longitudinal Deformation...... 105 4.2.7 Measurement Cortical Thickness and Change...... 106 4.3 Validation...... 107 4.3.1 Phantom Experiments ...... 107 4.3.2 MRI Experiments...... 109 4.3.3 Scan-Rescan Reproducibility ...... 111 4.3.4 Intra-operator Reproducibility Test...... 114 4.3.5 Accuracy using Post-mortem Images...... 116 4.3.6 Sensitivity using Simulation of Atrophy...... 120 4.3.7 Comparison of Image Sequences ...... 122 4.3.8 Effect of Reference Image ...... 123 4.3.9 Comparison with FreeSurfer ...... 125 4.3.10 Bias in Estimating Atrophy...... 127 4.3.11 Comparison to a Cross-sectional Method...... 127 4.3.12 Effects of MS Lesions...... 129 4.4 Discussion...... 131 4.4.1 Validation Tests ...... 132 4.4.2 Conclusion ...... 138 Chapter 5 Cortical Atrophy in MS...... 140 5.1 Materials and Methods...... 140 5.2 Results...... 141 5.2.1 Baseline Characteristics ...... 141 5.2.2 Cortical Thinning ...... 143

5.2.3 Clinical Correlation of Cortical Thickness...... 144 5.2.4 Disability Progression vs. Thickness and Change...... 145 5.2.5 Predictors of Cortical Thickness and Thinning...... 146 5.2.6 Sample Size Estimate...... 147 5.3 Discussion...... 147 Chapter 6 Summary and Future Work ...... 153 6.1 Summary...... 153 6.2 Limitations ...... 155 6.3 Future Work...... 157 6.3.1 Higher Field MRI...... 157 6.3.2 Treatment Effect on Gray Matter Atrophy...... 157 6.3.3 Investigation on Pseudoatrophy Effect ...... 158 6.3.4 Regional Correlation of GM Atrophy and WM Damage...... 159 6.3.5 Regional Correlation of GM Atrophy and Local Damage...... 159 6.4 Conclusion ...... 160 Bibliography ...... 161

List of Tables

Table 1-1: MS subtypes...... 15 Table 1-2: Classification of cortical lesions ...... 23 Table 1-3: Gray matter atrophy in the literature...... 39 Table 2-1: Result of accuracy test using simulated BrainWeb MRI...... 61 Table 2-2: MRI sequence...... 62 Table 3-1: MRI sequence for longitudinal study...... 77 Table 3-2: Baseline characteristics...... 79 Table 3-3: Characteristics of subjects by EDSS progression ...... 83 Table 3-4: Characteristics of subjects by MSFC progression ...... 83 Table 3-5: Predictors of atrophy from multiple regression analysis ...... 84 Table 4-1: Comparison between longitudinal and cross-sectional methods ...... 128 Table 5-1: Baseline characteristics...... 142 Table 5-2: Clinical and MRI correlations...... 145 Table 5-3: Baseline patient characteristics based on MSFC progression...... 146 Table 5-4: Sample size calculation for cortical thinning...... 147

List of Figures

Figure 1-1: Common progression pattern in MS...... 16 Figure 1-2: Photographs of fixed normal and MS brain slices...... 20 Figure 1-3: An example of proteolipid protein stained MS cerebrum...... 22 Figure 1-4: MRI of MS patient and definitions of tissue regions...... 26 Figure 1-5: Spectrum of brain atrophy from different MS patients...... 35 Figure 1-6: Serial MRI of RRMS patient over 8.5 years...... 36 Figure 2-1: Overall flowchart of the segmentation method...... 50 Figure 2-2: Segmentation of brain, outer contour, and T2 lesions from FLAIR...... 51 Figure 2-3: Correction of intensity variation ...... 52 Figure 2-4: Interleaved motion artifact correction...... 53 Figure 2-5: Co-registration of FLAIR and T1-weighted images...... 53 Figure 2-6: Intensity-based probability map...... 55 Figure 2-7: Anatomic probability map from the Harvard Brain Atlas ...... 57 Figure 2-8: Morphologic probability maps for deep gray matter cortical gray matter...... 59 Figure 2-9: Result of gray matter, lesion and brain segmentation...... 60 Figure 2-10: Comparison of automated and manual segmentation ...... 63 Figure 2-11: Plots of scan-rescan variability in gray matter fraction ...... 65 Figure 2-12: Comparison of three different segmentation techniques ...... 67 Figure 2-13: Effect of lesion size on GM volume ...... 69 Figure 3-1: Plot of mean annualized rates of atrophy in white matter fraction, gray matter fraction, and brain parenchymal fraction...... 81 Figure 3-2: Plot of gray matter fraction and EDSS ...... 82 Figure 3-3: Plot of gray matter fraction and MSFC...... 82 Figure 4-1: Schematic 2D examples of different measures of thickness...... 91 Figure 4-2: Flowchart of cortical longitudinal atrophy detection algorithm ...... 92 Figure 4-3: Averaged T1-weighted image in a CLADA-reference space...... 94 Figure 4-4: Input, corrected, registered, and averaged images for CLADA ...... 96

Figure 4-5: Segmentation steps in CLADA...... 98 Figure 4-6: Surface deformation for subject-specific surface model ...... 100 Figure 4-7: Comparison of computational times of self-intersection test with or without bounding box ...... 101 Figure 4-8: Laplace map in shrink-wrap algorithm...... 103 Figure 4-9: Nonlinearly transformed and filtered regional labels ...... 106 Figure 4-10: Effects of self-intersection test and surface relaxation steps ...... 108 Figure 4-11: Laplace map in phantom experiment...... 109 Figure 4-12: Example of CLADA images...... 110 Figure 4-13: Comparison of surfaces with and without the relaxation step ...... 110 Figure 4-14: Example of longitudinal deformation in atrophied brain...... 111 Figure 4-15: Boxplots of percent error in global and regional cortical thickness measurements from scan-rescan tests ...... 113 Figure 4-16: Surface map of scan-rescan absolute cortical thickness difference...... 114 Figure 4-17: Intra-operator variability in CLADA...... 116 Figure 4-18: Figurative flowchart of the rapid autopsy program ...... 117 Figure 4-19: Selection of region of interest from in situ MRI and from corresponding photograph of fixed brain slice ...... 119 Figure 4-20: Plot of cortical thickness measurements from in situ MRI and photographs of post-mortem brain slices...... 120 Figure 4-21: Result of sensitivity test...... 121 Figure 4-22: Plot of annualized cortical thickness change measured from T1SE and MPRAGE images...... 123 Figure 4-23: Effect of different reference spaces on cortical thickness and cortical thinning ...... 125 Figure 4-24: Comparison of cortical thickness measured from CLADA and FreeSurfer ...... 126 Figure 4-25: Scatter plot of scan-rescan variability from full longitudinal and cross- sectional versions of CLADA...... 129 Figure 4-26: Effect of MS lesions on the annualized rates of cortical thinning ...... 131 Figure 5-1: Baseline cortical thickness difference maps ...... 142

Figure 5-2: Boxplots of cortical thickness at baseline...... 143 Figure 5-3: Annualized change in global cortical thickness and longitudinal evolution of cortical thickness ...... 144

Acknowledgements

I would like to thank my advisor, Dr. Elizabeth Fisher, for her guidance, support and mentorship, not only during my graduate studies but also during my undergraduate times. This work would not be possible without her patience and kindness. Her expertise in the scientific design of experiments, image analysis, manuscript writing, and presentation has made me a better researcher.

I thank my committee members, Drs. Andrew Rollins (academic advisor), Bruce

D. Trapp, and David L. Wilson who provided me with helpful suggestions to improve my research and dissertation.

I am also grateful for the members of the Fisher laboratory, past and present, for many intellectual discussions. I am thankful to the collaborators in the Mellen Center for

Multiple Sclerosis Treatment and Research, Department of Neuroscience, and

Department of Quantitative Health Sciences as well as the patients and volunteers for their time and cooperation for the studies.

The studies were supported by NIH National Institute of Neurological Disorders and Stroke (P01-NS38667) and National Multiple Sclerosis Society (RG 3099A, RG

3548A2).

List of Abbreviations

25FTW: 25-Foot Timed Walk 9HPT: Nine-Hole Peg Test BPF: Brain Parenchymal Fraction CIS: Clinically Isolated Syndrome CLADA: Cortical Longitudinal Atrophy Detection Algorithm cMRI: Conventional Magnetic Resonance Imaging CNS: Central Nervous System CSF: Cerebrospinal Fluid CTh: Cortical Thickness DMT: Disease Modifying Therapy EDSS: Kurtzke Expanded Disability Status Scale FCM: Fuzzy C-Means Clustering Method FLAIR: Fluid-Attenuated Inversion Recovery FOV: Field-Of-View GdLV: Gadolinium-Enhancing Lesion Volume GM: Gray Matter GMF: Gray Matter Fraction ICS: Inner Cortical Surface MPRAGE: Magnetization Prepared Rapid Gradient Echo MRI: Magnetic Resonance Imaging MS: Multiple Sclerosis MSFC: Multiple Sclerosis Functional Composite MTR: Magnetic Transfer Ratio NC: Normal Control NABT: Normal-Appearing Brain Tissue NAWM: Normal-Appearing White Matter OASIS: Open Access Series of Imaging Studies OCS: Outer Cortical Surface

OCV: Outer Contour Volume PASAT: Paced Auditory Serial Addition Test PD: Proton Density PPMS: Primary Progressive Multiple Sclerosis ROI: Region of Interest RRMS: Relapsing-Remitting Multiple Sclerosis SPM: Statistical Parametric Mapping SPMS: Secondary Progressive Multiple Sclerosis T1LV: T1 lesion volume T1SE: T1-weighted Spin Echo T2LV: T2 lesion volume WM: White Matter WMF: White Matter Fraction

MRI Analysis to Detect Gray Matter Tissue Loss in Multiple Sclerosis

Abstract

KUNIO NAKAMURA

Multiple sclerosis (MS) has been traditionally characterized by primary demyelination and inflammation in white matter (WM). However, recent histopathologic studies have shown that gray matter (GM) of MS patients is also abnormal. My aim was to develop methods for quantifying GM damage in terms of GM atrophy and cortical atrophy, to investigate the evolution in various MS disease stages, and to assess relevance to clinical status. First, I developed an automated algorithm that segmented GM and WM in magnetic resonance images (MRI) and measured the normalized GM volume. The algorithm was designed to be applicable to MRI of MS patients, which had focal lesions and significant atrophy. The algorithm was validated and applied in a longitudinal study that included patients with clinically isolated syndrome (CIS), relapsing-remitting

(RRMS) and secondary progressive (SPMS) MS as well as healthy normal controls.

Other conventional MRI markers of focal damage and clinical measures were available to explore the correlations and predictors of GM atrophy. We found that (1) the rate of GM atrophy increased in a stage-dependent manner, which was similar to that of whole brain atrophy, (2) GM atrophy had moderately strong correlations with the disability measures, and (3) predictors of GM atrophy changed from RRMS to SPMS. Next, I developed a

registration and deformable model-based longitudinal method (CLADA, Cortical

Longitudinal Atrophy Detection Algorithm) that had high reproducibility and could

measure global and regional cortical atrophy in terms of cortical thickness and its change.

CLADA was validated and applied to the full longitudinal MRI dataset to explore the

evolution of cortical thinning, its clinical correlations, and predictors. The rate of cortical

thinning increased with advancing disease, correlated with clinical disability, and

distinguished stable and worsening patients with more significance than GM atrophy. In

summary, my research showed that GM and cortical atrophy could be measured reliably

with the new techniques. Furthermore, application of these methods in MS patients

demonstrated that GM and cortical atrophy measurements were (1) relevant both clinically and biologically (2) able to provide insights on MS pathogenesis; and (3) suitable for future clinical trials of potential MS therapies.

CHAPTER 1 INTRODUCTION

In this chapter, an overview of multiple sclerosis (MS) including its clinical, pathologic, and imaging description is provided. Next, various advanced imaging techniques that are sensitive to abnormality in gray matter (GM) are described. The measurement of GM atrophy is presented as a clinically relevant mean to detect overall

GM damage, and finally the aims of my project are outlined.

1.1 Overview of Multiple Sclerosis

Multiple sclerosis is a chronic, inflammatory and demyelinating disease of the central nervous system (CNS). Its pathologic hallmarks include inflammation, demyelination, gliosis, and axonal degeneration. Conventional magnetic resonance imaging (MRI) typically shows multiple focal abnormalities in the brain and spinal cord.

MS affects approximately 2.5 million worldwide and 300,000 in the United States

(Anderson et al., 1992) incurring an approximate annual cost of $10 billion (Whetten-

Goldstein et al., 1998). It is the most common cause of non-traumatic neurologic disability among young adults (Weinshenker, 1996). The average age at disease onset is around 30, and women are 2-3 times more susceptible than men. MS symptoms, prognosis, and disease course vary significantly among patients. But, many patients have

a relatively normal life expectancy despite their disability, and only a small portion of

patients die prematurely (National Multiple Sclerosis Society, 2011).

1.2 Clinical Features of MS

1.2.1 Symptoms and Diagnosis

Multiple sclerosis is a highly variable disease in terms of symptoms and

progression, but it is generally characterized by unpredictable acute attacks of

neurological impairment (relapses) followed by chronic progression of disability. There

is no surrogate marker that reliably predicts the relapses or progression.

The relapses are highly variable within and among patients. These symptoms

include sensory disturbances (numbness, pain, tingling sensation, and coordination

problems), motor dysfunctions (spasticity, gait ataxia, and weakness in arm and leg), and visual impairments (Noseworthy et al., 2000). Other MS symptoms include seizure,

bladder, bowel, and sexual dysfunctions (Compston and Coles, 2008). Cognitive

dysfunction (slow processing speed and inability to concentrate) and memory impairment

are also common (Chiaravalloti and DeLuca, 2008). Depression and fatigue are also widely reported (Ziemssen, 2009). Relapses generally persist for days. Patients tend to

recover completely or partially at the early disease stage even without any treatment, and

a relatively long period of disease inactivity follows (remittance). Relapses become rare

during progressive worsening of disability.

Since its symptoms are similar to other conditions, MS can be difficult to

diagnose, and the diagnostic criteria for MS have been evolving. The traditional criteria

required at least 2 relapses for clinically definite MS (Poser et al., 1983). The current

diagnostic criteria (revised McDonald criteria) integrate MRI and require a dissemination of lesions in both time and space on MRI.1 Since relapses are rare (0.6-1.1 relapse per year (Scalfari et al., 2010)), the revised McDonald criteria facilitate early diagnosis of

MS, allowing early MS treatment.

1.2.2 MS Disease Course

Three MS subtypes are defined according to the pattern of relapse and progression of disability (Lublin and Reingold, 1996) as shown in Table 1-1.

Table 1-1: MS subtypes (Lublin and Reingold, 1996). MS Subtype Description Clearly defined relapses with full or partial recovery; Relapsing-remitting MS periods between disease relapses characterized by a lack (RRMS) of disease progression (remittance) Initial RR disease course followed by progression with or Secondary progressive without occasional relapses, minor remissions, and MS (SPMS) plateaus Primary progressive MS Disease progression from onset with occasional plateaus (PPMS) and temporary minor improvements allowed.

Relapsing-Remitting and Secondary Progressive MS

Approximately 85% of MS patients experience relapse onset type of MS where patients initially exhibit an RRMS pattern followed by SPMS (Confavreux and Vukusic,

2006). At the first relapse suggestive of CNS inflammation, patients are diagnosed as

clinically isolated syndrome (CIS). At the second relapse or imaging evidence of

1 The dissemination of lesions in time require either (1) a contrast-enhancing lesion at the different location, at least 3 months after the initial clinical event or (2) a new hyperintense lesion on T2-weighted MRI at least 30 days after the initial clinical event. The dissemination of lesions in space require 3 of the following 4 conditions: (1) at least 1 contrast-enhancing lesion or 9 hyperintense lesions on T2 scans if there is no contrast-enhancing lesion; (2) at least 1 infratentorial lesion; (3) at least 1 juxtacortical lesion; and (4) at least 3 periventricular lesions.

disseminated lesions by time and space, the CIS patients are diagnosed as RRMS

(Polman et al., 2005). Most CIS patients who become MS initially present with optic

neuritis or sensory symptoms (Weinshenker et al., 1989). RRMS patients generally

advance to SPMS where disability worsens with no or occasional relapses (Confavreux and Vukusic, 2008). A typical evolution of disability is illustrated in Figure 1-1 (Trapp et al., 1999a).

Figure 1-1: A common progression pattern in MS (dark dashed line). Patients do not experience symptoms until the first relapse (onset). With additional clinical or imaging evidence of CNS inflammation suggestive of MS, patients are diagnosed with relapsing- remitting MS (RRMS) and experience relapses with complete or partial recoveries. During secondary progressive MS (SPMS), patients become more disabled even though the relapses become rare. The gray line shows a hypothesized sub-clinical disease activity in RR-SP MS course. Modified from Trapp et al. (Trapp et al., 1999a).

It is believed that the clinical relapses are only the tip of the iceberg and sub-

clinical disease activity occurs even before the diagnosis and continues throughout the

disease. Patients may be capable of functionally compensating during RRMS but lose the

compensatory ability at a certain threshold, and convert to SPMS (Figure 1-1, gray curve)

(Trapp et al., 1999b). It is uncertain whether relapses predict long-term disability in MS

(Confavreux et al., 2000; Ebers et al., 2008; Scalfari et al., 2010). A potential surrogate marker of MS disease progression must be sensitive to such sub-clinical disease activity.

Primary progressive MS

The rest of MS patients (15%) are primary progressive MS (PPMS) who continuously become disabled without or very rare (in the order of 1 per decade) relapses from the disease onset (Confavreux and Vukusic, 2006; Kremenchutzky et al., 1999;

Lublin and Reingold, 1996). Compared to RR-SP MS, the onset age is older (mean=40) and the proportion of men is higher (50%) in PPMS (Miller and Leary, 2007). The pathological and biological difference underlying these clinical phenotypes is not clear.

1.2.3 Clinical Measures of Disability

There are standardized methods to evaluate the level of disability. The most common measure is Kurtzke expanded disability status scale (EDSS) (Kurtzke, 1983), which scores disability on an ordinal rating scale from normal (EDSS=0) to death due to

MS (EDSS=10). The pivotal scale is EDSS of 6 where patients require unilateral walking aids. At EDSS of 9.5, which is the last score before death, the patients are “totally helpless bed patient; unable to communicate effectively or eat/swallow.”

A more recent quantitative measure of disability is the multiple sclerosis functional composite (MSFC), which is a combined z-score of 3 tests: (1) 25-foot timed walk (25FTW) for lower extremity function, (2) nine-hole peg test (9HPT) for arm and hand function, and (3) paced auditory serial addition test (PASAT) for cognitive function

of concentration and processing speed. The benefits with MSFC are that it is objective, quantitative, and less expensive than EDSS because a trained staff can administer the test rather than a neurologist (Rudick et al., 1997). MSFC includes a test of cognition.

Both measures have disadvantages. EDSS is nonlinear, subjective (prone to inter- rater variability), and heavily dependent on ambulation after EDSS of 6. EDSS changes may be reversible over short time periods (Ebers et al., 2008). The drawbacks for MSFC are that PASAT is highly demanding and intolerable by some patients. In longitudinal studies, PASAT tends to improve because of a learning effect (Polman and Rudick, 2010).

Both measures are used in correlational studies of imaging markers. Progression on EDSS is often used in MS clinical trials as the primary outcome measure.

1.2.4 MS Treatment

There is no cure for MS. An ideal cure for MS should halt the disease progression and regain the neurologic deficits. Preventive measures are also not available. Current

MS treatments are either symptomatic treatments or disease-modifying therapy (DMT).

For severe acute relapses, high-dose corticosteroids are intravenously injected to reduce inflammation and shorten the relapse severity and duration. However, corticosteroids have not demonstrated superiority over other proven therapies in terms of reducing disease activity and slowing progression (Frohman et al., 2007; Shah et al.,

2007).

There are currently 7 DMTs approved in the United States for MS: interferon β-1a subcutaneously or intramuscularly, interferon β-1b, glatiramer acetate, mitoxantrone, natalizumab, and fingolimod. In RRMS, interferon β and glatiramer acetate are most

commonly used. These agents are only moderately effective, as some patients do not respond to treatment (Rudick et al., 2004). Natalizumab and mitoxantrone have potentially serious side effects and are typically used for patients who do not respond to other therapies.

Typically, DMT’s efficacy is shown using the relapse rate or disability progression. But the relapse rate is low (typically 0 or 1 relapse per year), and the changes in EDSS or MSFC are small during the short duration of clinical trials (2 years or less), thus requiring a large number of patients. DMT’s efficacy in preventing long- term disability from these short clinical trials is not well established (Ebers et al., 2008).

Therefore, clinical trials for new therapeutic candidates demand sensitive, cost-effective, non-invasive, and objective surrogate markers of disease progression.

1.3 Pathology in MS

MS brains typically contain characteristic multiple foci in WM and these focal areas have been extensively studied due to their appearance on MRI and on post-mortem specimen. However, there is significant damage in non-focal areas including normal- appearing white matter (NAWM) and GM.

1.3.1 Focal Pathology in White Matter

The most obvious pathology in MS is the multiple visible WM abnormalities called lesions or plaques (Figure 1-2, arrows). The lesions are typically located periventricularly or subcortically throughout the brain.

Figure 1-2: Normal brain (a) and MS brain with WM lesions (arrows, b). Courtesy of Dr. Ansi Chang.

The lesions are areas of demyelination and are extremely heterogeneous as there are various degrees of inflammation, demyelination, remyelination, gliosis, and axonal damage. These lesions are classified according to their inflammatory activity: (1) active lesions are hypercellular demyelinated area with abundant perivascular and parenchymal macrophages; (2) chronic active lesions are demyelinated and have hypercellular border, a hypocellular center, and perivascular cuffs. The hypercellular border often contains macrophages and myelin debris; and (3) chronic inactive lesions are hypocellular and demyelinated throughout the area. These lesions often have gliosis and large extracellular spaces (De Groot et al., 2001; Trapp et al., 1998). Remyelinating lesions are characterized by thinly wrapped and irregularly formed myelin. These lesions are also called shallow plaques (Bruck et al., 2003).

Axonal damage is important pathology in MS and occurs early (Trapp et al.,

1998). Possible mechanisms of axonal damage include bystander effects of inflammatory activity (Bo et al., 1994), disturbed energy balance and chronic necrosis (Trapp and Stys,

2009). Unlike myelin content, which can increase with remyelination (Prineas and

Connell, 1979), axonal transection is irreversible and believed to be responsible for the progression of irreversible disability (Bjartmar et al., 2000; Trapp et al., 1999b).

1.3.2 Diffuse Pathology in White Matter

Normal-appearing white matter (i.e., outside of focal lesions, NAWM) is myelinated but not normal as activated microglia, gliosis, and axonal loss can be found

(Allen et al., 2001; Trapp et al., 1998). There is significant diffuse inflammation in

NAWM of SPMS and PPMS compared to acute or RRMS patients (Kutzelnigg et al.,

2005). There is also axonal loss in NAWM of MS compared to controls (Evangelou et al.,

2000; Ganter et al., 1999; Peterson et al., 2001), and an increased number of axonal terminal spheroids is found in progressive MS compared to acute or relapsing MS

(Kutzelnigg et al., 2005). These studies have shown that NAWM in MS is abnormal and may reflect the progression of disease. Importantly, conventional MRI is not sensitive to pathology in NAWM. The precise mechanism of axonal damage in NAWM is also not known. A possible mechanism is that once axons are damaged by nonspecific inflammatory process, the distal portion of axons undergoes Wallerian degeneration, resulting in axonal degeneration in NAWM. This phenomenon is evidenced by the empty tubes of remaining myelin sheaths (Dutta and Trapp, 2007; Trapp et al., 1998). Axonal

loss in NAWM is also irreversible and considered to contribute to the progression of

irreversible disability (Bjartmar et al., 2003).

1.3.3 Pathology in Gray Matter

In addition to pathology in WM, histopathological studies show that GM of MS

patients is also demyelinated, and GM demyelination can be equally or more extensive

than focal WM demyelination. Vercellino et al. detected cortical demyelination in 15% of

the total GM area and WM demyelination in 22% of the total WM area in coronal brain sections (Vercellino et al., 2005). Similarly, Bo et al. found 26.5% demyelinated GM area

versus 6.5% in WM (Bo et al., 2003); Gilmore et al found 33% GM and 20% WM

demyelinated areas (Gilmore et al., 2009).

Figure 1-3: A proteolipid protein (PLP) stained example of MS cerebrum. The figure shows densely myelinated white matter (dark), less myelinated cortex (less dark), white matter demyelination (crosses), and gray matter demyelination (arrows). Courtesy of Dr. Ansi Chang.

Cortical lesions are classified according to their locations (Bo et al., 2003; Kidd et

al., 1999). The most common classification is described in Table 1-2 (Bo et al., 2003).

Table 1-2: Widely recognized classification of cortical lesions defined by Bo et al. (2003). Types III and IV may be combined for subpial lesions. Classification Distribution of cortical lesion

Type I GM/WM mixed lesion (leukocortical lesions)

Type II Intracortical lesion; completely in the cortex

Subpial cortical lesion; extending from pial surface toward WM but Type III not touching it

The percent area of cortical demyelination with respect to the total cortical area

increases with the disease stage: from 3% in RRMS to 26% in SPMS in one study

(Vercellino et al., 2005), and from 3% in RRMS to 13% in SPMS in another study

(Kutzelnigg et al., 2005). Cortical lesions are more frequently found in cingulate gyrus

(Bo et al., 2003; Vercellino et al., 2005). In addition to the cerebral cortex (Bo et al.,

2003; Kidd et al., 1999; Kutzelnigg et al., 2005; Peterson et al., 2001), GM demyelination

can be found in other CNS regions including the spinal cord (Gilmore et al., 2006;

Gilmore et al., 2009; Gilmore et al., 2009), cerebellum (Gilmore et al., 2009; Kutzelnigg et al., 2007), hippocampus (Geurts et al., 2007; Papadopoulos et al., 2009; Vercellino et al., 2005), basal ganglia (Vercellino et al., 2005), thalamus (Cifelli et al., 2002), and hypothalamus (Huitinga et al., 2001).

Similar to the focal WM lesions, GM pathology in demyelinated areas includes neuronal, axonal and glial loss. There is also synaptic and dendritic loss in GM lesions

(Cifelli et al., 2002; Frischer et al., 2009; Peterson et al., 2001; Vercellino et al., 2005;

Wegner et al., 2006). However, GM lesions are less inflammatory than WM lesions as

there is no significant lymphocyte infiltration, complement deposition or BBB

breakdown (Brink et al., 2005; Kidd et al., 1999; Peterson et al., 2001; van Horssen et al.,

2007).

The pathology in myelinated cortex, i.e., regions outside of cortical lesions, is not

well established. Peterson et al. have reported slightly increased transected neurites in the

myelinated cortex compared to the control cortex (8±4 par mm3 and 1±1 par mm3, respectively), but this difference is subtle compared to the number of transected neurites in the demyelinated cortical area (4119±540 in active cortical lesions) (Peterson et al.,

2001). Wegner et al. also have shown that the neuronal density in the myelinated MS cortex is not statistically different from that in the cortex of control brains (86±13 vs.

89±22 cells/mm2, p>0.20) (Wegner et al., 2006). There is general thinning in the cerebral

cortex of MS compared to controls (2.55±0.2mm and 2.84±0.36mm, p=0.027) as

measured from 2D post-mortem brain slices, but no significant correlation was found

between cortical thickness and the extent of cortical demyelination (Wegner et al., 2006).

Whether GM neurodegeneration occurs as a retrograde degeneration from

damaged axons within lesions or independently of focal WM pathology remains highly

debatable (Frischer et al., 2009; Geurts et al., 2009; Geurts and Barkhof, 2008; Pirko et

al., 2007; Stadelmann et al., 2008). For example, the association between meningeal

inflammation and cortical demyelination in the cortex is unclear (Kooi et al., 2009;

Magliozzi et al., 2007). Regardless of the dependence on WM pathology, various types of

damage occur throughout GM in MS. Further study of GM pathology in living MS

patients is required understand its relationship to disability progression.

1.4 Imaging Markers in MS

Histopathologic studies are performed on biopsy and autopsy samples. Biopsy samples may be atypical MS because the purpose of biopsies is to rule out other rare conditions such as sarcoidosis at the diagnosis (National Multiple Sclerosis Society,

2011) and because biopsies are not required for monitoring. Autopsy cases are biased towards the chronic end-stage MS. In vivo imaging techniques such as MRI can non- invasively detect ongoing MS abnormalities. MRI has greater contrast for soft brain tissues and higher sensitivity for MS lesions than computed tomography (Nesbit et al.,

1991). MRI provides visual and quantitative measures that can be used to investigate pathologic events, to objectively monitor the disease progression, and to evaluate therapeutic outcomes.

In this section, conventional MRI (cMRI), new MRI techniques, and quantitative image analysis methods are reviewed. An example of MRI in MS and definitions of tissues are shown in Figure 1-4 and described in the following sections.

Figure 1-4: Examples of MRI in MS and definitions of tissues. (a) T2-weighted image, (b) fluid-attenuated inversion recovery image, (c) proton density-weighted image, (d) T1- weighted image without contrast, (e) T1-weighted image after contrast injection, (f) brain parenchyma, (g) T2 lesions, (h) gray matter, (i) white matter, (j) normal-appearing white matter, (k) magnetization transfer ratio map, (l) apparent diffusion coefficient map, which was calculated from diffusion tensor images. The images are registered and transformed with b- spline interpolation. (f-j) The tissue definitions are shown in colors.

1.4.1 Conventional MRI in MS

T2-weighted imaging

T2-weighted MR images of MS patients show hyperintense regions in WM (T2 lesions), typically in periventricular and subcortical WM, which correspond to focal MS lesions found at autopsy (Stewart et al., 1984). A problem of distinguishing cerebrospinal

fluid (CSF) and WM lesions on T2-weighted spin echo (T2SE) images (Figure 1-4a) can

be resolved by using T2-weighted fluid-attenuated inversion recovery (FLAIR) images

because of dark CSF and increased contrast of T2 WM lesions (Figure 1-4b). MRI has a

superior sensitivity over clinical relapses as MRI can detect clinically silent disease

activity. In fact, MRI-detected disease activity can outnumber clinical activity by 10-fold

(Confavreux and Vukusic, 2006). MRI-guided histopathologic comparison reveals that

only 55% of T2 lesions are demyelinated, indicating that T2 lesions are nonspecific

marker of MS pathology (Fisher et al., 2007). Furthermore, T2 lesion volume, which

increases from RRMS to SPMS, eventually levels off (“plateauing”) and does not explain

the accumulation of neurologic disability (Li et al., 2006). Indeed, T2 hyperintense

lesions are only weakly correlated with clinical disability in most studies (Barkhof, 1999;

Filippi et al., 1995; Li et al., 2006; Mammi et al., 1996).

Proton density-weighted imaging

Proton density (PD)-weighted images (Figure 1-4c) are sensitive to the density of

water protons. PD-weighted images show greater GM-WM contrast, but hyperintense MS lesions may appear similar to GM. PD-weighted MRI can be acquired without an additional scan time when the PD/T2-weighted dual echo sequence is used to acquire T2- weighted MRI. Using both PD/T2-weighted dual echo MRIs is helpful when segmenting

T2 lesions.

T1-weighted imaging

T1-weighted spin echo images (Figure 1-4d) show that the subsets of T2 lesions

are hypointense lesions (T1 lesions). Chronic T1 lesions are called black holes. They significantly correlate with the percent of residual axons stained with Bodian staining and moderately correlate with degree of matrix destruction stained with hematoxylin and eosin, indicating that T1 lesions represent regions of severe tissue destruction (van

Walderveen et al., 1998). Contrast enhancement of T1-weighted images using

gadopentetate dimeglumine (Gd-DTPA) shows hyperintense lesions (i.e., Gd-lesion or

enhancing lesions; Figure 1-4e) at regions where there is blood-brain barrier breakdown.

Thus these enhancing lesions represent active inflammation (Kermode et al., 1990). A

weekly MRI study has shown that each enhancing lesion lasts about 2-4 weeks (Cotton et

al., 2003) with or without clinical decline, emphasizing MRI’s greater sensitivity over

clinical symptoms (Lai et al., 1996). These enhancing lesions are also seen at the first

appearance of T2 lesions (Kermode et al., 1990). Like T2 lesion volume, enhancing

lesions are commonly used in clinical trials due to the high sensitivities (Frank et al.,

1994; Miller et al., 1988). However, the correlation between lesion measures and

disability is only modest at best, and the number of enhancing lesions decreases in the progressive phase of the disease (Filippi and Rocca, 2007).

1.4.2 Advanced MRI to Detect GM Abnormality

cMRI is highly sensitive to the focal MS pathology in WM, but insensitive to GM

pathology. Only 0 to 10% of cortical lesions were detected using cMRI (Geurts et al.,

2005b; Kidd et al., 1999; Newcombe et al., 1991). Therefore, non-conventional,

advanced MRI techniques have been implemented to improve the detection of GM

damage. GM damage is important because it can be more extensive than WM (Bo et al.,

2003; Gilmore et al., 2009; Vercellino et al., 2005) and because conventional imaging markers (i.e., T2 lesion and enhancing lesion volumes) do not fully explain the progression of disability (Filippi and Rocca, 2007; Li et al., 2006). In the following sections, examples of advanced MRI methods are discussed, including magnetic resonance spectroscopy, magnetization transfer ratio, and diffusion tensor imaging as well as new MRI methods to improve the sensitivity of abnormal GM. After these examples, brain atrophy is separately described as a quantitative method to detect overall brain damage.

Magnetic resonance spectroscopy

Magnetic resonance spectroscopy (MRS) is an in vivo spectroscopic method to quantify regional metabolite compositions using MRI. Among various MRS-detectable metabolites (e.g., choline, creatine, inositol, alanin, and lactate), N-acetylaspartate (NAA) is particularly important in MS because NAA is localized in neurons, neuronal processes and axons; thus, NAA is a specific marker of neuronal damage and neuronal function

(Bjartmar et al., 2001).

Acute lesions show a reduction in NAA level which may partially recover after acute phase, suggesting that NAA is decreased in lesions but reversible in the surviving axons (De Stefano et al., 1995b). Alterations in the lipid resonances are shown to precede the appearance of lesions in NAWM (Narayana et al., 1998). Another MRS study has shown that axonal pathology begins early as the NAA level is significantly decreased in

NAWM of CIS patients compared to healthy normal controls (Wattjes et al., 2007).

Several studies have correlated the decreased NAA levels to disability levels (Davie et al.,

1995; De Stefano et al., 2003). Furthermore, MRS abnormalities are detected in deep GM

including thalamus (Cifelli et al., 2002) and hippocampus (Geurts et al., 2006). In the cortex where MRS measurements are difficult (Geurts et al., 2004), the findings are

inconsistent (Chard et al., 2002b; Geurts et al., 2006; Kapeller et al., 2001; Sharma et al.,

2001).

Several disadvantages of MRS are (1) MRS-based NAA measures are often

normalized with respect to creatine concentration, which may vary (De Stefano et al.,

1995b), (2) NAA decrease may be reversible in the surviving axons (De Stefano et al.,

1995a), (3) large MRS voxels are heavily contaminated by the partial volume effects,

which is critical in the cortex, (4) reproducibility is low and varies regionally (Geurts et

al., 2004), (5) it is difficult to apply in multi-center clinical trials, and (6) sensitivity to

longitudinal change and treatment response are currently uncertain (Barkhof et al., 2009).

Nevertheless, MRS is highly specific and provides insights into pathology relevant in MS.

Additional metabolites including GABA (gamma-aminobutyric acid), vitamin C and

glutamine are also investigated for their potential relevance in MS (Bakshi et al., 2008;

Srinivasan et al., 2005).

Magnetic transfer ratio

Magnetization transfer (MT) results from interactions of magnetized protons in

water and macromolecules (e.g., lipid membrane on myelin). Application of an off-

resonance MT pulse can selectively saturate magnetization on macromolecular protons

(Arnold, 2007). A magnetization transfer ratio (MTR) map can be generated from MT images acquired with and without off-resonant MT pulse as follows:

M − M MTR = 0 t Equation 1-1 M0

where M0 is the image without and Mt is the image with the MT saturation pulse (Wolff and Balaban, 1994). An example of MTR map is shown on Figure 1-4(k).

In post-mortem MRI-histopathologic correlational studies, MTR is strongly correlated to myelin content in WM (r = −0.84, p < 0.001) (Schmierer et al., 2004) and even sensitive to remyelination (Barkhof et al., 2003; Chen et al., 2007; Schmierer et al.,

2004). In a longitudinal study, the local inhomogeneity in MTR has predicted subsequent

MTR change in the lesions and NAWM because active lesions tend to appear inhomogeneous (inactive center with active edge) while chronic inactive lesions have homogeneous appearance (Chen et al., 2007). MTR alterations occur months before the appearance of contrast enhancing lesions. Therefore, MTR change may be the first MRI- detectable event as MS lesions emerge (Filippi et al., 1998). Other cross-sectional (i.e., single time-point across subjects) studies have shown significant differences between MS patients and healthy normal controls (NC) in whole brain MTR (Loevner et al., 1995) as well as in NAWM and GM MTR, indicating that MTR can detect diffuse pathology

(Cercignani et al., 2001; De Stefano et al., 2006; Ge et al., 2001b).

Image acquisition for MTR is not difficult, but MTR values strongly depend on the MRI sequence parameters and saturation pulses (Ropele and Fazekas, 2009). Current studies are investigating MTR’s reliability in multi-center clinical trials (van den

Elskamp et al., 2010) and its sensitivity in detecting subpial cortical lesions (Chen et al.,

2010).

Diffusion tensor imaging

Diffusion tensor imaging (DTI) is sensitive to random water motion or water diffusion (Basser et al., 1994). The water diffusion is represented by ellipsoidal tensor

consisting of three orthogonal eigenvectors (λ1, λ2, λ3 where λ1 > λ2 > λ3). Various DTI- based parameters can be calculated from the eigenvectors: for example, the trace of apparent diffusion coefficients (sum of 3 eigenvectors = λ1 + λ2 + λ3), axial diffusivity

(λ1), mean diffusivity (mean of 3 eigenvectors = (λ1 + λ2 + λ3) / 3), and fractional anisotropy, which is related to the standard deviation of the eigenvectors (anisotropy).

In the CNS, water diffusion is predominantly restricted by the semi-permeable cell membranes of axons and myelin (Pierpaoli et al., 1996), and DTI can indirectly show myelin and axonal abnormalities in NAWM (Ceccarelli et al., 2007). An example of apparent diffusion coefficient map is shown in Figure 1-4(l).

In MS, DTI studies show decreased fractional anisotropy and increased axial and mean diffusivity in lesions and NAWM compared to normal controls (Bammer et al.,

2000; Ciccarelli et al., 2001; Filippi et al., 2001). Similar abnormalities have been detected in GM (Ceccarelli et al., 2007; Cercignani et al., 2001; Ciccarelli et al., 2001), but DTI’s pathologic relevance is unclear as there are no directional fibers in GM.

Imaging of gray matter damage

Gray matter of MS patients generally appears normal on cMRI, and cMRI’s sensitivity to cortical lesions with respect to histopathology is extremely low (0-10%)

(Geurts et al., 2005b; Kidd et al., 1999; Newcombe et al., 1991). The sensitivity is higher for leukocortical lesions (type I, ~30%) due to their involvement in WM, but for intracortical (type II) and subpial lesions (type III), the sensitivity is extremely low (~5%)

(Geurts et al., 2005b) despite their abundance as 65-85% of all cortical lesions are the

subpial type (Albert et al., 2007; Bo et al., 2003; Peterson et al., 2001). Several factors

may explain low sensitivity of cMRI: (1) partial volume effects significantly reduce the

contrast due to the complex folding of the thin cortical sheet; (2) there is relatively little

myelin in the cortex, and thus the effect of demyelination is small; (3) GM has a higher

cellular density, and thus the extracellular space may not be as expandable as in WM and

may not significantly affect the density of MRI-detectable protons (Kidd et al., 1999);

and (4) cortical lesions are less inflammatory and cause less edema (smaller change in

proton density), which also reduces the contrast.

Advanced imaging methods such as double inversion recovery (DIR) (Geurts et

al., 2005a), and phase-sensitive inversion recovery (PSIR) (Nelson et al., 2007) have

improved the detection of cortical lesions by 150-500% with respect to cMRI, but the majority of cortical lesions are still missed when compared to the histopathologic standard (Seewann et al., 2010). There is no established standard DIR sequence yet, and since DIR is noisy, there is significant inter-rater variability in lesion detection (Geurts et al., 2011). Nevertheless, DIR-based cortical lesions have been detected in all types of MS

(Calabrese et al., 2009; Calabrese et al., 2010b; Calabrese et al., 2007a). Preliminary results have shown that the number and volume of DIR cortical lesions increase over 3

years in RRMS and SPMS patients (Calabrese et al., 2010a) and in PPMS (Calabrese et al., 2009), suggesting that cortical lesions may be associated with accumulation of disability.

There is no sensitivity improvement with high-field MRI (4.7T) (Geurts et al.,

2008), but ultrahigh-field MRI at 8-9.4T may reveal more cortical lesions (Kangarlu et al.,

2007; Schmierer et al., 2010). High-field or multi-sequence MRI has improved classification of cortical lesions (Bagnato et al., 2006; Nelson et al., 2007; Nelson et al.,

2008; Tallantyre et al., 2010). However, whole brain coverage is difficult, and high-field

MRI scanners are not used for routine clinical imaging studies.

In a post-mortem high-field (4.7T) MRI study of the spinal cord, high-resolution

PD-weighted MRI detected 73% of GM lesions. The sensitivity was higher than GM lesions in the cerebral cortex because most of the spinal cord GM lesions were leukocortical and tended to be large along the spine (Gilmore et al., 2009). In vivo spinal cord studies are difficult because (1) spinal cord is small and its MRI is affected by partial volume effects with surrounding bone, fat, and CSF, and (2) spinal cord MRI is also influenced by motion artifact due to respiratory and cardiac movements.

A reliable method to detect GM damage and the translation of advanced technology into clinical practice remain a challenge in MS. Therefore, various quantitative image processing methods that do not require acquisition of sophisticated

MRI or high field MRI scanner have been proposed.

1.5 Brain Atrophy in MS

Another imaging marker designed to capture overall disease severity is whole brain atrophy (Rudick et al., 1999). Whole brain atrophy represents the end result of destructive pathology, both focal and diffuse, and can be quantified from cMRI (Fisher and Rudick, 2003). The precise mechanisms of brain atrophy in MS are not known, but demyelination, axonal degeneration, and neurodegeneration as well as synaptic and dendritic loss are believed to lead to brain atrophy.

Measures of whole brain atrophy, such as brain parenchymal fraction (BPF), have been developed to estimate brain volume normalized to head size (Rudick et al., 1999).

BPF is decreased in MS (Rudick et al., 1999), and the rate of whole brain atrophy in MS patients is higher than in healthy normal controls (Fisher and Rudick, 2003); the brain volume decreases by 0.2-0.5%/year in normal aging (Fox and Schott, 2004) and by 0.6-

1.5%/year in MS (Barkhof et al., 2009; Fisher and Rudick, 2003). Examples BPF measurements in cross-sectional and longitudinal studies are shown in Figure 1-5 and

Figure 1-6, respectively.

Figure 1-5: Examples of spectrum of brain atrophy from different MS patients. Brain parenchymal fraction (BPF) from left to right is 0.87 (a, no atrophy), 0.82 (b, moderate

atrophy), and 0.70 (c, severe atrophy). Images are transformed to a standard space using rigid- body registration.

Figure 1-6: Registered serial axial FLAIR images from a single RRMS patient over 8.5 years. Brain parenchymal fraction (BPF) is shown with follow-up interval in months (in parenthesis) below each image.

Another commonly used measure of whole brain atrophy is the percent brain

volume change (PBVC), calculated by SIENA (Structural Image Evaluation, using

Normalisation, of Atrophy) (Smith et al., 2002) in FMRIB Software Library (FSL).

SIENA is a longitudinal, registration-based method that calculates shifts in the brain edges. In previous studies, SIENA has been more robust than cross-sectional methods

(Altmann et al., 2009) such as SIENAx (Smith et al., 2002) and central cerebral volume

(Losseff et al., 1996).

There are several confounding factors in the brain atrophy measurements.

Technical factors include inconsistency in the image acquisition (e.g., MRI sequence parameters, scanner hardware, software, image resolution, and subject repositioning).

Biological factors such as the level of (de)hydration can change the brain volume by 0.2-

0.6% (Duning et al., 2005; Kempton et al., 2009). An important factor relevant in MS is pseudoatrophy effect. This effect is an accelerated brain volume decrease presumably due to resolution of inflammatory edema. Indeed, a significant pseudoatrophy effect has been observed during clinical trials of anti-inflammatory agents (e.g., interferon β-1a and

natalizumab) where whole brain atrophy rates are greater in the treated group compared

to the placebo group in the first year. The atrophy rates are reversed and show the

treatment effect (i.e., reduced atrophy) in the second year (Miller et al., 2007; Rudick et

al., 2000). In another study, the changes in WM volume appear to be influenced by

inflammatory activity in WM (Tiberio et al., 2005).

Compared to lesion measurements, whole brain atrophy is more strongly

correlated to clinical measures such as EDSS and MSFC (Ge et al., 2000; Rudick et al.,

2001). Brain atrophy measures such as BPF and PBVC are routinely used as secondary or

exploratory outcome measures in large multi-center MS clinical trials (Cohen et al., 2008;

Miller et al., 2007; Rudick, 2004).

1.6 Gray Matter Atrophy in MS

Unfortunately, whole brain atrophy measurements may not accurately represent

“true” irreversible tissue loss because of the pseudoatrophy effect. Since GM lesions are

less inflammatory (Peterson et al., 2001), the effect of pseudoatrophy is believed to be

small in GM (Zivadinov et al., 2007). Therefore, GM atrophy may better reflect true

irreversible tissue loss. Furthermore, the neurodegenerative aspect of MS may be more

specifically measured by GM atrophy than whole brain atrophy because most of neurons

and neuronal cell bodies are in GM (Azevedo et al., 2009; Kandel et al., 2000), and little

myelin is found in GM. Therefore, GM atrophy is a potential marker of irreversible tissue

loss that is related to MS disease progression and may be useful in clinical trials of anti-

inflammatory and neuroprotective therapies.

1.6.1 Volumetric Gray Matter Atrophy

Typical volumetric measurement of GM atrophy involves (1) segmentation of

brain GM using high-resolution structural T1-weighted MRI, (2) measurement of

normalized GM volume, or GM fraction (GMF), and (3) calculation of annualized change

in GMF between two or more time-points.

Previous cross-sectional studies have shown that normalized GM volume is

decreased in MS compared to healthy normal controls (Chard et al., 2002a; Inglese et al.,

2004; Sastre-Garriga et al., 2004). In longitudinal studies, the rate of GM atrophy is estimated to be approximately 1-2% per year in MS (Horakova et al., 2008; Sastre-

Garriga et al., 2005; Tedeschi et al., 2009; Tiberio et al., 2005; Valsasina et al., 2005), which is greater than that due to normal aging (less than 1%) (Evans et al., 2010; Ge et al.,

2002; Thompson et al., 2003).

GM atrophy occurs early in MS as evidenced by decreased GMF found in early

RRMS compared to NC (Chard et al., 2002a). A longitudinal study reported a higher rate

of atrophy in CIS patients who converted to MS (Dalton et al., 2004), suggesting that GM atrophy is sensitive to sub-clinical damage at the very earliest stage of MS.

The correlations between GM atrophy and disability can vary (Table 1-3). Some small cross-sectional studies and short longitudinal studies have not found correlations between GM atrophy and disability. It is possible that these studies are too small (Tiberio et al., 2005) and that the level of disability is too homogeneous (Chard et al., 2002a).

Studies may require a longer duration because the short term change in disability is small and may not be very reliable (Ebers et al., 2008). In fact, a moderately strong correlation of correlation coefficient (r) = –0.58 is found between GMF and EDSS in a large cross- sectional study (n = 597) (Tedeschi et al., 2005). In a longer longitudinal study, Dalton et al. have compared the various rates of atrophy (whole brain, GM, WM, and ventricle) from a group of CIS patients who have converted to MS within 3 years of first relapse and a group that has remained as CIS, and they have found that GM atrophy has the largest difference (3.3% vs. 1.1%, CIS converting to MS and remaining CIS), indicating that GM is clinically relevant (Dalton et al., 2004). A long-term study of CIS followed over 20 years has shown that GM atrophy reflects the disease subtypes and disability to a greater extent than T2LV or WM atrophy (Fisniku et al., 2008).

Table 1-3: Table of gray matter atrophy (gray matter fraction, normalized gray matter volume, or cortical gray matter volume) in the literature.

Author Method Subject (#) GM Atrophy Rates GM-Clinical Correlations (year) RR (20), SP Liu (1999)§ Cavalieri NA None (20), NC (10) Ge 3D RR (30), NC NA No correlation with EDSS (2001a)† VIEWNIX (25) Chard RR (26), NC SPM99 NA No correlation with EDSS (2002a) (27) Quarantelli RR (50), NC RMC NA No correlation with EDSS (2003)† (54)

De Stefano RR (65), PP GMV-EDSS (RR)§: −0.27* SIENAx NA (2003) (25), NC (18) GMV-EDSS (PP)§: −0.64*** Chard RR (13), NC MS: −0.0086/y None except ΔGMF SPM99 (2004) (9) NC: −0.0021/y ARR 18mo prior to study ** Dalton SPM99 CIS (27), GM atrophy in converted CIS −0.017/3y (2004) (T2WI) CISÆMS (31) than remaining CIS GMV-Cognitive test score: Amato RR (41), NC −0.58** SIENAx NA (2004) (16) GMV-PASAT: 0.65** cGMV: ci < cp Sastre- PP (43), NC GMF-MSFC: 0.34* Garriga SPM99 NA (45) GMF-9HPT: 0.33* (2004) Sanfilipo MS (41), NC GMF-EDSS: −0.34* SPM99 NA (2005) (18) GMF-25FTW: −0.45** Oreja- Guevara SIENAx RR (26) NA No ΔEDSS correlation (2005) Sastre- SPM99 / Garriga PP (145) −1.50±1.6%/y NA SIENA (2005) Tedeschi RR (427), SP Predictor of GMF atrophy: (2005)† / RMC (140), PP( 30), NA EDSS + age at disease onset RMC NC (104) Tiberio RR (21), NC NC: −1.0%/y No correlations: GMF-EDSS SPM99 (2005) (10) MS: −2.1%/y nor MSFC RR (77), Benedict GMV-PASAT: 0.47** SIENAx SP(42), NA (2006) cGMV-neuropsych: 0.29-0.58 NC(27) No correlation with EDSS Morgen VBM + RR (19), NC NA GMV-PASAT: 0.54 (2006) SPM2 (19) GMV: ci < cp Prinster VBM + RR (51), NC No correlation: regional GMV NA (2006) RMC (34) and EDSS Portaccio RR (41), NC SIENAx NA NA (2006) (16) Oreja- No correlation: GMV,WMV- Guevara SPM99 RR (22) NA ∆EDSS (2006) Horakova SIENA RR (36) –1.4%/y NA (2008) NC: –0.02%/y NC (17), CIS GMF-EDSS: −0.48*** CIS: −0.028%/y (7), CIS→RR GMF-MSFC: 0.52*** Fisher CISÆRR:−0.096%/y FCM (8), RR (28), GMF-25FTW: −0.43*** (2008) RR: −0.23%/y GMF-9HPT: 0.46*** RR→SP (7), RRÆSP: −0.35%/y SP (19) GMF-PASAT: 0.43*** SP: −0.39%y GMV-EDSS: -0.41** NC (25), CIS GMV-MSFC: 0.55*** Fisniku SIENAx (29), RR (33), NA GMV-9HPT: 0.44** (2008)† SP (11) GMV-25FTW: -0.49** GMV-PASAT: 0.32* Tedeschi RMC RR/SP (267) −1.97%/2y NA (2009)

ΔGMV-ΔMSFC: 0.42** Furby ΔGMV-Δ9HPT: 0.31* SPM5 SP(117) −1.18%/y (2010) ΔGMV-ΔPASAT: 0.33** No correlation for ΔWMF Khalil CIS (44), RR cGMV-neuropsych: 0.5*** SIENAx NA (2011) (80) GMV- neuropsych: 0.5*** Abbreviations: 25FTW: 25-foot Timed Walk; 9HPT: Nine-Hole Peg Test; ARR: Annualized Relapse Rate; cGMV: Normalized Cortical Gray Matter Volume; ci: cognitively impaired; CIS: Clinically Isolated Syndrome; CISÆRR: patients who converted from Clinically Isolated Syndrome to Relapsing Remitting MS; cp: cognitively preserved; EDSS: Expanded Disability Status Scale; GMV: Gray Matter Volume; MSFC: MS Functional Composite; NC: Healthy Normal Control; NA: Not Available; PASAT: Paced Auditory Serial Addition Test; PP: Primary Progressive MS; RMC: Relaxometric Characterization; RR: Relapsing Remitting MS; RRÆSP: Relapsing Remitting patients who converted to Secondary Progressive MS; SP: Secondary Progressive MS; SPM: Statistical Parametric Mapping; SRTD: Selective Reminding Test (Delayed); Structural Imaging Evaluation of Normalized Atrophy from Oxford fMRI Laboratory Software Library (FSL); SIENAx: Cross-sectional version of SIENA with normalized volumes; VBM: Voxel-based Morphometry; WLG: Word List Generation; WMF: White Matter Fraction

Regional GM atrophy can correlate better with specific functional deficits. Using a voxel-based morphometry (VBM) method (Ashburner and Friston, 2000), early regional GM volume loss has been found in deep GM even at the first clinical presentation (Henry et al., 2008), supporting the ability of regional GM atrophy to detect sub-clinical damage. VBM can be used to explore the spatial relationship between local

GM tissue loss and focal damage (Ceccarelli et al., 2008; Morgen et al., 2006; Prinster et

al., 2006; Sepulcre et al., 2006; Sepulcre et al., 2009) and potentially diffuse NAWM

damage. However, VBM requires many statistical voxel-wise comparisons and is

believed to be less sensitive to regions where inter-subject variation is naturally high (e.g.,

cortex). Therefore, methods that measure the volume and its change directly from subject

images may be preferred over VBM methods when detecting cortical atrophy (Anderson et al., 2006; Hutton et al., 2009). Using such direct methods, several studies found correlations between cortical GM volume loss and disability (De Stefano et al., 2003) as well as cognitive impairment (Amato et al., 2004; Benedict et al., 2006; Portaccio et al.,

2006), and the cortical GM volume loss predicted most of the variance in cognitive

dysfunction (Benedict et al., 2006).

1.6.2 Cortical Thickness Measurements in MS

In addition to GM volumetric studies, recent development of sophisticated image

processing tools and high-resolution images have allowed measurement of cortical

thickness (CTh) (Aganj et al., 2009; Dale et al., 1999; Das et al., 2009; Goldenberg et al.,

2002; Han et al., 2004; Hutton et al., 2008; Jones et al., 2000; Lohmann et al., 2003;

MacDonald et al., 2000; Memoli et al., 2004; Miller et al., 2000; Rueda et al., 2010; Scott

et al., 2009; Srivastava et al., 2003; Xu et al., 1999; Yezzi and Prince, 2003; Zeng et al.,

1999). One study has shown that the brain size is more strongly associated with the cortical surface area than with CTh (Im et al., 2008). Since the association between brain

size and CTh is small, CTh does not strictly require head size normalization.

Most current CTh measurement techniques such as FreeSurfer (Dale et al., 1999),

registration-based CTh (DiReCT) (Das et al., 2009), cortical reconstruction using implicit

surface evolution (CRUISE) (Han et al., 2004), and Anatomic Segmentation using

Proximity (ASP) (MacDonald et al., 2000) require high-resolution images, which are not

typically acquired for MS patients, and these techniques cannot be applied in low- resolution MRIs. Some are longitudinal methods (e.g., version 4.0 or above of FreeSurfer

(Han et al., 2006) and DiReCT (Das et al., 2009), image gradient-based method (Chen et al., 2004), or 4D minimal line integral method (Li et al., 2010)). Han et al. have reported that a longitudinal CTh measurement method possesses smaller variability errors than the independent application of cross-sectional methods using FreeSurfer (Han et al., 2006)

suggesting that the longitudinal methods are preferred for measuring the CTh change.

One of the registration-based longitudinal CTh methods can be applied in low-resolution

MRIs, but the measurements tend to be restricted to regions where the cortical surface

borders CSF (gyral apices and open sulci) and may not cover the entire cortex (Chen et

al., 2004). Therefore, there is a need for a longitudinal method that can measure CTh

from the entire brain and can be applied to low-resolution T1SE MRIs.

Using current methods, alterations in CTh are reported in normal aging (Resnick

et al., 2003; Salat et al., 2004; Thambisetty et al., 2010) as well as in a variety of

pathologic conditions such as Alzheimer’s disease (Lerch et al., 2005), frontotemporal

dementia (Das et al., 2009), Huntington’s disease (Rosas et al., 2002), schizophrenia

(Kuperberg et al., 2003), and MS (Sailer et al., 2003).

In MS, a cross-sectional CTh study has found thinner cortex compared to NC

(mean CTh of 2.30±0.14mm and 2.48±0.11mm, respectively, p<0.001), particularly in

frontal and temporal lobes (Sailer et al., 2003). Like total GM tissue loss, CTh is reduced

early in the disease (Calabrese et al., 2007b), and early thinning is observed in frontal and temporal lobes and spread to other lobes in later advanced stages (Calabrese et al., 2010c;

Calabrese et al., 2007b; Sailer et al., 2003). In a large cross-sectional study, CTh is correlated with EDSS in 425 RRMS patients (Charil et al., 2007). Regionally, CTh correlates with a fatigue level (Pellicano et al., 2010) and cognitive impairment

(Calabrese et al., 2010c). Thinner cortical areas corresponding to the initial clinical symptom in CIS patients suggestive of MS are found (e.g., patients with optical neuritis onset have thinner occipital lobes, and patients with the pyramidal onset had the thinner motor cortex) (Calabrese et al., 2007b; Jenkins et al., 2011). There has been only one

longitudinal study of CTh in MS patients, and it has found that the rate of cortical thinning is greater in progressively disabling patients compared to stable patients (−3.13±

2.88%/yr and 0.06±2.31%/yr, p=0.002, respectively) in 30 MS patients over a year (Chen et al., 2004), further emphasizing that CTh is clinically relevant. However, long-term evolution of cortical thinning has not been reported in MS.

In summary, GM and cortical atrophy are potential surrogate markers of MS disease progression. Since whole brain atrophy may be significantly affected by inflammatory edema, and since most of the neurons are in GM, GM atrophy may be a more specific marker for irreversible tissue destruction than whole brain atrophy is. The temporal evolution of GM and cortical atrophy, as well as their clinical relevance in MS, need to be investigated. A better understanding of the changes in GM would provide more complete information about pathologic processes in MS than one could obtain using only whole brain atrophy measures. Important questions such as the relationship between

GM atrophy and focal WM lesions or between GM atrophy and diffuse WM tissue damage could then be addressed.

1.7 Objectives

My research had two aims: (1) to develop and validate an automated cross- sectional method to segment GM, to apply this segmentation method in a longitudinal study, to measure GM atrophy in terms of GMF, and to investigate the clinical relevance of GM atrophy; (2) to develop and validate a longitudinal method for measuring CTh and its change, to apply this method to the same longitudinal study, and to explore its clinical relevance in MS.

In Chapter 2, a new automated cross-sectional method to segment GM is

described. This segmentation algorithm is designed to be applicable in MRIs of MS

patients, which often exhibit significant atrophy and focal abnormal areas. In Chapter 3,

the developed segmentation algorithm is applied to baseline and fourth year data of a

longitudinal MRI study to characterize the evolution of overall GM tissue loss in a

variety of MS patients (CIS, CIS patients who converted to RRMS, RRMS, RRMS

patients who advanced to SPMS, and SPMS) with respect to the healthy normal controls.

In Chapter 4, a more sophisticated image processing algorithm, Cortical Longitudinal

Atrophy Detection Algorithm (CLADA), is described. CLADA is specifically designed

to measure the subtle temporal change in CTh. Unlike the segmentation algorithm in

Chapter 2, CLADA is a registration and deformable model-based “longitudinal” method and can achieve highly consistent results. Various validation tests are conducted to evaluate its reliability. In Chapter 5, the full longitudinal (semi-)annual MRI data is used to investigate the patterns of cortical atrophy in MS. Finally in Chapter 6, my work is summarized, and possible future directions are provided.

CHAPTER 2 GRAY MATTER SEGMENTATION

ALGORITHM

2.1 Introduction

This chapter describes the gray matter (GM) segmentation algorithm for the measurement of GM atrophy in MS patients. The general overview of the algorithm

(Section 2.1.3) is described first, followed by the detailed description of each component

(Section 2.2). Various validation tests and their results are reported (Section 2.3) next.

The developed algorithm will be applied in a longitudinal study in the next chapter.

The majority of content in this chapter was published in NeuroImage (Nakamura and Fisher, 2009).

2.1.1 MRI Format and Viewing

All MRI images were obtained as either DICOM or vendor specific format and were converted to an in-house format (BIP, Biomedical Image Processing format,

Cleveland Clinic). Digital image processing software was written using the BIP library in the C language and compiled on SGI Irix64 and Linux (RedHat and CentOS).

A customized MRI viewer was developed using GTK+ for orthogonal viewing of

BIP images on Windows and CentOS Linux. Unlike previous BIP viewers, which were

essentially 2D slice viewers, the new viewer showed 2 orthogonal planes (axial plus

sagittal or coronal planes) for better visualization of the cortex. The viewer was also

capable of showing color images, which facilitated visual verifications of segmentation

results.

2.1.2 Background

Measurement of GM atrophy using normalized GM volume requires

segmentation of GM tissue and calculation of GMF over time. Segmentation of brain

tissue from MRI is an extensively studied field and many automated software exist. Few

algorithms rely solely on image intensity (Schnack et al., 2001) because these approaches

are overly sensitive to image artifacts such as radiofrequency (RF) inhomogeneity, B0 inhomogeneity, and aliasing, and cannot adequately account for overlapping intensity distributions across structures. Therefore, to improve segmentation accuracy, most tissue segmentation algorithms combine intensity information with other techniques, such as the use of a priori anatomic information (Chalana et al., 2001; Van Leemput et al., 1999) or edge information through deformable contours (Davatzikos and Prince, 1995; Xu et al.,

1999; Zeng et al., 1999). Intensity information is analyzed differently in each approach, including Gaussian mixture models (Andersen et al., 2002; Ashburner and Friston, 2005;

Marroquin et al., 2002; Zhang et al., 2001), discriminant analysis (Amato et al., 2003), k- nearest neighbor classification (Mohamed et al., 1999), and fuzzy c-means clustering

(Ahmed et al., 2002; Pham, 1999; Suckling et al., 1999; Zhou and Bai, 2007; Zhu and

Jiang, 2003). The use of multiple images has advantages over a single image because the different contrasts can be enhanced between tissues. For example, FLAIR images have

desirable contrast between MS lesions and the normal-appearing brain tissue and can be

combined with other images to obtain GM and WM segmentation (Sajja et al., 2006).

There are a few widely available and commonly used brain tissue segmentation

methods that use both intensity and a priori anatomic information. These algorithms, such

as the segmentation tool in SPM (Ashburner and Friston, 2005) and FAST in FSL (Zhang

et al., 2001), have been designed for general use, and therefore, are not necessarily

optimized for specific pulse sequences or for application to images from patients with a

specific disease. The use of such general programs to segment MR images of MS patients

often results in misclassification of MS lesions as GM or CSF due to overlapping

intensities, which then requires time-consuming manual editing (Sanfilipo et al., 2005)

and introduces operator variability into the measurements. These methods are also prone

to classification errors due to partial volume effects between MS lesions and normal

tissue. Furthermore, for retrospective image analysis, where image data may not have

been acquired using optimal sequences for use with one of the widely available

segmentation tools, a customized segmentation method may be required to obtain the

most accurate results. SPM also requires high-resolution images of ~1×1×1mm3, and thus not well suited for typical clinical MRIs in MS, which typically have lower resolution

(slice thickness = 3-5mm). Therefore, there is a need for new segmentation method that can be applied in routine clinical MRIs of MS for accurate measurements of GM atrophy.

The normalization of GM volume in GMF accounts for the inter-subject variability of head size, and the normalization factor is calculated by the sum of GM,

WM and CSF volumes when CSF volume can be reliably measured (Equation 2-1). An alternative method suited for T1-weighted or FLAIR images is to use the outer contour

volume (OCV), which is the volume inside the smoothed brain parenchymal surface

(Equation 2-2).

[GMV] GMF = Equation 2-1 [][][]GMV WMV ++ CSFV

[GMV] GMF = Equation 2-2 []Outer Contour Volume

The annualized rate of GM atrophy is calculated in longitudinal studies by taking the difference of GMF from at least two time-points and dividing by the interval between the two scans.

2.1.3 Overview of Segmentation Algorithm

The overall flow chart, including preprocessing and GM segmentation steps, is shown in Figure 2-1.

Figure 2-1: Overall flowchart of the segmentation method: (1) image acquisition of FLAIR and T1 spin echo (T1SE); (2) co-registration of FLAIR and T1SE if needed; (3) segmentation of brain, MS lesions, and outer contour using FLAIR; (4) preprocessing of T1SE including intensity correction and de-noising; (5) fuzzy c-means clustering for the intensity-based probability map (PrMap); (6) affine standard space registration between subject’s T1SE and the Harvard Brain Atlas and the transformation of atlas-based anatomic probability maps; (7) morphologic probability from the brain mask; (8) combining the 3 probability maps and partial volume correction; (9) gray matter volume calculation; and finally (10) calculation of gray matter fraction.

Prior to GM segmentation, the brain is segmented from non-brain tissues on

FLAIR or proton density- / T2-weighted (PD/T2) dual echo images using a fully automated knowledge-based segmentation method, as previously described (Fisher et al.,

1997), and an example of segmentation results are shown in Figure 2-2. The outer contour of the brain (a smoothed surface that includes the ventricles and CSF surrounding the brain) is also segmented. The total brain volume (BV) and volume within the outer contour (outer contour volume, OCV) are calculated. BPF is calculated as the ratio of BV and OCV, BPF = BV / OCV. For MRIs of MS patients, focal T2 hyperintense lesions (T2

lesions) are segmented in the same software using a modified version of the iterated

conditional modes (ICM) algorithm (Besag, 1986).

Figure 2-2: Example of input axial FLAIR image (a), segmented brain segmentation (b), outer contour of the brain (c), and focal T2 hyperintense lesions (d).

The GM segmentation step involves T1-weighted images and relies on three

probability maps (Figure 2-1 steps 5-7): intensity-based probability, anatomic probability,

and morphologic probability. The probability maps are combined to create the tissue

masks for GM and WM, which are corrected for partial volume effects. Finally, gray matter fraction (GMF) is calculated as the ratio of GM volume and OCV (Equation 2-2).

2.2 Segmentation Algorithm

2.2.1 Image Correction

The image correction steps after segmentation of brain, lesions and outer contour,

include inter-slice intensity correction, non-parametric nonuniformity normalization (N3)

(Sled et al., 1998), and anisotropic diffusion filtering (Perona and Malik, 1990). The

inter-slice intensity correction algorithm estimates and reduces the signal loss due to the

B0 inhomogeneity in the superior and inferior slices. A single multiplication factor is

estimated for each slice by calculating the mode of the pixel-by-pixel ratios between

contiguous slices within the brain. The N3 algorithm corrects the smooth intensity

variation on MR images by deconvoluting the intensity histogram and mapping the

estimated ‘true’ intensity back to the image. The anisotropic diffusion filtering reduces

the noise while preserving edges. An example of intensity correction using inter-slice

intensity correction and N3 algorithms is shown in Figure 2-3.

Figure 2-3: The intensity variation in original image (a) was corrected with inter-slice intensity correction program and N3 algorithm (b).

In datasets with motion between interleaved sets, the images are corrected by

splitting, registering, transforming, and merging the interleaved sets. The registration

algorithm has 6 degrees of freedom (6-dof or rigid-body), uses normalized mutual

information (NMI) as the measure of image similarity (Pluim et al., 2003), and calculates

the optimal parameters using the simplex downhill optimization (Press et al., 2002). The

transformation is divided into 2 equivalent midway transformations so that the

interpolation artifact is equally distributed among the interleaved sets. Figure 2-4 shows interleaved axial FLAIR image with patient motion (left). The mismatch in interleave sets is visible in the sulci, hyperintense lesions, and brain boundaries (left) and is successfully corrected (right).

Figure 2-4: An example of interleaved motion artifact (left) and corrected image (right). The 2D axial FLAIR image was acquired with two interleaved sets from a MS patient who moved between the interleave sessions. The motion correction algorithm split, rigidly aligned, transformed, and combined the interleaved sets.

For the patient motion between acquisitions of FLAIR and T1-weighted images, the images are co-registered using the same 6-dof NMI registration method. Figure 2-5 shows an example of co-registration of FLAIR and T1-weighted spin echo (T1SE). The brain boundary is calculated from the FLAIR image and overlaid on T1SE. The pre- registration misalignment in (a) is corrected after co-registration using NMI registration

(b).

Figure 2-5: FLAIR brain mask boundary is overlaid on T1-weighted image before and after co-registration.

2.2.2 Fuzzy C-means Clustering Method

The first of three probability maps is derived from the grayscale intensities of T1-

weighted images. To calculate the intensity-based probability, the grayscale values for all

normal-appearing brain tissue (NABT) voxels (i.e., voxels included in the brain mask but

not in the lesion mask) are analyzed with a modified fuzzy c-means (FCM) clustering

method (Bezdek et al., 1993; Clark et al., 1994). FCM iteratively calculates the fuzzy

memberships and the cluster centers (mean tissue intensity) using the following

equations:

−1 ⎡ 2 ⎤ c ⎛ − vx ⎞ u = ⎢ ⎜ ik ⎟ ⎥ Equation 2-3 ik ⎢∑⎜ − vx ⎟ ⎥ j=1 ⎝ jk ⎠ ⎣ ⎦ n 2 ∑()xu kik k =1 Equation 2-4 vi = n 2 ∑()uik k =1

where uik is the fuzzy membership of tissue i at voxel k, calculated from xk, which is the

intensity at voxel k, and vi, which is the mean intensity for tissue i. There are three tissue

classes for GM, WM, and CSF. In addition to the above standard FCM, this algorithm

includes additional parameters to factor in non-Gaussian tissue characteristics and local information, as shown:

−1 ⎡ 2 ⎤ c ⎛ − vxw ⎞ ⎢ ⎜ iki ⎟ ⎥ Equation 2-5 uik = ∑ ⎢ j=1 ⎜ − vxw ⎟ ⎥ ⎝ jkj ⎠ ⎣⎢ ⎦⎥ N ' β Equation 2-6 uik ()1 β uik +−= ∑uikn N n=1

The additional parameters are w for weighting one tissue over the other to account for

various tissue intensity characteristics, and β, to account for the local fuzzy memberships.

The fuzzy membership with local information is u'ik, where uikn is the fuzzy membership

for class i of voxel k's local neighbor n. Since cortical GM could be very thin, a 2D local

region with 4 neighbors is used. The parameters β and w are empirically determined

based on the best results obtained as compared to manual tracing in terms of the

similarity index (Zijdenbos et al., 1994). The similarity index is calculated as below.

2TP Similarity Index= Equation 2-7 2TP +FP +FN

where TP is the number of voxels correctly classified as GM, FN is the number of voxels

incorrectly classifying as non-GM, and FP is the number of voxels incorrectly classifying as GM. The similarity index ranges from 0 (no overlap) to 1 (perfect match) with greater than 0.7 is considered good agreement.

Figure 2-6: T1-weighted MRI (a) and its intensity-based probability map from fuzzy c-means clustering method. The slice is at the same location as the FLAIR image in Figure 2-2.

2.2.3 Anatomic Probability Map

The second GM probability map, the anatomic probability map, is commonly

used in brain tissue segmentation (Ashburner and Friston, 2005). The probability maps

may come from an average of many brains or a single template. In this GM segmentation

method, the Harvard Brain Atlas is used as the template brain (Kikinis et al., 1996). The

GM and WM masks from the atlas are morphologically dilated and blurred for the GM

and WM probability maps.

Registration between subject image and the Harvard Brain Atlas template image

uses an affine 12-dof registration. Unlike typical standard space registration methods,

which only uses grayscale MRI from the subject and template, this registration method

uses both grayscale image and segmented brain mask. The 12 registration parameters are

estimated in five sequential steps as follows: (1) two rotations (RY and RZ) based on

subject’s left and right brain hemispheric symmetry on grayscale image; (2) global

scaling factor (SG) based on the ratio of characteristic brain radii (the peak positions in the brain radius histogram) from the subject brain and atlas brain; (3) x-rotation (RX) and

three translations (TX, TY, TZ) with fixed Ry, Rz, and SG; (4) three orthogonal scaling (SX,

SY, SZ) with fixed translation and rotation parameters; and (5) three skewing (KX, KY, KZ) with fixed translation, rotation and scaling parameters. The registration parameters in steps (3)-(5) are determined based on maximizing the overlap between the subject’s brain and the template brain. The anatomic probability map is transformed to the image space using trilinear interpolation. The sequential determination of each registration parameter and the use of both grayscale MRI and brain mask result in good alignment for an affine registration.

Figure 2-7: The registered (same slice location as in Figure 2-2) anatomic probability map derived from the Harvard Brain Atlas.

2.2.4 Morphologic Probability Map

The third GM probability map is the individualized morphological probability map and is created from morphologic models of cortical GM and deep GM. The anatomic probability map based on the affine template registration alone cannot account for significant atrophy, and the morphologic probability is designed to capture the structural shifts that occur due to marked ventricular enlargement common in MS brains.

Furthermore, because of lower contrast in deep GM than cortical GM on typical clinical

T1SE MRIs, a separate morphologic deep GM probability map is used to improve the

detection of deep GM.

The morphologic deep GM model consists of two 3D ellipsoids Figure 2-8(a).

Their shape and location are determined from the directions of the principal axes of the

left and right lateral ventricles and the brain’s centroid position. Defining each ellipsoid

requires 6 parameters, x0, y0, z0, a, b, and c as follows:

()− xx 2 ()− yy 2 ( − zz )2 0 + 0 + 0 = 1 Equation 2-8 a 2 b 2 c 2

where (x0, y0, and z0) are the ellipsoid’s centroid location and (a, b, and c) are the radii along 3 axes. These 6 parameters are pre-determined during the algorithm calibration using manually placed ellipsoids from 7 MRIs. After the training, the ellipsoid centroid is calculated as [x0, y0, z0] = Cbrain + T where Cbrain is the 3D coordinate of the brain centroid and T is the offset of [16.5, 23.5, 6.0]mm for the ellipsoid in the right hemisphere and

[−16.5, 23.5, 6.0]mm for the left ellipsoid. The ellipsoidal radii are also pre-determined using the length along the principal axes with scaling factors: (a,b,c) = [0.1031, 0.1355,

0.1362] L where L is the length of principal axes. Finally, each ellipsoid is transformed along the principal axis of the lateral ventricle on the same side. The ellipsoids cover all the deep GM structures (i.e. caudate, putamen, and globus pallidus) except for the diencephalon, which can be captured by the cortical GM model because it borders the brain surface.

The cortical GM model represents the probability of cortical GM as a function of the distance from the approximated brain surface Figure 2-8(b): the function is 100%- probability for distance ≤ 4mm and linearly decreases to 0% by 10mm. This brain surface is determined by the mid-sagittal plane and by the edges on initial brain segmentation excluding the lateral ventricle edges. The mid-sagittal plane is detected through minimization of left and right hemispheric intensity differences. The lateral ventricles is segmented by a series of morphologic operators (3mm-dilation, seed-fill, 1.5mm-opening, and 1mm-median filter) on the segmented brain mask.

Figure 2-8: Morphologic probability maps using the ellipsoidal model for deep gray matter region (a) and using the brain mask and mid-sagittal plane for cortical model (b).

2.2.5 Partial Volume Correction

The three probability maps are multiplied and normalized from 0 to 100% to calculate the posterior probability, and then GM and WM masks are created by setting thresholds at 50%. The GM volume is measured after partial volume correction (Santago and Gage, 1995), and gray matter fraction (GMF) is calculated as the ratio between GM volume and OCV as shown in (Equation 2-2). White matter fraction (WMF) is calculated as BPF – GMF. GM segmentation results are visually verified using a color image that includes partial volume effects (Figure 2-9).

Figure 2-9: Result of gray matter segmentation in blue, lesion segmentation in green and brain segmentation (non-brain in red).

2.3 Validation

Five tests were performed to evaluate the performance of the GM segmentation algorithm: (1) segmentation of simulated MRI datasets and comparison to the gold standard to determine accuracy; (2) segmentation of real MRI datasets and comparison to results from manual tracings as another evaluation of accuracy; (3) segmentation of scan- rescan images to determine the reproducibility; (4) comparison to other widely used segmentation methods; and (5) segmentation of the same image with simulated lesion growth to determine the effects of MS lesions on GM volume.

2.3.1 Accuracy Test with Simulated MRI

The simulated BrainWeb MRI data (Collins et al., 1998) was used to evaluate the accuracy of segmentation by comparing the segmented tissue masks to the gold standard segmentation. BrainWeb offers various noise levels (0-9%) and RF inhomogeneity levels

(0-40%). Twelve combinations of noise and inhomogeneity levels were used. The volumetric errors including absolute and percent difference were measured. Since the volumetric difference was a global measure and may average out to small errors, the similarity index (Equation 2-7) was also calculated to evaluate accuracy.

The volumetric error was 1.2% for GM volume at the default 3% noise and 20%

RF inhomogeneity level. For the same image, the similarity index for GM was 0.964 where similarity index of 0.7 or above is considered good agreement.

There was a large decrease in the similarity index when the noise level was increased from 7% to 9%. Nevertheless, the segmentation algorithm performed well in the presence of RF inhomogeneity at the various noise levels as evaluated by the similarity index (Table 2-1).

Table 2-1: Similarity indices from the accuracy test using simulated BrainWeb MRI at different RF inhomogeneity and noise levels RF inhomogeneity levels 0% 20% 40% Average 0% 0.974 0.975 0.967 0.972 1% 0.971 0.974 0.966 0.970 3% 0.962 0.964 0.956 0.960 5% 0.945 0.949 0.941 0.945 7% 0.911 0.920 0.913 0.915

Noise levels Noise levels 9% 0.866 0.873 0.858 0.866 Average 0.938 0.943 0.933 0.938

2.3.2 Accuracy Test with Manually Traced Real MRIs

MRIs from 3 MS patients and 3 normal healthy controls were used to evaluate the segmentation accuracy by using manual tracings in real MRIs as the gold standard. The

images were acquired on a 1.5T Siemens Vision scanner and included a FLAIR and

T1SE with the sequence parameters listed in Table 2-2:

Table 2-2: MRI sequence. Image type Sequence detail Voxel size 30 contiguous slices FOV = 172×230 mm T2-weighted fluid- Matrix size = 192×256 attenuated inversion TR = 6000 ms 0.9×0.9×5mm recovery (FLAIR) TI = 2000 ms TE = 105 ms NSA = 2 30 contiguous slices FOV = 172×230 mm Matrix size = 192×256 T1-weighted spin echo 0.9×0.9×5mm TR = 800 ms TE = 15 ms NSA = 1 Abbreviations: FOV = field of view, TR = repetition time, TI = inversion time, TE = echo time, NSA = number of signal average

The images were processed through the automated GM segmentation algorithm and, separately, GM was manually traced in each image.

The leave-one-out cross-validation approach was used to avoid testing and training on the same dataset. Because the optimization of the segmentation parameters (w and β) used the same manually traced images, these optimized parameters were not used for this part of validation. Instead, for each of the six cases, the parameters were optimized separately using only other five cases and excluding the test data.

An example of accuracy test with manually traced real MRI is shown in Figure

2-10.

Figure 2-10: An example of input image (a), manual segmentation (b), and automated segmentation (c) from MS patient. Visually, the automated segmentation method showed good correspondence compared to the manual segmentation. (d) There was an excellent correlation between the automated and manual volume measurements from 3 MS patients and 3 controls. The mean absolute error for GM volume was 3.1±2.6%, and the Pearson correlation coefficient was 0.993. (e) The Bland-Altman plot showed that there is a bias between manual and automated segmentations.

The mean absolute error (standard deviation) for GM volume was 3.1% (2.6), and the automated and manually derived tissue volumes for each image set were highly correlated (Pearson correlation coefficient, r = 0.993). The similarity index of automated segmentation compared to manually tracing was also good; the mean similarity index was

0.841 and 0.836 for controls and MS patients, respectively. Bland-Altman plot showed a bias (Figure 2-10e) where brains with larger GM volume had larger absolute difference in

GM volume. Since there is no intrinsic bias in the automated segmentation method, the

observed bias may be related to the manual tracing. A post hoc analysis showed that BPF and GM volume were correlated in these 6 subjects (r = 0.87). It was possible that with the 5mm slice thickness, manual tracings tended to be too conservative along the GM-

CSF border in deep sulci, and this conservative manual tracing might have led to an underestimation of true GM volume. While the accuracy of manual tracing is a common pitfall in validation, the correlation coefficient of 0.993 and similarity indices of 0.841 and 0.836 assured that our segmentation is accurate.

2.3.3 Reproducibility Test

For the reproducibility test, 9 MS patients were imaged at 3 different time-points within 2 weeks. Each image set was processed separately through the automated GM segmentation algorithm. The reproducibility of the algorithm was evaluated by the coefficient of variation (COV = standard deviation / mean) for GM volumes and GMF.

The COV for GM volumes and GMF was 1.0% and 1.1%, respectively (Figure

2-11). This reproducibility of approximately 1% appeared to be sufficient to detect GM atrophy in MS, which was approximately 1-2%/year.

Figure 2-11: (a) Gray matter fractions and (a) difference from the individual mean gray matter fraction from scan-rescan MRI of 9 MS patients. Each subject was scanned 3 times over a 2 week-period.

2.3.4 Comparison to Other Segmentation Methods

The results of the segmentation algorithm were quantitatively and qualitatively compared to the results of other common freely available segmentation algorithms:

FMRIB's Automated Segmentation Tool (FAST in FSL) (Zhang et al., 2001) and

Statistical Parametric Mapping (SPM) (Ashburner and Friston, 2005). For the BrainWeb images, which had a high-resolution 1mm isotropic voxel size, the similarity indices were compared to the published values (Ashburner and Friston, 2005; Ferreira da Silva, 2007).

For low-resolution T1SE images, there were no published values. Therefore, 3 methods

(the proposed method, FAST and SPM) were applied on T1SE from the 6 MRIs that were manually traced. The similarity index was calculated using the manual segmentation as the gold standard. As a qualitative assurance, a blinded neuroradiologist reviewed the segmentation results and scored them as 1 = excellent, 2 = good, 3 = acceptable, and 4 = unacceptable.

The similarity index measured from applying our GM segmentation method to

BrainWeb image (0.94) was higher than the published values: 0.93 and 0.90 for SPM and

FAST, respectively (Ashburner and Friston, 2005; Ferreira da Silva, 2007). Therefore, our algorithm performed slightly better than the commonly available segmentation methods using the BrainWeb image.

On T1SE images from 6 subjects, the average similarity index (s.d.) was 0.8381

(0.01), 0.7887 (0.04), and 0.7084 (0.08) for our method, SPM, and FSL, respectively. For

T1SE MRIs, our method out-performed other commonly used methods when compared to the manual tracings. An example of segmentation with input images is shown in Figure

2-12.

Figure 2-12: Example of comparison among 3 different segmentation techniques. (a) input FLAIR image; (b) input T1-weighted image; (c) SPM5 GM segmentation; (d) FSL GM segmentation; and (e) the developed method.

An experienced neuroradiologist was asked to perform a blinded review of the

GM segmentation results for each of the 3 methods for 4 brains. The results showed that the new method resulted in acceptable GM segmentation in all cases while some results from other methods were unacceptable due to the misclassifications of lesions as GM and non-GM classified as GM. The mean subjective scores demonstrated consistency with the quantitative evaluation (new method: 2.0, FSL: 2.5, and SPM: 2.5; 1 = excellent, 2 = good, 3 = acceptable, and 4 = unacceptable), showing at least in these images, the new method is superior to other commonly used segmentation methods.

2.3.5 Sensitivity to Lesion Size

Finally, the segmentation algorithm was tested for the effects of MS lesions in

WM on FLAIR images (T2LV). Masks of segmented MS lesions were morphologically dilated with 3D spherical kernels of various sizes (1.2, 1.5, 1.9, 2.1, 2.6, 2.8, 2.9, 3.5, and

3.9 mm) in order to simulate different sized MS lesions within the same MRIs. The dilated lesion masks were masked by the brain tissue mask to ensure that ‘lesion’ voxels did not extend into non-parenchymal space and masked out from the GM mask. The

automated segmentation was then performed on the same MRI repeatedly using different lesion masks and GM volume was measured. This testing was repeated in 18 different

MS patients to ensure that lesions of various starting sizes and locations were included.

The linear regression was calculated between GM volume and T2LV to measure the effect of enlarging MS lesions.

The linear regression analysis showed there was a clear and fairly consistent effect of total T2LV on the measured GM volume for each case; as the lesion size was systematically increased, the GM volume decreased (Figure 2-13). The lesion volume and the GM volume were inversely proportional with a mean (s.d.) slope of −0.26 (0.07).

Since T2LV are highly variable even in the same patient, this systematic effect suggested that the GM volumes need to be corrected.

Figure 2-13: Effect of lesion size on GM volume. From T2-weighted image (a), lesions were segmented and overlaid on T1-weighted image (b), which were subsequently dilated by (1.9mm in c and 2.9mm in d) spherical kernel. Using T1-weighted image (e), gray matter was segmented within the normal-appearing brain tissue as shown in (f) original segmentation, (g and h) segmentation after simulated lesion growth by 1.9- and 2.9mm-dilation, respectively. (i) The plot showed the effect of lesion size on automated gray matter volume measurements in 18 different MS patients. The average slope was −0.26.

2.4 Discussion

2.4.1 Segmentation Method

This chapter described a new, fully automated method for GM segmentation in brain MRIs of MS patients. The method combined probability maps derived from intensity information, anatomic information, and morphology and segmented GM from low-resolution clinical brain MRIs.

Effect of MS lesions

The use of both anatomic and morphologic probability maps was a unique aspect of this algorithm. These maps provided patient-specific information about the locations of cortical GM and deep GM structures that was more precise than the use of either the anatomic or morphologic map alone. This was a key step for correct tissue classification in MS brains since T1 hypointense lesions and their surrounding partial volume voxels would be misclassified as GM. The misclassification of lesion voxels was a common problem with the application of general-use brain segmentation software, such as SPM

(Ashburner and Friston, 2005) and FAST (Zhang et al., 2001), to segment GM in MS patients. Correcting the misclassification required time-consuming manual editing to obtain an acceptable segmentation (Sanfilipo et al., 2005).

The misclassification of lesions as GM was a potential problem for both cross- sectional and longitudinal studies. In cross-sectional studies, comparison of SPM-derived

GM volumes between different subjects without correcting misclassified GM voxels might lead to a result that is actually a composite of both GM and lesion volumes, which, in turn, might lead to misinterpretation of the results. As shown in Figure 2-13(i), initial

GM volumes were not strongly correlated to lesion volume, but uncorrected GM volumes might artificially appear to be correlated to lesion volume if the misclassification of T1 hypointense lesions was left uncorrected.

In longitudinal studies, this dependence of GM volume on lesion volume presented an even bigger problem because MS lesions are highly dynamic, and the misclassified lesion volume change might be even greater than the true GM volume change. By performing the lesion segmentation step separately, masking out the lesions, and combining intensity, anatomic, and morphologic probability maps to segment GM, the misclassification of MS lesions could be avoided.

Effect of atrophy

The use of two different probability maps, anatomic and morphologic, was also important for segmentation of MS brains because many patients had a significant degree of atrophy, leading to the enlarged lateral ventricles. For this reason, a simple affine transformation of an anatomic reference might not provide a useful estimate of GM location, particularly for deep GM structures. The morphologic probability map improved the estimate of GM location because it was derived directly from the patients’ image data and, therefore, the degree of brain atrophy for that individual was taken into account.

2.4.2 Validation Tests

Accuracy tests

In a comparison to published results using BrainWeb images, the new segmentation method performed comparably well (Ashburner and Friston, 2005;

Shattuck et al., 2001). Another accuracy test was the comparison to manual segmentation of real MRIs. This was done because the simulated BrainWeb images were very different from typical MRIs acquired for MS patients. The typical clinical images acquired for MS studies are low-resolution MRIs (slice thickness = 3-5mm) as in the images used here.

While the developed algorithm clearly agreed with manually segmented images, the numerical accuracy results were only moderately good. This was not surprising given that the “gold standard” for comparison was manual segmentation, which was known to be error-prone and highly subjective.

In fact, there was an observed bias in the comparison between manual and automated segmentation (Figure 2-10e) where the volume difference was greater for larger brains. This bias appeared to be caused by the differences in the ability to distinguish the GM-CSF border in deep sulci. Brains with low GM volume typically had a significant degree of whole brain atrophy and vice versa. These brains with significant atrophy had very large sulci with clear GM-CSF edges, whereas the brains without significant atrophy had ambiguous GM-CSF edges due to partial volume effects in the tight sulcal spaces. The brains that were manually traced for this validation study had a very wide range in brain parenchymal fraction (BPF = 0.74 − 0.87), and therefore a wide range in GM-CSF separation in the deep sulci, which might have made this bias evident.

Scan-rescan reproducibility test

The scan-rescan test was performed in order to evaluate the applicability of the new segmentation method for longitudinal studies of GM atrophy, wherein the extent of

GM tissue loss could be estimated by the difference in GM fractions obtained at 2

different time-points. The COV of approximately 1% from repeated scans obtained within 2 weeks demonstrated that the algorithm was highly reproducible and similar to other methods (0.4% (Benedict et al., 2006), 0.4% (Sastre-Garriga et al., 2005), 0.7%

(Chard et al., 2002a), 0.8% (Sanfilipo et al., 2005), 1.0% (Furby et al., 2010), 1.9% (Ge et al., 2001a)). This variability was greater than that of BPF, which was about 0.2% (Fisher and Rudick, 2003). The typical rate of GM atrophy in MS patients was estimated to be about 1-2% per year (Horakova et al., 2008; Sastre-Garriga et al., 2005; Tedeschi et al.,

2009; Tiberio et al., 2005; Valsasina et al., 2005) as shown on Table 1-3. Therefore, this method would be most appropriate for application to longitudinal studies of duration 2 years, or longer. For short-term studies, a more precise method for measurement of GM tissue loss would be needed (e.g., longitudinal registration-based methods (Chapter 4)).

Comparison to other methods

The developed method was directly compared to other typical segmentation methods (SPM and FAST). The results showed that while the new method performed equally well when applied to high-resolution images (i.e., BrainWeb), it out-performed the other methods when applied to low-resolution images. This was critical because typical MRIs acquired for MS are lower-resolution. Therefore, the new method would be applicable for retrospective analysis of images from MS patients such as existing clinical trial datasets, and would be capable of determining the rates of GM atrophy without acquiring new images.

Effect of T2 lesions on GM volume

The last validation test investigated if the GM volume was affected by changes in

T2LV. Unlike the issue of misclassified lesion voxels discussed above (Chard et al.,

2002a; Sanfilipo et al., 2005), this effect stemmed from the fact that unsupervised clustering algorithms were sensitive to the voxel intensities that were fed to it as input.

Thus, even a relatively minor change in the input voxels could have an effect on the final clustering result. In the developed algorithm, lesion voxels were masked out before the

FCM algorithm. Therefore, as the lesions grew or shrank, the actual voxels that were used as input in the intensity-based classification step would change. This was shown by the simulations to have a clear effect on the resulting GM volumes. If left uncorrected in a longitudinal study of MS patients, some portion of the change in GM volume would appear to be correlated to T2LV and its change, simply from this technical issue. A systematic test was performed to calculate the extent and significance of this association, so that the effects of T2LV could be estimated using the mean slope of the GM volume versus T2LV regression lines.

2.4.3 Conclusion

In summary, the results indicated that the new segmentation algorithm could be used for reliable measurement of GM atrophy in MS patients if the follow-up duration was 2 years or longer. It also suggested that the individual GM volume should be corrected for T2LV before statistical group comparisons. The issues inherent to the analysis of MRIs from MS patients had been addressed directly in the design of the algorithm.

CHAPTER 3 MEASUREMENT OF GRAY MATTER

ATROPHY IN MS

3.1 Introduction

In this chapter, the newly developed GM segmentation algorithm was applied to a longitudinal MRI study to characterize GM atrophy in MS patients with the most common phenotypes who had entered into a prospective longitudinal study. Specifically, the rates of GM atrophy were measured from patients with clinically isolated syndrome

(CIS), relapsing-remitting MS (RRMS), and secondary progressive MS (SPMS), and compared to the GM atrophy rate of age- and gender-matched healthy normal control

(NC) subjects. This chapter addressed the pattern of GM tissue loss in these patients over the course of 4 years. The correlations between GM tissue loss and other MRI measures of tissue damage, and between GM tissue loss and clinical worsening were also investigated.

The majority of content in this chapter was published in Journal of the

Neurological Sciences (Rudick et al., 2009) and Annals of Neurology (Fisher et al., 2008).

3.2 Materials and Methods

3.2.1 Subjects

MS patients and healthy NC were recruited from the Mellen Center for Multiple

Sclerosis Treatment and Research at Cleveland Clinic. MS patients were diagnosed according to the standard criteria (Polman et al., 2005). In detail, RRMS patients had two or more discrete relapses with significant neurological recovery in the prior 3 years, and

SPMS patients needed to experience continued deterioration for at least 6 months, with or without superimposed relapses with a prior history of at least two relapses. Patients were allowed to be on disease modifying treatment. CIS patients had an episode of neurologic dysfunction typical of MS (e.g., optic neuritis, transverse myelitis, and brainstem syndrome). Conversion of CIS to MS was based on a clinical relapse. Age-matched healthy subjects with a normal neurological examination were also recruited.

Clinical evaluations for patients included Kurtzke Extended Disability Status

Scale (EDSS) (Kurtzke, 1983), 25-foot timed ambulation (25FTW), 9-hole peg test

(9HPT), 3-second Paced Auditory Serial Addition Test (PASAT), relapse history, and medications. MS Functional Composite (MSFC) was determined from the 25FTW, 9HPT, and PASAT by normalization to a published MS reference group (Fischer et al., 1999).

Disease worsening was defined on MSFC as sustained change of 20% or more from baseline for any of the 3 components of the MSFC sustained for 2 consecutive visits and on EDSS as worsening from baseline by 1.0 EDSS point (or 0.5 point for those with

EDSS >=5.5 at baseline) also sustained at 2 consecutive visits. Thus, for both MSFC and

EDSS progression, worsening was required to persist for ≥6 months.

For group comparisons, patients were categorized into 5 groups using the baseline and the end-of-study classifications: (1) CIS patients, (2) CIS patients who converted to

MS, (3) RRMS patients, (4) RRMS patients who converted to SPMS, and (5) SPMS patients.

3.2.2 MRI

MRI was acquired axially on a 1.5-Tesla Siemens scanner and included a T2-

weighted, fluid-attenuated inversion recovery (FLAIR), proton density (PD)-weighted

images acquired with and without a magnetization transfer (MT) saturation pulse for

calculation of magnetization transfer ratio (MTR) map, and T1-weighted spin echo

(T1SE) images acquired before and after the injection of standard-dose gadolinium

contrast agent (Gd-DTPA, T1Gd). MRIs were acquired annually for NC and semi-

annually for the patients. Subjects were excluded if they had corticosteroid therapy within

2 months. The details of the image acquisitions are listed on Table 3-1.

Table 3-1: MRI image sequence for longitudinal study. Voxel Image type Sequence detail size TR = 6000s T2-weighted fluid-attenuated 0.9×0.9 TI = 2000ms inversion recovery (FLAIR) 5mm TE = 105 ms × Proton density-weighted 3D TR = 35 ms 0.9×0.9 gradient echo image without MT TE = 6 ms saturation pulse Flip angle = 3 degrees ×5mm TR = 35 ms Proton density-weighted 3D MT pulse: 7.68 ms 250 Hz Gaussian TE = 6 ms 0.9×0.9 gradient echo image with MT applied 1.5 kHz off resonance; H1 = Flip angle = 3 ×5mm saturation pulse 8.8 μT; once per TR interval degrees T1-weighted spin echo image TR = 800 ms 0.9×0.9 without contrast TE = 15 ms ×5mm Acquired 5 min after injection of T1-weighted spin echo image TR = 800 ms 0.9×0.9 gadolinium-DTPA (0.1 mmol/kg with contrast TE = 15 ms body weight) ×5mm

3.2.3 Image Analysis

Image analysis was used to estimate brain volumes including BPF, GMF, and

WMF. BPF was calculated using the previously described method as the brain volume

(BV) divided by the outer contour volume (OCV) (Fisher et al., 1997). The gray matter volume (GMV) was calculated using the new segmentation algorithm (Nakamura and

Fisher, 2009) and was adjusted to account for the lesion volume (T2-hyperintense lesion volume, T2LV) as derived in the previous chapter (Section 2.3.5): [GMV Adjusted] =

[GMV Measured] + 0.26 × [T2LV]. Thus, GMF was calculated as the final [GMV Adjusted] divided by the OCV. WMF was calculated as BPF minus GMF.

Other conventional measures were T2LV, T1-hypointense lesion volume (T1LV), and gadolinium-enhancing lesion volume (GdLV). Additionally MTR maps were calculated from the PD-weighted image pair acquired with and without an MT saturation pulse using Equation 1-1 (Dousset et al., 1992). The mean MTR of NABT (MTRNABT) and mean lesion MTR (MTRLesion) were calculated from the voxels included within the brain and T2 lesion masks, respectively. The mean MTRs were divided by the mean

MTR of normal-appearing WM (MTRNAWM) to calculate MTR contrast ratios (CR).

3.2.4 Statistical Analysis

The MRI measures were calculated for the baseline and 4-year data. Changes were calculated and compared among MS subgroups and healthy control subjects using analysis of covariance (ANCOVA). For categorical variables, a χ2 test was performed.

Spearman’s rank correlation coefficient (SRCC) was used for correlations between GM atrophy and clinical disability. Pearson’s correlation coefficient was used to assess correlations between GMF and WMF, and between GMF and age. To compare groups in

MSFC and EDSS progression status, t-tests were used. For this comparison, longer clinical data was used (mean = 6.6 years). To determine the predictors of GM atrophy, a multiple regression model was developed using a set of baseline MRI measures and their

4-year changes, and adjusted R2 was reported.

3.3 Results

3.3.1 Baseline Characteristics

Table 3-2 summarizes the baseline characteristics for each subject group as classified at the 4-year follow up.

Table 3-2: Baseline characteristics. CIS→ RRMS→ Characteristics NC CIS RRMS SPMS MS SPMS Number of subjects 17 7 8 28 7 19 44.9 Mean age (SD), yr 41.6 (8.1) 36.1 (7.1) 39.7 (8.4) 42.2 (7.0) 49.7 (7.4) (10.1) Female, (%) 10 (59) 6 (86) 5 (63) 23 (82) 4 (57) 14 (74) Mean disease 0.3 0.5 6.7 18.5 17.4 NA duration (SD), yr (0.1) (0.7) (5.1) (11.1) (5.0) Mean EDSS (SD) NA 0.9 (0.9) 1.2 (0.4) 2.0 (1.5) 4.8 (1.6) 5.4 (1.3) 0.36 0.51 0.39 −0.34 −1.04 Mean MSFC (SD) NA (0.37) (0.46) (0.62) (0.78) (1.49) Mean T2LV (SD), 20.7 42.4 43.8 NA 2.5 (2.3) 7.6 (8.3) ml (17.8) (24.1) (26.0) Mean T1LV (SD), 0.15 0.29 1.74 8.17 8.73 NA ml (0.22) (0.48) (2.49) (6.89) (9.00) Gd+ (%) 0 14 38 18 29 16

Mean MTR NABT 35.6 35.5 35.9 35.3 34.6 34.35 (SD) (0.75) (0.98) (0.81) (1.01) (0.35) (1.19) Mean MTR CR 0.89 0.92 0.92 0.88 0.90 Lesion NA (SD) (0.12) (0.05) (0.04) (0.03) (0.05) 0.862 0.861 0.855 0.840 0.810 0.801 Mean BPF (SD) (0.012) (0.008) (0.023) (0.027) (0.02) (0.04) 0.554 0.551 0.555 0.537 0.519 0.528 Mean GMF (SD) (0.015) (0.010) (0.016) (0.018) (0.017) (0.032) 0.308 0.309 0.300 0.304 0.291 0.280 Mean WMF (SD) (0.011) (0.009) (0.017) (0.016) (0.015) (0.016) Abbreviations: BPF = brain parenchymal fraction; CIS = patients who had a clinically isolated syndrome and did not meet the criteria for a diagnosis of clinically definite MS over 4

years; CIS→MS = CIS patients who converted to clinically definite MS over the course of 4 years; CR = contrast ratio; EDSS = Expanded Disability Status Scale; Gd+ = patients with gadolinium-enhancing lesions; GMF = gray matter fraction; MSFC = Multiple Sclerosis Functional Composite; MTR = magnetization transfer ratio; NA = not applicable; NABT = normal-appearing brain tissue; NC = healthy normal control; RRMS = relapsing-remitting MS patients (throughout study); RRMS→SPMS = relapsing-remitting MS patients who converted to secondary progressive MS over 4 years; SD = standard deviation; SPMS = secondary progressive MS patients (throughout study); WMF = white matter fraction.

3.3.2 Group Comparison of Gray Matter Atrophy

The annualized rates of WMF, GMF, and BPF are plotted in Figure 3-1 for each group. The figure shows that the rates of GM atrophy increased with MS stage in comparison to the NC: 3.4-fold greater in patients converting from CIS to RRMS, 8.1- fold greater in RRMS patients, 12.4-fold greater in patients converting from RRMS to

SPMS, and 14-fold greater in SPMS patients. The GM atrophy rates in the combined set of RRMS patients (CIS→RRMS and RRMS stable) and combined set of SPMS patients

(RRMS→SPMS and SPMS) were significantly greater than GM atrophy rate in NC (p =

0.05 and p = 0.005, respectively). In contrast, WM atrophy rates were similar in all MS subtypes, at approximately threefold greater than in healthy control subjects.

Figure 3-1: Plot of mean annualized rates of atrophy in white matter fraction, gray matter fraction, and brain parenchymal fraction. For abbreviations, see Table 3-2. (Fisher et al., 2008)

3.3.3 Clinical Correlation of Gray Matter Atrophy

GMF was correlated with clinical disability scores. At baseline, GMF had the strongest correlation with MSFC score (SRCC = 0.49, p<0.0001) while the strongest

WMF correlation was with 9HPT (SRCC = 0.63, p<0.0001). At the last visit, GMF correlation was greatest with MSFC (SRCC = 0.56, p<0.0001) and was similar between

RRMS and SPMS subgroups. GMF was also moderately correlated with EDSS for the group as a whole (SRCC = −0.48, p<0.0001). Correlations differed between the RRMS and SPMS groups for EDSS and for the individual components of the MSFC.

Figure 3-2: Plot of gray matter fraction and EDSS. There was a moderate correlation (Spearman’s rank correlation coefficient = 0.56, p < 0.0001). Red dots represent RRMS and blue dots represent SPMS. (Fisher et al., 2008)

Figure 3-3: Plot of gray matter fraction and MSFC. There was a moderate correlation (Spearman’s rank correlation coefficient = −0.48, p < 0.0001). Red dots represent RRMS and blue dots represent SPMS. (Fisher et al., 2008)

When MS patients were classified into those who worsened and remained stable according to the EDSS and MSFC scores over the duration of 6 years, the group with worsening disease had larger mean annualized rate of atrophy in GMF and WMF (Table

3-3 for EDSS and Table 3-4 for MSFC). This difference between worsening and stable patients was significant when the status was defined by MSFC whereas EDSS-defined

groups did not reach statistical significance. Atrophy in WM was similar with either of the worsening definitions (Rudick et al., 2009).

Table 3-3: Characteristics of subjects by EDSS progression. Characteristics Worse Stable p-value Number of subjects 21 42 Mean age (SD), yr 44.4 (7.24) 42.1 (9.83) 0.3 Mean disease duration at 10.8 (8.89) 10.5 (8.38) 0.88 baseline (SD), yr MS classification, (%) CIS 3 (14) 5 (12) 0.86 RRMS 11 (53) 25 (59) SPMS 7 (33) 12 (29) EDSS at baseline (SD) 3.10 (2.56) 3.36 (2.02) 0.69 MSFC at baseline (SD) −0.33 (1.36) 0.004 (1.04) 0.33 BPF at baseline (SD) 0.83 (0.04) 0.83 (0.04) 0.92 GMF at baseline (SD) 0.53 (0.03) 0.53 (0.02) 0.91 WMF at baseline (SD) 0.29 (0.02) 0.29 (0.02) 0.84 BPF % Δ/year (SD) −0.30(0.26) −0.26 (0.24) 0.65 GMF % Δ/year (SD) −0.34 (0.40) −0.25 (0.39) 0.38 WMF % Δ/year (SD) −0.28 (0.47) −0.25 (0.63) 0.86 Abbreviation: see Table 3-2.

Table 3-4: Characteristics of subjects by MSFC progression. Characteristics Worse Stable p-value Number of subjects 29 34 NA Mean age (SD), yr 44.6 (9.7) 41.3 (8.3) 0.15 Mean disease duration at 15.0 (8.9) 6.7 (6.0) < 0.001 baseline (SD), yr MS classification, (%) CIS 2 (7) 6 (18) < 0.01 RRMS 12 (41) 24 (71) SPMS 15 (52) 4 (11) EDSS at baseline (SD) 4.35 (2.05) 2.35 (1.91) < 0.001 MSFC at baseline (SD) −0.51 (1.34) 0.24 (0.84) 0.012 BPF at baseline (SD) 0.81 (0.04) 0.84 (0.03) 0.01 GMF at baseline (SD) 0.53 (0.03) 0.54 (0.02) 0.08 WMF at baseline (SD) 0.29 (0.02) 0.30 (0.02) < 0.01 BPF % Δ/year (SD) −0.38 (0.24) −0.20 (0.22) 0.01 GMF % Δ/year (SD) −0.40 (0.47) −0.18 (0.28) 0.03 WMF % Δ/year (SD) −0.22 (0.44) −0.30 (0.68) 0.57 Abbreviation: see Table 3-2.

3.3.4 Predictors of Gray Matter Atrophy

MRI predictors of atrophy were different between the RRMS and SPMS groups, and for GM atrophy and WM atrophy, as shown in Table 3-5. Changes in MRI measures over the course of the study were not predictive of BPF change in the SPMS group.

However, in the RRMS group, on-study changes were strongly predictive of concurrent change in BPF, accounting for 72% of the variance in atrophy rate in the RRMS patients.

The predictors of GM atrophy included both focal (ΔT2LV and baseline MTRLesion), and diffuse damage (baseline MTRNABT) in the RRMS group. Interestingly, however, no predictor was found for GM atrophy in the SPMS group.

Table 3-5: Predictors of atrophy from multiple regression analysis. Dependent RRMS SPMS variable Predictor STB p R2 Predictor STB p R2

MTRNABT,0 0.68 <0.0001 T2LV0 −0.93 <0.0001 ΔMTR 0.57 <0.0001 MTR −0.40 0.027 BPF, % Δ/year Lesion 0.72 Lesion,0 0.52 ΔT2LV −0.53 <0.0001 GMF0 −0.78 <0.001 ΔT1LV 0.30 0.01 MTRNABT,0 0.56 <0.0001 GMF, % Δ/year MTRLesion,0 −0.46 0.001 0.62 none ΔT2LV −0.46 0.0004 MTR 0.35 0.015 T2LV −0.64 0.001 WMF, % Δ/year NABT,0 0.43 0 0.32 ΔMTRLesion 0.63 <0.0001 WMF0 −0.43 0.02 Abbreviation: Δ = change; subscript “0” = baseline data.

3.4 Discussion

In this study, patients with CIS, RRMS, and SPMS were studied concurrently with age- and sex-matched NC subjects over 4 years in an observational protocol. The results showed that GM atrophy accelerated in patients with more advanced stages of MS.

The annualized rates of whole brain atrophy increased with disease stage, from less than

0.2% in CIS patients converting to RRMS to almost 0.4% in patients with SPMS.

Expressed as fold increase from NC, whole brain atrophy increased progressively from 2- fold normal in CIS patients converting to RRMS to 5.8-fold normal in SPMS patients.

WM atrophy was relatively constant at 3-fold normal across all disease categories. GM atrophy increased from 3.4-fold normal in CIS patients converting to RRMS to 14-fold normal in SPMS patients.

Although several short-term studies reported greater rates of GM atrophy than

WM atrophy in MS patients at various stages of disease (Chard et al., 2004; Dalton et al.,

2004; Oreja-Guevara et al., 2005; Sastre-Garriga et al., 2005; Tiberio et al., 2005;

Valsasina et al., 2005), an increasing contribution of GM tissue loss to whole brain atrophy was never seen. This finding suggested that the pathological process changed as the disease advanced. The results suggested that MS initially affected both WM and GM equally but became more prominent in GM over time. This was an important finding since MS was conventionally thought as a WM disease.

A second important finding was that GM atrophy correlated with disability.

Numerous prior reports have documented correlations between whole brain atrophy and neurological disability, and varying degrees of correlations between GM atrophy and disability, as shown in Table 1-3. The results here showed moderately strong correlations between GMF and MSFC, 25FTW, 9HPT, PASAT and EDSS. In addition to correlations between atrophy and disability scores, there was a trend toward a greater rate of GM atrophy, but not WM atrophy, in patients with sustained EDSS worsening. The findings of greater GM atrophy rates with increasing disease duration and advancing disease stage, together with the disability correlations, would underscore the clinical relevance of GM atrophy in MS and support the use of GM atrophy in clinical trials, particularly in SPMS.

A third important finding was that GM atrophy was more closely related to the disease progression status on MSFC than EDSS. This is not surprising since EDSS heavily depends on ambulation while MSFC includes a test of cognitive function where cognitive dysfunction is generally attributed to GM damage. This finding is consistent with previous reports, which found moderate to fairly strong correlation between GM tissue loss and cognitive tests (Amato et al., 2004; Benedict et al., 2006).

Lastly, the predictors of atrophy were different between RRMS and SPMS as well as between GM and WM. T2LV changes, MTRLesion, and MTRNABT accounted for a large proportion of the variance in whole brain atrophy and WM atrophy in both RRMS and

SPMS. These same markers explained some of the variance in GM atrophy in RRMS. In contrast, there were no observed correlates of GM atrophy in SPMS patients. It is currently unclear if neurodegenerative GM tissue loss in MS occurs as retrograde degeneration from damaged axons or as a process that is independent of focal WM pathology (e.g., cortical/meningeal inflammation or primary neurodegeneration) (Frischer et al., 2009; Geurts et al., 2009; Geurts and Barkhof, 2008; Pirko et al., 2007; Stadelmann et al., 2008; Vercellino et al., 2005). These results might shed light on the role of primary or secondary neurodegeneration in MS. The results here suggested that the processes responsible for GM atrophy changed over the course of disease. In the early stage of disease, GM atrophy appeared to be partly related to WM tissue damage. However, as the disease progressed, pathological processes resulting in GM tissue loss evolved but were no longer related to WM tissue damage because none of the conventional lesion or MTR measures reflecting WM damage predicted GM tissue loss in SPMS. It was possible that

GM lesions, not visible by conventional MRI, were the major contributor to GM atrophy in SPMS (Kutzelnigg et al., 2005).

One potential limitation in investigating GM atrophy in clinical trial is that the duration of typical MS clinical trials is shorter (1-2 years). In this 4-year longitudinal study, the average GM atrophy was approximately 0.3% per year, or 1.2% over 4 years, and the scan-rescan reproducibility was 1.0% (Section 2.3.3), indicating that this cross- sectional GM segmentation method was sensitive enough for detecting GM atrophy.

However, in shorter studies, such cross-sectional methods might not be sensitive enough in detecting GM atrophy or might require a large sample size. Therefore, a more robust longitudinal registration-based method might be needed.

CHAPTER 4 CORTICAL LONGITUDINAL ATROPHY

DETECTION ALGORITHM

4.1 Introduction

This chapter describes a new algorithm (Cortical Longitudinal Atrophy Detection

Algorithm, or CLADA), to measure cortical atrophy by measuring global and regional cortical thickness (CTh) and change in CTh using serially acquired MRIs. CLADA is designed to be sensitive to small changes in MRI. The chapter is organized as follows: description of CLADA (Sections 4.1-4.2), validation tests and results (Section 4.3), and discussion (Section 4.4).

The majority of this chapter was published in NeuroImage (Nakamura et al.,

2011)

4.1.1 Background

The measurement of CTh from brain MRI is a fairly new technique, and few algorithms have been published. Generally, these methods can be classified into voxel- based and surface-based methods. The voxel-based methods include the gradient-based approach (Chen et al., 2004), Eulerian partial differential equation (PDE) approach

(Acosta et al., 2009; Das et al., 2009; Hutton et al., 2008; Yezzi and Prince, 2003),

Laplace equation approach (Jones et al., 2000), level set approach (Han et al., 2004; Zeng et al., 1999), or minimal line integral method (Aganj et al., 2009). The gradient-based method uses intensity gradient perpendicular to the brain surface and can be robust against smoothly varying intensity inhomogeneity. The Eulerian PDE and Laplace equation-based methods define CTh as the trajectory between two surfaces (as in an electric field). The level set methods use an implicit explicit model where the voxel values represent the distance from the surface, and CTh can be calculated from the distance map. The minimal line integral methods define CTh as the minimal straight distance from one surface to another. Typically, the voxel-based methods are computationally fast but heavily dependant on the accuracy of segmentation, are difficult to impose constraints (e.g., surface smoothness and thickness smoothness), and require separate topological processing (Das et al., 2009; Han et al., 2003; Rueda et al., 2010).

On the other hand, methods that use explicit surface models (collections of points and polygons) (Dale et al., 1999; MacDonald et al., 2000; Xu et al., 1999) are time- consuming but can apply constraints to find the surfaces, are inherently topologically correct (MacDonald et al., 2000), and sub-voxel in their nature.

The definition of CTh differs among various methods and includes the distance along the surface normal (normal distance) (Chen et al., 2004), minimal distance (Aganj et al., 2009; Li et al., 2010; Miller et al., 2000), minimal distance averaged from both surfaces (Dale et al., 1999), distance from solving the Laplace equation (Laplace distance) (Hutton et al., 2008; Jones et al., 2000; Yezzi and Prince, 2003), and straight distance between corresponding points (link distance) (MacDonald et al., 2000)

These definitions can result in the same thickness values when two surfaces are parallel (Figure 4-1a where the red and yellow curves represent the surfaces; e.g., the yellow curve for GM-WM surface and the red curve for GM-CSF surface) or when their curvature is the same (Figure 4-1b). In other situations, thickness varies among the four definitions. The normal distance is intuitive, but the lines along the surface normal can cross each other when the curvature of the two surfaces is different (Figure 4-1c; white lines = normal distance from the yellow surface and black = link distance). In this case, the normal distance measured from the red surface will be similar to the link distance

(black). The normal distance does not guarantee the correspondence as shown in Figure

4-1f where white arrows do not return to the same point. Thus, the thickness measurements can be different in each direction (i.e., the distance from GM-WM surface to the GM-CSF surface versus the distance from GM-CSF surface to the GM-WM surface). The minimal distance is simple but also does not guarantee the correspondence.

The direction of the distance may not be perpendicular to the surface (Figure 4-1d, white lines = minimal distance and black lines = link distance). A more mathematical definition is the Laplace distance, which guarantees that the direction is perpendicular to the surface at the surface and that the end points correspond to each other (Figure 4-1e; white curves

= Laplace distance and black lines = link distance). The curve connecting the surface points can bend according to this definition. On the other hand, the link distance does not curve and may not be perpendicular to the surface (Figure 4-1a-e, black lines).

Figure 4-1: Schematic 2D examples of different measures of thickness (distance along the surface normal, minimal distance from one surface, distance between corresponding points, and distance from solving the Laplace equation) in various situations where two surfaces are represented by yellow and red curves. (a) When the surfaces are parallel, thickness is the same among the 4 methods. (b) When the curvature on the 2 surfaces is the same, thickness still remains the same among the 4 methods. (c) When the surfaces are not parallel, the distance along the surface normal from yellow surface (white) and the distance between corresponding points (black) may differ. Note that the distance along the surface normal from the red surface will be similar to the distance between corresponding points (black). (d) When the curvatures on 2 surfaces are different, thickness using the minimal distance (white) from yellow surface may be different from the distance between corresponding points (black). Note that the minimal distance from the red surface is not necessarily the same as the minimal distance from the yellow surface. (e) When both surfaces are convex, the distance from Laplace equation (white) guarantees that the curve is normal to both surfaces at the surface whereas the straight distance between corresponding points is not normal to the surface. Note that the distance from solving Laplace equation can curve. (f) The non-correspondence is shown here as the arrows do not return to the same point.

One study has shown low sensitivity with the minimal distance measure and high variability with the normal distance measure while the link distance is the most reliable

(Lerch and Evans, 2005).

CLADA is an explicit surface method and its CTh is measured using the definition of the link distance. Therefore, CTh is measured at each time-point on each corresponding pair of vertices. Finally, the rate of change in CTh is calculated for the

measurement of cortical atrophy.

4.1.2 Overview of CLADA

A simplified flowchart of CLADA is shown in Figure 4-2. CLADA combines images from all time-points, creates averaged template image, generates a model of the cerebral cortex consisting of 2 explicit surfaces (for inner GM-WM interface and GM-

CSF interface, Figure 4-12(d) and (e) respectively) based on the template image, and then deforms the model for each individual time-point for the measurements of CTh and CTh change.

Figure 4-2: Simplified overall flowchart of cortical longitudinal atrophy detection algorithm (CLADA), starting from serial acquisition of T1-weighted images, registration, averaging, segmentation, subject-specific deformable model, longitudinal deformation, and finally cortical thickness measurements.

4.1.3 Deformable Model Format and Viewing

Explicit surfaces are modeled with Object Oriented Graphics Library (OOGL) in

OFF file format and implemented in the C language. The OFF file format essentially is a collection of 3D points (vertices) and polygons represented by the vertex indices. In

CLADA, all surface polygons are triangular to ensure a unique normal vector for each polygon. An edge file, which saves the list of neighboring vertex indices, is additionally

created to reduce the computation time. The OFF files are rendered and viewed in geomview2.

4.2 Algorithm Description

4.2.1 Image Acquisition

CLADA requires serially acquired T1-weighted images as input. CLADA can process both high-resolution MRI (e.g., magnetization prepared rapid gradient echo

(MPRAGE), fast low angle shot (FLASH), or spoiled gradient recalled (SPGR)) and low- resolution 2D images (e.g. T1 spin echo (T1SE)), but the consistency in serial images is critical (i.e., scanner, scanner hardware, software, and imaging sequence). Additional imaging modalities such as FLAIR or T2-weighted images are helpful, but not required, for differentiating the brain, non-brain and abnormal MS lesions.

4.2.2 Pre-processing

Input images are first visually inspected and re-oriented to the default BIP image direction (image orientations of x = left-to-right, y = anterior-to-posterior, and z = superior-to-inferior) as shown in Figure 4-3. Image artifacts (e.g., motion, aliasing, ghosting, and ringing) are individually inspected and recorded. For 2D interleaved sequences, the interleaved patient motion is corrected as in the pre-processing for GM segmentation (Section 2.2.1).

2 geomview: http://www.geomview.org/

Figure 4-3: An example of the averaged T1-weighted image in a CLADA-reference space. From left to right, axial, coronal and sagittal planes.

The images are then corrected for B0 and B1 field inhomogeneity by 3 different methods. The first intensity correction recovers the signal drop-off by determining a multiplication factor for each slice, as previously described in Section 2.2.1. The second pre-processing is the N3 correction and also previously described (Sled et al., 1998).

Thirdly, a parametric bias field correction (PABIC) is applied (Styner et al., 2000). This algorithm estimates the true intensities by a third-order parametric Legendre polynomial model. The image intensities are initially normalized using a simple scaling during preprocessing and more accurately using a differential bias correction method at the longitudinal deformation (Lewis and Fox, 2004). This differential bias correction do not necessarily eliminate or reduce the intensity nonuniformity in each image, but the resulting bias fields in serial images become more consistent, thus making the measurement of longitudinal atrophy more reliable.

The baseline image is rigidly transformed to a reference space as in Figure 4-3.

Since CLADA is a cerebral cortical analysis algorithm, the partial coverage for the cerebellum is not a problem. The same standard space registration method is used as in

Section 2.2.3.

4.2.3 Registration and Averaging

During this step, serially acquired images are registered and averaged to create a template image and to improve the signal-to-ratio. For intra-subject registration, non- brain constrained symmetric registration algorithm (NBCSR) is used to register the follow-up images to the baseline image (Chen et al., 2008). Briefly, NBCSR involves 6 steps: (1) whole head rigid-body registration, (2) brain-only affine registration, (3) affine registration in cropped nonbrain region where the top of skull and neck were cropped to eliminate variable appearance, (4) brain-only 6-degrees-of-freedom registration with fixed scaling and skewing from the previous step, (5) repeating steps 1-4 in the backward direction and (6) combination of the resulting forward and inverse transformations to achieve symmetric registration. The correlation coefficient is used as the cost function.

This registration matrix is multiplied to the baseline-to-reference registration matrix, and all images are transformed with spline interpolation in a resampled reference space of

0.75×0.75×0.75mm3 (Figure 4-4).

The field-of-view (FOV) masks are also created and transformed to the same reference space to remove an artifact of averaging missing voxels in case of incomplete brain coverage. Therefore, T1-weighted images from different time-points are added voxel-by-voxel and divided by the sum of the FOV masks for an average image. If

FLAIR or T2-weighted images are acquired, they are also registered and transformed to the reference image using NBCSR with normalized mutual information as the cost function.

), corrected, registered, and averaged images. images. averaged registered, and ), corrected, Example of the input (months 0, 6, 12, 18 and 24 18 and 6, 12, 0, (months of the input Example Figure 4-4: 4-4: Figure

4.2.4 Segmentation

The goal of the segmentation here is to create a cortical GM-WM boundary mask

(Figure 4-5i-j for axial and coronal views), which will be used to initialize the explicit

GM-WM surface in the next step. The overall scheme of segmentation is similar to

FreeSurfer (Dale et al., 1999) in that the image was segmented into GM, WM, and CSF and a cortical GM–WM boundary is generated through a combination of segmentation and atlas-based classification (Dale et al., 1999). For consistency and efficiency in

CLADA, this segmentation step is carried out only once for each subject rather than for each of the images separately.

The subject-specific average image is segmented using the method from Chapter

2. The segmented brain mask (Figure 4-5b) is further processed using morphologic operations to fill in the lateral ventricles as described in Section 2.2.4 (Figure 4-5c). Next, the GM-WM boundary mask, which contains the cerebral white matter, lateral ventricles, and deep gray matter, is created for initial surface deformation. The deep gray matter structures (caudate, putamen, and globus pallidus), cerebellum and brainstem are isolated once in the Harvard Brain Atlas (Kikinis et al., 1996). ART (Automatic Registration

Toolbox, version 2009-03-123 nonlinear registration is used to register the subject image and the atlas (Ardekani et al., 2005). Using the resulting transformation, the masks for the cerebellum and brainstem are nonlinearly transformed to the subject space, disconnected, and eliminated from subsequent analysis steps (Figure 4-5 g and h, respectively). The deep gray structures are added to the GM-WM boundary mask; the transformed masks of caudate (red) and putamen (turquoise) are shown in Figure 4-5f and the segmented WM

3 http://www.nitrc.org/projects/art/

mask in Figure 4-5e. If necessary, MS lesions are segmented on FLAIR or PD/T2 dual echo images using iterative conditional modes (Besag, 1986) and added back into the

GM-WM boundary mask (Figure 4-5d). On MRI of MS patients, it is critical that juxtacortical lesions are properly segmented so that the surface deformation into lesions is avoided (Figure 4-5 k and l; the inlets show zoomed areas of the same juxtacortical lesion with axial and coronal views, respectively). Typical subcortical and periventricular lesions do not interface with the cortex and do not affect the initial deformation. Manual editing is performed to correct any remaining segmentation errors due to lesion holes (if no FLAIR image is available) and misclassification of blood vessels as white matter, which may occur in low-resolution images due to partial volume effects.

Figure 4-5: Examples of detailed segmentation steps in CLADA. (a) Averaged MPRAGE, (b) brain mask overlay, (c) lateral ventricle overlay, (d) MS lesion overlay, (e) white matter overlay, (f) coronal view of deep gray matter structures (red = caudate and turquoise = putamen), (g) cerebellum overlay, (h) brainstem overlay, (i,j) axial and coronal views of the final GM-WM boundary mask, and (k,l) axial and coronal views of the same juxtacortical lesion, which require proper segmentation as shown by the inlets.

4.2.5 Deformable Model

As in other explicit surface models (Dale et al., 1999; MacDonald et al., 2000), two explicit surfaces, one for the inner cortical surface (ICS, between GM and WM) and

one for the outer cortical surface (OCS, between GM and CSF) are constructed next. The

OCS is difficult to detect due to small separations between opposing surfaces within tight sulci and partial volume effects, and thus the ICS is created first with the GM-WM boundary mask (Dale et al., 1999; Teo et al., 1997).

The approach used in CLADA combines some aspects of both FreeSurfer (Dale et al., 1999) and anatomic segmentation using proximities (ASP, (MacDonald et al., 2000)).

Instead of simultaneously deforming the two surfaces from the start, as in ASP, a single spherical surface shrink-wraps to the GM-WM boundary mask first, to define the ICS, then it deforms outward to define the OCS. Only after the two surfaces are generated, simultaneous deformation is applied to both surfaces.

An example of deformation is shown in Figure 4-6 starting from (a) sphere, (b) intermediate shrink-wrap with long 1.5mm edges, (c) ICS after shrink-wrap, (d) initial

OCS, and (e) final subject-specific cortical surface model after simultaneous surface deformation. The shrink-wrap, OCS initialization and simultaneous deformation are described next.

Figure 4-6: Example of surface deformation for subject-specific dual cortical surface model. The top row shows the rendered deformed surface, and the bottom row shows image with the surfaces shown in yellow and red for inner cortical surface and outer cortical surface, respectively. From left to right, (a) initial spherical surface (yellow), (b) after the shrink-wrap at large edge length (2mm), (c) further shrink-wrap with 1mm-edge length, (d) initial estimate of outer cortical surface, and (e) final subject specific cortical deformable model after simultaneous deformation.

Shrink-Wrap Method for Inner Cortical Surface

CLADA's shrink-wrap algorithm is initialized from a sphere (Figure 4-6a) and employs (1) multi-resolution remesh function, (2) method for fast surface intersection detection, and (3) method for surface relaxation on proximal surfaces to improve efficiency and accuracy.

For remeshing, the algorithm described by Botsch and Kobbelt is implemented to manipulate the mesh configuration and distribution (Botsch and Kobbelt, 2004). Briefly, this remesh algorithm iteratively (1) splits edges longer than 4/3 of the target length, (2) collapses edges shorter than 4/5 of the target, (3) flips edges to reduce the variations in the number edges among vertices, and (4) tangentially smoothes the surfaces. For efficiency, a multi-resolution approach is implemented by decreasing the target length from 5 mm to 1 mm. The algorithm is computationally efficient, maintains consistent topology, has small geometric distortions due to its use of an iterative tangential relaxation, and results in consistent distribution of vertices (i.e., 6 or 7 edges per vertex and approximately 60°-angles in many triangles). This remesh function reduces the variability in the number of vertex neighbors, as seen by typical marching cube algorithms (Lorensen and Cline, 1987).

A fast triangle–triangle intersection test is used to detect self-intersections (Moller,

1997) and to prevent physically impossible overlaps. The intersection tests are performed

100

with fast recursive bounding boxes (Zomorodian and Edelsbrunner, 2000) to decrease the computation time. Figure 4-7 shows the log-plot of the computation time versus the input object’s number of vertices (approximately ranging from 3,200 to 320,000). The bounding boxes reduces the computation time from O(n2) to O(n).

Figure 4-7: A log-scaled plot comparing the computational times from the basic self- intersection test method and improved method with bounding boxes (BBox).

The surface relaxation on proximal surfaces minimizes the local surface area by smoothing when an object has a tunnel, a common artifact in low-resolution images with a significant partial volume effect. The algorithm detects proximal surface regions by extending the surface deformation and testing for surface intersections but not actually applying the deformation. If a potential surface intersection is detected, a local smoothing operation is performed on this surface region by averaging the neighboring vertex positions.

A Laplace map is created from the brain and GM-WM boundary masks to guide the direction of the deformation (Figure 4-8). It is particularly helpful in deformation into deep sulci. This map is similar to the gradient vector flow (Xu and Prince, 1998) and is a

101

non-increasing contour map with its maximum at the edge of initial brain mask and the minimum at the edge of the target GM-WM boundary mask. The Laplace equation is a second-order PDE of a scalar field “u” and solved using a finite difference method by approximation (Haidar et al., 2005).

Δu = 0 ∂2u ∂2u ∂2u N Equation 4-1 2 + 2 + 2 = ∑ un ∂x ∂y ∂z n=1

where Δ is the Laplace operator, u(x,y,z) is the 3-dimensional scalar field, and N is the number of neighbors, which is set to 6 in CLADA. In a matrix form, this equation becomes

v v = buA where

⎧ 1 if on boundary ⎪ Equation 4-2 ⎪ 6 if == ji Aij = ⎨ −1 ⎪ if == ji neighbor ⎪ ⎩ 0 otherwise

and where the matrix A is an NV×NV sparse matrix, NV is the number of voxels within the brain mask, b is the value of u(x,y,z) for boundary condition or 0 if inside the object, and matrix elements can be either 1, 6, −1, or 0. The scalar field, u(x,y,z), is calculated for the voxels inside the brain but outside the GM-WM border mask. At the brain boundary, u(x,y,z) is set to a large number (e.g., 65535) and the GM-WM boundary u(x,y,z) is set to

0. The masks are resampled to even higher resolution at 0.4mm isotropic voxel, so the exact solution of this matrix at higher resolution requires an inversion of a 2million-by-

2million matrix, which is not practical. Therefore, a conjugate gradient method is used to approximate the solution of this matrix. An example of the Laplace map is shown in

102

Figure 4-8. After the calculation of the Laplace map, sobel directional filters are applied to calculate the direction of deformation (shown by small black vectors on Figure 4-8 b).

Figure 4-8: (a) The log10-transformed Laplace map, log10(u(x,y,z)), is shown and color-coded from dark blue (u=0 around GM-WM border), green (u=32768), yellow (u=45875) to dark red (u=65536 near the b rain surface). (b) The sobel dir ectional filters were applied on the Laplace map to calculate direction of surface deformation (small black vectors). The region came from the medial frontal lobe (white rectangle in (a)).

At the end of the shrink-wrap, the ICS typically contains approximately 200,000 vertices and matches the GM-WM boundary mask (Figure 4-6c).

Outer Cortical Surface

After ICS detection, the initial OCS is estimated by iteratively expanding ICS outward (Figure 4-6d). A target OCS vertex is established for each ICS vertex using the surface model, averaged image, and sobel-filtered gradient image according to the following rules: (1) the intensity at the OCS vertex must be above the estimated mean

OCS intensity, μOCS, which is the mean intensity of GM-CSF edge voxels. The edge voxels must have high gradient, above the 75th percentile for all gradients in the brain

103

mask, and intensity between mean GM and CSF intensities (obtained from segmentation).

(2) The first derivative of the intensity profile along the OCS surface normal must be negative (i.e., decreasing intensity). (3) The OCS must not intersect another surface. (4)

The OCS vertex must be contained within the brain mask, but not in the cerebellum or brainstem masks. (5) The distance between the ICS vertex and the target OCS vertex must be less than predefined distance of 3 mm. Although the cortex is thicker than 3 mm in some regions, the initial estimate is confined to 3mm to provide a stable and realistic starting point that is close to the final surface. As in ASP (MacDonald et al., 2000), the correspondence between each inner surface vertex and outer surface vertex is maintained for calculation of CTh.

Simultaneous Surface Deformation

Finally, the two cortical surfaces are deformed simultaneously with an additional internal energy term to ensure CTh smoothness:

∂v = αFsurf+ βF thck+ γF imag Equation 4-3 ∂t F = v − v surf local Equation 4-4 ˆ Fthck = CTh local n − v Equation 4-5 λ Fimag = {}2HI()()v− μCS −1 Equation 4-6 ∇Iv + λ () where ∂∂ tv is the iterative deformation of 3D vertex position, v, determined by the internal surfa ce smoothness force (Fsurf), internal thickness smoothness force (Fthck), external image force (Fimag) and the weighting constants (α, β, γ). The surface

104

smoothness term (Fsurf) pulls the vertex towards the local average point (vlocal); similarly, the thickness smoothness pulls the vertex towards the point that corresponds to the

ˆ averaged local thickness (CThlocal) along the surface normal ( n ). The image force is determined by the image gradient at the vertex (∇I(v)), a constant lambda, λ, which regulates the gradient, and a Heaviside function (H) of image intensity (I(v)) and estimated edge intensity (μCS). Lambda (λ) is the 95% value of the image gradient within the brain. The weighting constants, α, β and γ, are set to 0.5, 1.5, 2.0, respectively, as determined through trial-and-error to produce good results. The integral solution is calculated by the fourth order Runge-Kutta method, and the deformation iterates until the deformation size (ΔC) is smaller than predefined threshold (ΣΔCn/dn/N < 1%) where dn is the distance between n'th vertex on ICS and OCS and N is the total number of vertices.

The resulting model is the subject-specific cortical model based on the averaged image and is used as initialization for longitudinal deformation in individual images. The remesh function is not applied so as to maintain consistent ICS-OCS correspondence.

4.2.6 Longitudinal Deformation

The subject-specific cortical model is subsequently deformed to fit the images from individual time-points. The cortical model rather than the image is transformed to prevent interpolation artifacts (Han et al., 2006). In this step, the remesh function is not applied in order to ensure that consistent ICS–OCS correspondence is maintained. The same deformation process as described above is applied iteratively for each time-point until the change is below a predefined 1% threshold as before.

105

4.2.7 Measurement Cortical Thickness and Change

Cortical thickness is determined for each surface vertex as the distance between corresponding vertices on the ICS and OCS surfaces (link distance). The global CTh is estimated as the surface-area-weighted average: CThglobal=ΣAndn/ΣΑn (where An is the triangular surface areas surrounding the n'th ICS and OCS vertices). Although the superior aspect of the brainstem as well as the diencephalon, hippocampus and amygdala are a part of the ICS and OCS surfaces, these non-cortical regions are excluded from the global CTh measurement by masking with the registered Harvard Brain Atlas (Kikinis et al., 1996) regions. For regional analysis, the atlas is transformed and filtered (in surface space, Figure 4-9). Surface-area-weighted regional CTh is measured for each anatomic region in the same manner as for global thickness. Regional labels remain constant across time-points for consistent longitudinal analysis.

Figure 4-9: Nonlinearly transformed and filtered regional labels for the frontal lobe (blue), parietal lobe (yellow), temporal lobe (orange), occipital lobe (red), and insula r area (cyan). The regional cortical thickness and thinning are measured by surface-area-weighted average within each label. The color is mapped on the inner cortical (GM-WM) surface (d).

106

4.3 Validation

Individual elements of the CLADA’s deformable model were tested using simulated data prior to testing on real brain MRIs. These tests were done to demonstrate efficacy of the self-intersection test, surface relaxation step used for tracking the convoluted inner surface, and the Laplace-based shrink-wrap algorithm. Further tests were performed to evaluate CLADA in terms of its (1) scan-rescan reproducibility (2) intra-rater reproducibility, (3) accuracy, (4) sensitivity to change in CTh, (5) sensitivity to input image type, (6) sensitivity to reference time-point, (7) comparison with FreeSurfer,

(8) bias in estimating atrophy, (9) comparison to a cross-sectional method, and finally

(10) effect of MS lesions on CTh measurements.

4.3.1 Phantom Experiments

Tests using simulated data and brain MRIs demonstrated the ability of CLADA to track convoluted surfaces (Figure 4-10). While deformation from a sphere to a smaller sphere did not cause any problems, even without the self-intersection test, a sphere with a tunnel caused problems when the surface was deformed without testing for self- intersections because the surface continued to grow on opposite sides. Although physically impossible in actual brain WM, such tunnels and disconnections can result from segmentation errors. This could be particularly problematic for low-resolution MRIs with significant partial volume effects. For spheres with straight tunnels, deformations that included the self-intersection test resulted in essentially the same final surface with or without the inclusion of the relaxation step. However, if there was a narrowing in the center of the tunnel, similar to the shape of the cortex when WM voxels are misclassified

107

due to partial volume effects, the result of the self-intersection method alone will depend on the position of the initial object. The relaxation step moved the self-intersection point by relaxing and minimizing the surface area, thus reducing the dependence on the initial condition.

Figure 4-10: Effects of including or excluding the self-intersection test and surface relaxation steps shown as a collection of cross-sections of 3D objects (black) with the initial starting point of the deformable surface (blue) and the resulting surface after deformation (red). The images show cross-sectional views of a deformation starting from the blue object into the red object with the binary image in black. Top row: With a sphere and well-located initial object, all methods perform equally well. Second row: When there is a tunnel in the middle of the sphere, the self-intersection test is needed to stop the deformation. Otherwise, the surfaces continue to move toward the other side. Third and fourth rows: When the tunnel narrowing is located at the center, the same results are obtained with or without relaxation. However, when the center is shifted to the right, as in the fourth row, the relaxation method accurately shifts to the right with the narrowing.

108

Next, the use of the Laplace map to determine the direction of deformation was evaluated by another phantom experiment (Figure 4-11). Given an initial surface (blue) and a target object (gray “L”-shaped object), the surface iteratively shrank inward as shown by colored contours (b). “t” is the number of iterations and colored coded from blue (t=0), light blue (t=9), pink (t=17), orange (t=21), to red (t=29). The surface eventually self-intersected and could no longer move at t=29. Using the gradient of

Laplace map (c, short vectors) in addition to surface normal, the deformable surface shrank fairly smoothly towards the target object (d).

Figure 4-11: The use of Laplace map was evaluated in a phantom experiment. Given an initial blue surface and L-shaped target object (a), the surface iteratively shrank inward as shown by colored contours (b). t is the number of iterations and colored coded from blue (t=0), light blue (t=9), pink (t=17), orange (t=21), to red (t=29). The surface eventually self- intersected and could no longer move at t=29. Using the gradient of Laplace map (c, short vectors) in addition to surface normal, the deformable surface shrank fairly smoothly towards the target object.

4.3.2 MRI Experiments

109

The typical CLADA result on actual MRI is shown in Figure 4-12.

Figure 4-12: Example of CLADA images showing (a) averaged T1-weighted MPRAGE images, (b) inner cortical surface (yellow) and outer cortical surface (red), (c) color-coded regional patterns of atrophy and growth in this subject, and (d and e) rendering of inner and outer cortical surfaces, respectively.

CLADA was also tested on actual low-resolution MRIs. Axial images with 5 mm thick slices were reformatted in the coronal plane and resampled to have isotropic voxels with significant partial volume averaging (Figure 4-13a). The effects of including both the self-intersection and the relaxation step (Figure 4-13c) can be appreciated over using the self-intersection test alone (Figure 4-13b, blue arrow).

Figure 4-13: Comparison of the resulting surface detected with and without the relaxation step included in the deformation. (a) resampled low-resolution MRI (b) deformed cortical surface detected with the inclusion of the self-intersection test but without surface relaxation (c) deformed cortical surface detected with the inclusion of the self-intersection test and surface relaxation step. The blue arrow indicates the error where the surfaces moved incorrectly due to a tunnel artifact.

Figure 4-14 shows CLADA results from a patient with Alzheimer’s disease taken from Open Access Series of Imaging Studies database (OASIS) (Marcus et al., 2007).

110

Significant atrophy is visible in the ventricles over the course of 1223 days (3.4 years).

Despite the significant blurring in the averaged image (upper left), CLADA was able to find ICS and OCS properly and successfully deformed both surfaces even in the presence of severe atrophy. The figure also shows that the intra-subject registration (NBCSR) performed well even in the presence of significant atrophy.

Figure 4-14: An example of longitudinal deformation in the subcortical region including the posterior lateral ventricle (v) of a demented subject. Despite blurring, CLADA was able to detect the cortical surfaces in the averaged image and to longitudinally follow both inner (yellow) and outer (red) edges even in the presence of severe atrophy.

4.3.3 Scan-Rescan Reproducibility

The reproducibility of CLADA’s global and regional CTh measurements was evaluated using scan-rescan MRI datasets. Reproducibility was determined for both high- resolution 3D MPRAGE and standard-resolution 2D T1SE images. Ten scan-rescan datasets of young controls from the OASIS database (Marcus et al., 2007) were used for the high-quality image analysis. Subjects were scanned twice within 10 days on a 1.5T scanner with the same MPRAGE sequence [repetition time (TR) = 9.7msec; echo time

111

(TE) = 4.0msec; flip angle (FA) = 10°; inversion time (TI) = 20msec; slice thickness

(THK) = 1.25mm; gap = 0mm; number of slices = 128; matrix size = 256×256; in-plane resolution (IPR) = 1mm×1mm; number of signal averages (NSA) = 4]. The lower resolution 2D T1SE images were acquired from 9 MS patients on a 1.5 T Siemens

Magnetom Vision with the following sequence parameters: TR=800 ms, TE=20 ms,

THK=5.0 mm; gap=0 mm; number of slices=30; matrix size=192×256, IPR=0.9 mm×0.9 mm. The MS patients were scanned twice, with one week between acquisitions.

CLADA was applied to analyze the images from both datasets and to measured global and regional CTh, and calculated absolute difference and percent error (= absolute difference / average) for each subject. The percent errors in global and regional CTh were plotted as boxplots.

CLADA performed better on MPRAGE than T1SE images. The mean (standard deviation) scan-rescan absolute error in global CTh was 0.015 mm (0.012) and 0.029 mm

(0.019) for MPRAGE image pairs and T1SE image pairs, respectively. The mean (s.d.) percent error in global CTh was 0.45% (0.38) and 0.77% (0.49) for MPRAGE and T1SE, respectively. Regional CTh measures had lower percent error for MPRAGE images in the temporal, parietal and occipital regions (Figure 4-15).

112

Figure 4-15: Boxplots of percent error in global and regional cortical thickness measurements from scan-rescan tests. Reproducibility was measured from datasets consisting of T1- weighted spin echo images of 9 MS patients (blue) and MPRAGE of 10 OASIS subjects (green).

For visualization of regional variability, surface error maps similar to Han and colleagues (Han et al., 2006) were generated for each subject. The intensity of each surface voxel was set to the difference in CTh from the nearest vertex and the resulting error image was smoothed (FWHM=0.9 mm). To create mean surface error maps, the images were registered to the MNI152 template using ANTS SyN nonlinear registration

(Avants et al., 2008). The error images were averaged across subjects in the MNI space, inverse transformed back to native space of one subject, and mapped to the cortical surface.

Scan–rescan CTh error maps are shown in Figure 4-16. Overall, MPRAGE images showed more homogeneous errors less than 0.3 mm, and lower-resolution T1SE images resulted in larger errors up to 0.7 mm. There were slightly larger errors in the inferior region of the brains, but in general the errors were homogeneous over the entire cortex, especially for MPRAGE.

113

Figure 4-16: The absolute cortical thickness difference was averaged in MNI space and mapped back to the cortical surface. The red surface indicates large errors and dark blue indicates no error (0mm difference in CTh). MPRAGE images (top) showed very low variability (less than 0.3 mm) whereas spin echo images (bottom) had differences of up to 0.7 mm.

4.3.4 Intra-operator Reproducibility Test

Intra-operator variability was assessed using a subset of these cases (five OASIS

MPRAGE and five MS T1SE). A single operator analyzed these scan-rescan cases twice using CLADA, approximately 4 days apart. Variability was calculated as the absolute difference in global and regional CTh and Pearson correlation coefficient of CTh change.

An intra-operator variability surface map was generated in the same manner as for scan– rescan variability.

The mean (s.d.) intra-operator absolute difference in global mean CTh from the

MPRAGE and T1SE image sets were 0.01 (0.02) mm and 0.01 (0.01) mm, respectively.

114

The mean (s.d.) absolute differences in CTh change were 0.0009 (0.0012) mm/yr and

0.001 (0.001) mm/yr for MPRAGE and T1SE, respectively. The repeated CTh change measurements were highly correlated in both MPRAGE (slope=1.07, r=0.997, p<0.0001) and T1SE (slope=0.999, r=0.988, p<0.0001). The intra-operator error was much smaller than the scan-rescan error, indicating that the effect of manual editing is small.

115

Figure 4-17: The intra-operator variability was measured from repeated analysis of longitudinal T1SE images from five MS patients (left column, a, c, and e) and longitudinal OASIS MPRAGE images from 5 demented or non-demented subjects (right b, d, and f). (a,b) Differences in absolute cortical thickness; (c,d) regional error map; (e,f) global and regional cortical thickness changes from repeated analyses. For both lower-resolution T1SE and higher-resolution MPRAGE images, the repeated measurements were highly correlated near the unity line (slope = 0.999, r = 0.988, p < 0.0001 for T1SE; slope = 1.07, r = 0.997, p <0.0001 for MPRAGE).

4.3.5 Accuracy using Post-mortem Images

To evaluate the accuracy of the CTh measurements, CLADA was used to measure

CTh from post-mortem MRIs acquired in situ and then compared to a “gold standard” thickness measured in 2D digital photographs of the same brain slices.

Our rapid autopsy protocol (Fisher et al., 2007) involved the following as described in Figure 4-18: (1) clinicians at the Mellen Center for Multiple Sclerosis

Treatment and Research recruited prospective patients, obtained consent, and entered the demographic information into the database. (2) Upon the death of an MS patient, appropriate personnel were called, the cadaver was transported to the Cleveland Clinic, and in situ MRI was obtained. The relevant sequence here was MPRAGE [TR=1900 ms,

TE=1.71 ms, THK=1.2 mm; number of slices=120; matrix size=256×256, IPR=1 mm].

(3) The brain was removed from the cadaver, and the left cerebral hemisphere was fixed in 4% paraformaldehyde for at least 4 weeks. (4) After the fixation was complete, the fixed brain was placed in a customized slicing box with MRI-sensitive markers and imaged again with the same MPRAGE sequence except with 1mm isotropic voxel size and 4 signal averages to accurately identify the marker locations. (5) Post-fixation

MPRAGE (Figure 4-18(5) left) with MRI-sensitive markers were acquired (yellow arrow on the contrast enhanced image (Figure 4-18 (5) right) –this post-fixation MRI was used

116

solely to locate the corresponding slicing plane in the post-mortem MRI, and thus no CTh measurements were made from post-fixation MRI. (6) Sliced brain tissue was photographed. (7) The left brain hemisphere was segmented in in situ MRI. (8) The in situ MPRAGE was registered to post-fixation MPRAGE using affine and nonlinear registration tools (FLIRT and FNIRT in FSL (Smith et al., 2004)) to account for brain shrinkage due to fixation. (9) Guided by the MRI-sensitive markers, the registered in situ

MPRAGE was re-oriented to the slicing plane. Each brain slice was visually inspected and corrected for misalignment.

Figure 4-18: A figurative flowchart of the rapid autopsy program at the Cleveland Clinic.

To compare CTh measurements obtained from CLADA using 3D MRI to CTh measurements from 2D photographs of the brain slices, four constraints were set to identify valid regions of interest (ROI): (1) Corresponding vertices from the inner and outer cortical surfaces had to be within 0.2 mm from the co-registered image/photograph

117

plane; (2) The absolute value of the dot product of unit normals from the brain slice plane and the cortical surfaces had to be less than 0.3 to ensure they were approximately perpendicular; (3) Large ROIs that spanned gyri and sulci had to be subdivided to minimize spatial variability as gyrus thickness tended to be larger than sulcal thickness

(2.9 mm and 2.2 mm respectively, (Fischl and Dale, 2000)); and (4) As determined by careful visual inspection, areas that appeared to have any misregistration were excluded.

ROIs were manually selected from all major (frontal, temporal, parietal, and occipital) lobes throughout the hemisphere. For each valid ROI determined in the MRIs based on the criteria above, the same ROI was manually selected in the photograph based on landmarks and the mean CTh was measured within the ROI. To measure CTh from the photographs, the WM–GM inner edge and GM–CSF outer edge were manually determined, and CTh was estimated using the Laplace method (Haidar and Soul, 2006).

Finally, the Pearson correlation and absolute differences between the photograph CTh and MRI CTh measurements were calculated.

Twenty-seven slices from four brains imaged post-mortem were used in the accuracy test. Out of a total of 113 ROIs initially selected for analysis, 10 ROIs were omitted from three slices due to misregistration. Thus, accuracy analysis was completed on the remaining 103 ROIs. An example of matching photograph and MRI is shown in

Figure 4-19.

118

Figure 4-19: An example of region of interest (arrows) where the CLADA-derived cortical thickness was compared with gold standard cortical thickness of the post-mortem brain slice.

CTh measurements from CLADA analysis of MRIs were linearly correlated to

CTh measurements from photographs of the same tissue slices (r=0.68, p<0.001). The slope of the regression line (1.06) was slightly greater than unity (Figure 4-20). The mean

(s.d.) difference between MRI and photograph CTh measurements was 0.17 (0.54) mm and the mean (s.d.) absolute difference was 0.43 (0.36) mm. Detailed inspections on the large errors revealed that CLADA included dura mater and resulted in increased CTh from CLADA compared to CTh from sliced brain. The mean error was fairly small, considering the voxel size of 1 mm in these MRIs.

119

Figure 4-20: The cortical thickness measurements from MRI and photographs correlated with slope = 1.06, R = 0.68, and p < 0.001. The best fit line is shown in black, and dotted red line is the ideal case.

4.3.6 Sensitivity using Simulation of Atrophy

The sensitivity of CLADA for detection of small changes was tested using simulations of regional brain atrophy or growth. ROIs were manually selected from MRIs of five OASIS datasets. The ROIs on the brain surface were very slightly deformed by a known amount at sub-voxel level, either inward for atrophy or outward for growth. The direction of deformation was normal to the brain surface mask and determined by 1.6 mm-FWHM Gaussian and sobel directional filters. The magnitude of the deformation was normalized to maximum of 0.3 mm at the brain edge after the 1.6 mm-FWHM

Gaussian filter. CLADA was applied to the original and deformed images to measure changes in CTh. The error was calculated as the mean absolute difference between the measured results and the actual applied deformations. CLADA's sensitivity was calculated by TP/(TP+FN) where TP is the number of surface points CLADA correctly

120

deformed more than 0.1 mm and FN is the number of surface points CLADA did not deform even though the image was deformed more than 0.1 mm. A plot of the actual versus measured deformations was plotted, and the slope of best-fit line was calculated.

For the sensitivity test, a total of 8172 surface points were collected from 20 ROIs in five OASIS datasets (Figure 4-21). The maximum applied deformation was 0.3 mm.

The sensitivity or the true positive rate with a cut-off of 0.1 mm detection was 86%, and the slope of the best-fit line was 0.75. Measured changes were strongly correlated to applied simulated changes (r=0.958, p<0.001, Figure 4-21). The mean absolute error

(s.d.) between measured and applied changes in CTh was 0.02 (0.02) mm. This result showed that CLADA was sensitive to very small, sub-voxel change.

Figure 4-21: An example of the original (a), deformed (b), and difference (c) images, and (d) The results of the sensitivity test. The black solid line is the best fit line of the measured thickness change and the applied deformation (slope = 0.75, Pearson correlation coefficient = 0.958, p<0.001). The red dotted line is the ideal line.

121

4.3.7 Comparison of Image Sequences

Next, CLADA results from conventional 2D T1SE images and from high-contrast

3D MPRAGE images were compared. The slice thickness for 2D T1SE was 3mm and

MPRAGE was 1.2mm. Longitudinal MRI from 15 secondary progressive multiple sclerosis (SPMS) patients included axial T1SE images (TR=675 ms; TE=17 ms; THK=3 mm; number of slices=48; matrix size=256×192; in-plane resolution= 0.9mm×0.9 mm;

NSA=1) acquired on a 1.5 T Siemens Symphony scanner and axial MPRAGE images

[TR=1900 ms; TE=1.71 ms; FA=8°; TI=900 ms; THK=1.2 mm; number of slices=128; matrix size= 256×256; IPR=1×1mm; NSA=1] acquired on the same day on a 3 T

Siemens Magnetom Trio scanner. Images were acquired at baseline and weeks 2, 4, 26,

52, and 78. The T1SE and MPRAGE images were analyzed separately. The annualized global CTh change was calculated as the difference between CTh measurements at baseline and last follow-up divided by the time interval (78 weeks or 1.5 years). The

Pearson correlation was determined between annualized global CTh change measurements obtained from T1SE and MPRAGE images.

CTh measurements differed for T1SE and MPRAGE input images (Figure 4-22).

The mean global CTh at baseline was 2.71±0.21 for MPRAGE and 3.57±0.23 mm for

T1SE (p<0.0001). The mean annualized changes in CTh measured from MPRAGE and

T1SE images were −0.031±0.02 mm and −0.024±0.02 mm, respectively. The correlation was 0.569 (p=0.027). The percent change in CTh for MPRAGE and T1SE was −1.14% for MPRAGE and −0.69% for T1SE. Although the measured rates of cortical atrophy was smaller with T1SE datasets compared to MPRAGE datasets, CLADA was able to

122

detected atrophy in all patients from both sequences. The results also showed that CTh significantly differed when different sequences were used.

Figure 4-22: Scatter plot of annualized cortical thickness change measured from T1SE and MPRAGE images in 15 MS patients. The slope was 0.423 (r=0.569, p=0.027).

4.3.8 Effect of Reference Image

Registration-related variability was investigated by changing the reference time- point from baseline to follow-up. Eight longitudinal OASIS datasets (each with 2 time- points) were analyzed (Marcus et al., 2007). CLADA was run twice for each set: once with the baseline image used as the reference space, and then again with the follow-up image used as the reference. The absolute difference was calculated for the global and regional CTh measurements, and a surface error map was generated to compare the annualized CTh changes measured each way.

123

The results from repeated CLADA analysis with either baseline or follow-up image as the reference space showed a mean absolute difference in mean global CTh of

0.013±0.005 mm. The regional CTh absolute differences (s.d.) were 0.015 (0.016) in frontal, 0.023 (0.021) in temporal, 0.014 (0.013) in parietal, and 0.011 (0.012) mm in occipital lobes. The boxplot and the surface error map are shown in Figure 4-23.

The mean error due to the choice of reference was 0.013mm, which was comparable to the scan-rescan mean absolute difference of 0.015mm. Therefore, the effect of reference image was very small.

124

Figure 4-23: Effect of different reference spaces: (a) boxplot of absolute global and regional cortical thickness differences; (b) plot of change in cortical thickness using baseline as the reference space and using the follow-up image as the reference space (slope = 1.00, r = 0.997, p < 0.0001); (c) surface error map indicating regions where cortical thickness differed.

4.3.9 Comparison with FreeSurfer

CLADA was compared to FreeSurfer since it was widely used, freely available and extensively validated. CLADA and FreeSurfer were applied on a subset of dataset

(seven of the SPMS patient) described in the “Sequence Comparison” section. FreeSurfer version 4.5.0 was used on the baseline MPRAGE images. FreeSurfer CTh was measured within the same regions as CLADA to enable direct comparison of the global and regional CTh results. Regional labels (global, frontal, temporal, parietal and occipital

ROIs) created from CLADA were used in FreeSurfer as described in FreeSurfer ROI analysis workflow4. For FreeSurfer analysis, semi-automatically segmented MS lesions were filled with a mean normal-appearing WM intensity estimated during segmentation.

FreeSurfer was applied to the lesion-filled MRIs without manual intervention. To reduce analysis time, this step was performed in place of correcting the lesions after cortical reconstruction. The results were visually inspected for errors related to the lesions. A detailed CTh difference map was generated in MNI152 space by nonlinear registration to visualize regional differences in CTh measured by CLADA and FreeSurfer.

CLADA CTh measurements were correlated with FreeSurfer CTh measurements, both globally (r=0.73, p=0.04) and regionally (Figure 4-24). The global mean absolute difference between CLADA and FreeSurfer CTh was 0.37 mm. Regional CTh differences were slightly higher in temporal lobes (0.42 mm) than in other lobes (frontal=

4 http://surfer.nmr.mgh.harvard.edu/fswiki/VolumeRoiCorticalThickness

125

0.37 mm, parietal=0.35 mm, and occipital=0.31 mm). The detailed surface error map showed that differences were generally less than 0.5 mm, but the temporal and frontal lobes showed larger differences. Overall, CLADA CTh measurements were higher than

FreeSurfer CTh measurements. In this cross-validation test, accuracy could not determined, but the high correlation and relatively small difference showed that CLADA was in agreement with a previously validated method.

Figure 4-24: Cortical thickness measured using CLADA and FreeSurfer in seven SPMS patients (a). The cortical thickness measurements were correlated (r=0.70, p<0.04) and

126

consistently higher than FreeSurfer. The surface error map shows averaged absolute differences in cortical thickness which were slightly higher in temporal lobe (b, red surface).

4.3.10 Bias in Estimating Atrophy

To further study the bias in estimating atrophy, the method described by

(Yushkevich et al., 2010) was applied to the data from the Reproducibility Test (Section

4.3.3). This approach analyzed scan-rescan data as if the data were from an actual longitudinal study where for each subject, the repeated scans are randomly assigned to be either “baseline” or “follow-up.” Changes in global and regional CTh were statistically compared to truth (no change) using t-tests.

The estimation of bias using the scan-rescan datasets was not significantly different from zero. For the high-resolution OASIS images, the p-values were 0.657,

0.673, 0.354, 0.684, and 0.672 for global, frontal, temporal, parietal, and occipital CTh, and for the low-resolution scan-rescan images the p-values were 0.383, 0.112, 0.983,

0.758, and 0.633 for global, frontal, temporal, parietal, and occipital CTh, respectively.

The results showed that the bias in estimating atrophy by CLADA was not statistically significant.

4.3.11 Comparison to a Cross-sectional Method

Next, the full version of CLADA was compared to a cross-sectional version of

CLADA where images from individual time-points were analyzed independent of each other (i.e., no intra-subject registration and no averaging) in terms of variability errors, similar to the study by Han and colleagues (Han et al., 2006). In our study, global and

127

regional (frontal, parietal, temporal, and occipital) CTh were measured from 3 scan- rescan studies: (1) high-resolution, high-contrast (4 signal average) MPRAGE from

OASIS database, as in “Scan-rescan Reproducibility (4.3.3),” (2) high-resolution

MPRAGE (NSA = 1) from “Comparison of Image Sequence (4.3.7),” acquired within a month (baseline, week 2, and week 4), (3) low-resolution T1SE from “Scan-rescan

Reproducibility (4.3.3),” also acquired within a month (baseline, week 2, and week 4).

The variability error was defined as mean absolute deviation (Σ|CThn−μCTh|/N where n=1,2,⋅⋅⋅N, N was the number of images, and μCTh was the average CTh from different sessions) and the percent error was calculated as [mean absolute deviation] / μCTh × 100%.

The t-tests were used to compare the percent error between CLADA and cross-sectional method.

When the variability error in global CTh from dataset 1 (high-resolution, high- contrast MRI) was compared between cross-sectional version of CLADA and full

CLADA, there was no significant improvement (0.25% vs. 0.27%; p = 0.84; difference of

0.02% in Table 4-1: Column = Global and Row = Dataset 1; and Figure 4-25: green circles). However, CLADA out-performed when lower quality MRIs are used (Table 4-1).

Regionally, variability errors from full CLADA were consistently smaller than those from the cross-sectional method.

These results indicated that the longitudinal CLADA benefited from its longitudinal processing and that this benefit would be greater for MRIs with lower signal- to-noise ratio or lower resolution.

Table 4-1: Results of comparison between longitudinal and cross-sectional methods. The difference in the variability error is tabulated here; thus larger numbers indicate more

128

improvements for the longitudinal method. Note that there is no negative number indicating that longitudinal method consistently out-performed the cross-sectional method. Global Frontal Parietal Temporal Occipital Data 1 0.02% 0.13% 0.39% 0.15% 0.49% Data 2 0.80% 0.88% 1.02% 0.80% 1.49% Date 3 1.85% 2.18% 2.13% 1.80% 3.18%

Figure 4-25: The scan-rescan variability errors from the full longitudinal CLADA and cross- sectional version of CLADA. Points below the black solid line represent data where full longitudinal CLADA was better. Green circle = high-resolution, high-contrast MPRAGE (dataset 1). Blue triangle = high-resolution MPRAGE (dataset 2). Red square: low-resolution T1SE (dataset 3).

4.3.12 Effects of MS Lesions

Like the systematic effect of MS lesions (T2LV) on GMF (Section 2.3.5), the effect of MS lesions on CTh was investigated. Nine MS patients who had varying levels of T2LV were selected. The enlarging lesions were simulated by dilating the T2 lesion

129

masks by 1mm-, 2mm-, and 3mm-spherical kernels. CTh was measured using CLADA with original or dilated lesion masks.

The variability in baseline CTh was measured by the coefficient of variations. The effect of MS lesions on cortical thinning was determined by plotting the mean absolute difference in annualized rate of cortical thinning. The coefficient of variation was not measured for cortical thinning because the atrophy rate could be zero where the coefficient of variation became undefined.

The effect of MS lesions on the CTh measurements showed the average coefficient of variation (s.d.) of 0.29% (0.21%). The linear regression of CTh and T2LV showed a slope of 0.0007 mm/ml. The effect on the annualized rate of cortical thinning showed that the variation was very small (mean s.d. = 0.003; Figure 4-26). Therefore, there was no significant effect of lesions on the measurement of CTh and change in CTh.

130

Figure 4-26: The effect of MS lesions on the annualized rates of cortical thinning. This plot shows that the variation due to the enlarging MS lesions was very small: averaged variance was 7.50e-7. An increase in T2LV is indicated by lighter plot colors.

4.4 Discussion

This chapter described an efficient new method for measurement of cortical atrophy in MRIs acquired for longitudinal studies. Existing software packages for CTh measurements had been intended for analysis of high-contrast images with approximately

1 mm3 voxels. A unique feature of CLADA was that it was designed specifically to measure CTh change from lower-resolution 2D images with 3-5 mm slice thickness. The deformable model incorporated relaxation on self-intersection regions to account for WM segmentation errors, which were common in lower-resolution images with significant partial volume effects. Our tests showed that the algorithm was still successful in detecting cortical surfaces and estimating CTh change in 2D images with 3-5 mm slice thickness. However, as expected, the errors from lower-resolution images were higher than those obtained with high-resolution 3D MPRAGE images. These results suggested that although studies using older, lower-resolution MRIs might require larger sample sizes or longer duration, CLADA was applicable for analysis of retrospective datasets, such as those typically acquired for MS clinical trials.

CLADA used explicit rather than implicit surfaces for the cortical model because explicit representations seemed better suited to model highly convoluted surfaces (Zeng et al., 1999) with smoothness constraints (Acosta et al., 2009). An important feature of our deformable cortical surface detection algorithm was the incorporation of remeshing and self-intersection tests. These components led to an efficient multi-resolution shrink-

131

wrap algorithm that maintained surface topology. Topological inconsistency is a common problem in cortical surface detection algorithms initialized from a WM mask (Fischl et al.,

2001; Han et al., 2006). The multi-resolution remeshing approach achieved highly complex shapes while still maintaining topology. These functions were important because deformable models require good initialization to avoid local minima. The savings in computation time were substantial compared to equivalent methods using explicit surfaces. Measurement of CTh with CLADA takes approximately 15 hours for a single subject with six time-points, whereas similar methods, such as FreeSurfer or ASP, require approximately 20-30 hours for a single image at a single time-point (Dale et al.,

1999; MacDonald et al., 2000).

The strength of CLADA was the longitudinal nature of the algorithm, as demonstrated by the high reproducibility. CLADA used images from all time-points to create an average image with high SNR and to estimate the cortical surface by creating a subject-specific cortical surface model. This model was then deformed to fit the patient’s images from individual time-points. User interaction was kept to a minimum –only the mask for the subject’s average image was manually edited, if necessary. Thus, CLADA took less time to analyze follow-up images once the initial model has been created.

4.4.1 Validation Tests

Various validation tests were performed to quantitatively evaluate our new algorithm in terms of reproducibility, accuracy, sensitivity, performance with different types of MRIs, effect of reference space, comparison with FreeSurfer, bias in estimating atrophy, comparison with a cross-sectional method, and effect of enlarging MS lesions.

132

Reproducibility

The reproducibility test results demonstrated a relatively low scan-rescan error (<

1%) in all regional CTh measurements. This was expected given that CLADA is a longitudinal algorithm, which incorporates images over all time-points. The intra- operator variability error was in the order of scan-rescan reproducibility, suggesting that the effect of manual editing is very small. The intra-operator variability in longitudinal change was consistent and small (up to 0.007 mm/yr in temporal lobe).

Accuracy test

Accuracy validation through comparison with post-mortem brain tissue is a very difficult task (Das et al., 2009). The majority of earlier reports on CTh measurements from autopsied brains did not account for three dimensions (Mann, 1991; Wegner et al.,

2006; Wiley et al., 1991). Rosas et al. (2002) used autopsied brain tissue to validate the accuracy of MRI CTh measurements, but restricted measurements to the crowns of gyri where the curvature is minimal. To compare 3D MRI-based CTh measurements to 2D tissue-based CTh, careful co-registration of image slices to tissue slices and selection of valid ROIs from MRI were used to ensure that the 3D thickness vectors were in the plane of the corresponding tissue slices. Otherwise the measurements would not be expected to be comparable. These steps allowed the comparison of measurements even in curved regions and deep sulci.

The results demonstrated that CTh as measured by CLADA was correlated to

CTh measured in fixed tissue at the same locations. The slope indicated that MRI-based

133

measurements were slightly higher than those in the fixed tissue, likely due to fixation effects. The absolute error of 0.43 mm was fairly low, considering the voxel size of 1.2 mm3 and complex effects of fixation. The CLADA measurements were higher than photographs only by 0.17 ± 0.54mm on average (median = 0.09mm, mode = 0.0mm with bin size = 0.1mm). While the positive mean CTh offset appeared to show that the

CLADA CTh might be consistently higher, the fact that the median offset was 0.09mm with a 0.0mm-mode relative to the photograph-derived CTh showed that a few data points with large errors might have skewed the means.

Indeed several points had large positive errors due to the inclusion of dura mater: four ROIs with the largest measurements errors included dura mater in CTh measured by

CLADA. The dura mater is difficult to remove with a single MPRAGE image (van der

Kouwe et al., 2008), and additional imaging modalities may be helpful in future applications of CLADA.

Other complex factors might have affected the results of this validation test. First, there was an attempt to minimize fixation effects by using affine registration but it might not be entirely removed because CLADA-based CTh was larger. In addition to tissue shrinkage, the fixation effect becomes nonlinear in the focal abnormal areas, as more damaged tissues tended to deform more during the fixation. Furthermore, the fixed brain hemispheres tended to be flattened as the midsaggital plane was placed at the bottom of the container of fixative. Second, two different types of CTh measurements were compared: the distance between corresponding points with CLADA in MRI, and the

Laplace thickness in photographs. CLADA thickness was defined by straight lines while the Laplace thickness might curve, which may have artificially decreased the error at the

134

curved areas (Figure 4-1) (Jones et al., 2000). There are various ways of defining the corresponding points between the inner and outer cortical surfaces for calculation of thickness as discussed in Section 4.1.1. The Laplace method was utilized here because it had unique surface correspondence for CTh measurements even in regions of high curvature. Another important limitation is the accuracy of the gold standard. While CTh was carefully and accurately measured using high-resolution photographs (pixel size <

0.1 mm), factors like fixation and unclear tissue edges can affect the accuracy of the gold standard.

Sensitivity Test

The sensitivity test showed that CLADA was able to detect sub-voxel level change. The detected change tended to be slightly less than the applied deformation partly due to the surface and CTh smoothness terms incorporated into CLADA. There was also a trend for larger errors with greater deformation. This observation may be related to an artifact of the applied deformation, which may be misaligned with the cortical surface normal, and the absolute errors increase proportionally with the deformation field. As demonstrated by the comparison of measured to applied deformations, CLADA appeared to have lower sensitivity for positive change (increases in CTh) than for negative change (cortical thinning), with slopes of 0.67 and 0.78, respectively. This difference was primarily due to the limited range of motion for applied deformations in the positive direction in cortical regions with small sulci, which are common in these young normal controls. Measurements of positive deformations selectively applied to gyri surrounded by larger sulci confirmed this explanation. In these

135

restricted post hoc analyses, the positive deformations had similar sensitivity as negative deformations (slope = 0.82). Finally, part of the deviation from ideal slope (1.0) may be attributed to bias introduced by interpolation. In this test, the warped image was interpolated but the original image was not, and biased interpolation has been shown to lower sensitivity (Yushkevich et al., 2010). Our simulation method avoided deep sulci and used crowns of gyri to properly simulate atrophy or growth with simple image warping. Rather than to estimate accuracy, the main goal of these simulations was to determine if CLADA was sufficiently sensitive to detect very small changes in CTh. If necessary, more sophisticated methods for simulating global or regional atrophy using finite element or topology preserving models (Camara et al., 2006; Karacali and

Davatzikos, 2006) could be implemented for future studies.

Comparison of Image Sequences

The image sequence comparison showed that CTh was (on average) 0.8 mm larger when measured in MRIs with 1×1×3 mm3 voxels as compared to MRIs with ~1 mm3 voxels. Most likely, blurred tissue boundaries due to partial volume effects led to apparently thicker cortex. This comparison confirmed the importance of scanning subjects with the same sequence in longitudinal studies.

There was surprisingly good correlation between CLADA measurements from

MPRAGE images and the lower-resolution T1SE images acquired in the same session, with the exception of one outlier in the patient group. The outlier in Figure 4-22 appeared to be related to significant motion artifacts evident in the T1SE images at both time- points. The difference in the scanner field strength may have affected these results to

136

some extent. However, the objective of this test was to compare CLADA results from lower-resolution MRIs to that of higher-resolution, higher-contrast images in a longitudinal setting. These data support the use of CLADA to analyze retrospective MRI datasets.

Comparison with FreeSurfer

CLADA also showed good agreement with FreeSurfer, the most widely used software package for estimation of CTh. These comparisons showed that CLADA measurements are consistently higher than FreeSurfer by 0.36±0.14mm, on average

(median = 0.34 mm, mode = 0.3 mm with bin size = 0.1 mm). This test is a cross- validation analysis to demonstrate comparability with a widely used method rather than a comparison to a gold standard. One possible interpretation is that the FreeSurfer measurements are accurate and CLADA has a consistent sub-voxel offset of approximately 0.3 mm. However, it may also be the case that neither method is completely accurate. Correlational agreement is informative in cross-validation comparisons where the absolute truth is not known (Smith et al., 2007).

Bias in Estimating Atrophy

The bias in estimating atrophy was evaluated in several ways. First the reference image was changed to investigate the registration-related bias. The choice of reference image did have a small effect on CTh measurements, despite our efforts such as image averaging with symmetric linear registration, cortical model construction from the averaged template, and transformation of cortical model rather than the image. The bias

137

estimation using scan-rescan in contrast to the clinical datasets was small, indicating that

CLADA can be reliably applied to longitudinal CTh studies. Thus, construction of an unbiased template remains a difficult task (Guimond et al., 2000; Joshi et al., 2004;

Reuter and Fischl, 2011). Comparison of CTh measurements using either baseline or the follow-up image as the reference space resulted in small differences, which may be the source of scan-rescan error. The longitudinal change analysis was less affected by the choice of reference space. The differences in CTh change were on the order of 0.001 mm/yr, which was smaller than the mean atrophy rates of −0.05 mm/yr in controls (after converting from percent change) (Chen et al., 2004), −0.0036 mm/yr in MS, −0.0021 mm/yr in controls (Sampat et al., 2010) in longitudinal studies; in cross-sectional studies, the estimated rates of cortical atrophy were −0.0016 mm/yr in controls (Salat et al., 2004),

−0.0086 mm/yr in controls (Hutton et al., 2009), and −0.01 mm/yr in RRMS (Charil et al.,

2007).

4.4.2 Conclusion

Overall, the results show that CLADA can be used for reliable quantitative measurement of CTh change in longitudinal MRI studies.

In MS, histopathological studies show that the cortex is demyelinated (Kidd et al.,

1999) without evidence of inflammation (Peterson et al., 2001). Therefore, cortical atrophy most likely reflects severe cortical damage and neuronal loss rather than effects related to the resolution of edema or pseudoatrophy (Nakamura et al., 2010; Zivadinov et al., 2007).

138

There is also increasing MRI evidence of cortical damage. Cross-sectional and short-term longitudinal studies have found that MRI-detected CTh change is related to patient disability (Calabrese et al., 2007b; Charil et al., 2007; Chen et al., 2004) and cognitive impairment (Amato et al., 2004; De Stefano et al., 2003). The overall clinical relevance and predictive value of CTh changes for determination of disability progression over the longer term (>1 year) are yet to be determined, and the next chapter describes 4- year longitudinal study to address the clinical relevance of cortical thinning in MS.

139

CHAPTER 5 CORTICAL ATROPHY IN MS

This chapter describes a longitudinal clinical study of cortical atrophy in MS. The newly developed CTh measurement algorithm was applied to the same longitudinal data as in Chapter 3 to study the evolution of cortical thinning in CIS, RRMS, SPMS, and NC.

Clinical correlations and predictors of cortical thinning were also investigated.

This chapter has been prepared for a manuscript submission to Annals of

Neurology.

5.1 Materials and Methods

The subjects, MRI, and image analysis are described in Chapter 3. An important difference here is that CLADA was applied to the full longitudinal data. Thus, for each subject, CTh was measured at nine time-points for MS subjects and five time-points for

NC. The annualized change in global cortical thickness (∆CTh) was estimated by the slope of the regression line between CTh and the interval time from baseline.

Measurements for baseline CTh and ∆CTh were compared between MS and NC as well as between subgroups (CIS, RRMS, and SPMS) using analysis of covariance

(ANCOVA). Baseline CTh and ∆CTh were also compared between the stable versus worsening patients using ANCOVA with age as a covariate.

140

We calculated Spearman’s rank correlation coefficients for the correlational analysis for CTh and ∆CTh with imaging variables (BPF, GMF, WMF, T2LV, T1LV,

GdLV, and MTRNABT) and clinical measures (EDSS, MSFC and three components of

MSFC). We then used multiple regression analysis to determine the independent estimators of baseline CTh and ∆CTh among the imaging variables (T2LV, T1LV, GdLV,

MTRNABT, MTRNAWM, and MTRLesion).

The sample size was estimated for 1-year proof-of-concept study to detect a 60% effect size with 80% power at a 5% significance level. UnifyPow, an SAS module for sample-size analysis was used. The results of measurements from RRMS, SPMS, and combined CIS+RRMS+SPMS were used to calculate the sample size.

5.2 Results

5.2.1 Baseline Characteristics

The baseline characteristics are shown in Table 5-1, and CTh difference map is shown in Figure 5-1: (a) RRMS minus NC, (b) SPMS minus NC, and (c) SPMS minus

RRMS. CTh was significantly lower in the MS group compared to NC (4.19 ± 0.38 mm versus 4.52 ± 0.28 mm, p < 0.001), and there was a trend for lower baseline CTh in patients with more advanced disease, where CTh was 4.44 ± 0.22 mm in CIS, 4.21 ± 0.35 mm in RRMS, and 3.96 ± 0.42 mm in SPMS subjects (Figure 5-2).

141

Figure 5-1: Baseline cortical thickness difference maps (a) RRMS minus controls (NC), (b) SPMS minus NC, and (c) SPMS minus RRMS. Blue indicates smaller thickness, and red show growth.

Table 5-1: Baseline characteristics. NC CIS RRMS SPSM (n=17) (n=7) (n=36) (n=27) 44.2 Mean (SD) Age, years 41.6 (8.1) 39.8 (8.4) 45.7 (7.8) (10.9) Mean (SD) disease duration, NA 0.25 (0.1) 5.7 (5.4) 17.0 (7.6) years Mean (SD) EDSS NA 1.0 (0.76) 1.9 (1.5) 5.1 (1.5) 0.55 0.36 0.43 −0.82 Mean (SD) MSFC (0.27) (0.35) (0.58) (1.34) 18.1 42.7 Mean (SD) T2 lesion volume, ml NA 2.5 (2.3) (17.2) (24.9) 0.15 1.72 Mean (SD) T1 lesion volume, ml NA 8.3 (8.31) (0.22) (2.89) 0.862 0.862 0.843 0.804 Mean (SD) BPF (0.012) (0.008) (0.026) (0.04) 0.554 0.553 0.541 0.520 Mean (SD) GMF (0.015) (0.010) (0.019) (0.028) 0.308 0.309 0.302 0.283 Mean (SD) WMF (0.011) (0.009) (0.017) (0.016) Abbreviations: NC = healthy normal control, CIS = clinically isolated syndrome, RRMS = relapsing-remitting MS, SPMS = secondary progressive MS, SD = standard deviation, EDSS = extended disability status scale, MSFC = MS functional composite, CTh = cortical thickness.

142

Figure 5-2: Boxplots showing the results of group comparisons of baseline cortical thickness. Boxes indicate the quartiles, vertical lines are the extreme values, and the thick bars are the medians. There was a trend for lower cortical thickness for more severe disease subtype: healthy normal controls (NC, 4.52 ± 0.28mm), clinically isolated syndrome (CIS, 4.44 ± 0.22 mm), relapsing-remitting (RR, 4.21 ± 0.35mm), and secondary progressive (SP, 3.96 ± 0.42mm) MS.

5.2.2 Cortical Thinning

The annualized changes (ΔCTh) were −0.010 ± 0.018, −0.013 ± 0.012, −0.018 ±

0.016, and −0.025 ± 0.022 mm/yr for NC, CIS, RRMS, and SPMS, respectively (Figure

5-3a). ΔCTh in SPMS subjects was significantly different from NC (p = 0.008) but other

differences did not reach statistical significance. The longitudinal evolution of global

cortical thinning for each subgroup is shown in Figure 5-3b.

143

Figure 5-3: (a) Boxplots showing the annualized change in global cortical thickness (∆CTh) for each compared group (NC: 0.010 ± 0.018mm/year, CIS: 0.013 ± 0.012mm/year, RRMS: 0.018 ± 0.016mm/year, SPMS: 0.025 ± 0.022mm/year). (b) The same data is plotted as a function of time (error bars are standard errors). Missing data is estimated by linear regression.

5.2.3 Clinical Correlation of Cortical Thickness

At baseline, there were moderate correlations between CTh and EDSS as well as between CTh and MSFC (Table 5-2). All components of MSFC were correlated with

CTh and the strongest correlations were observed with 9HPT.

144

Table 5-2: Spearman rank correlation coefficients Baseline Cortical Cortical Thickness Thickness Change Correlations with Demographic and Disease Variables Age −0.46 (p<0.0001) −0.10 (p=0.43) MS Duration −0.60 (p<0.0001) −0.21 (p=0.08) EDSS −0.58 (p<0.0001) −0.20 (p=0.1) MSFC 0.54 (p<0.0001) 0.20 (p=0.1) 25’ walk −0.55 (p<0.0001) −0.11 (p=0.37) 9-hole peg test 0.62 (p<0001) 0.17 (p=0.16) PASAT 0.30 (p=0.01) 0.12 (p=0.33) Correlations with MRI Measures: BPF 0.68 (p<0.0001) 0.22 (p=0.07) GMF 0.55 (p<0.0001) 0.12 (p=0.33) WMF 0.59 (p<0.0001) 0.22 (p=0.07) T2LV −0.62 (p<0.0001) −0.38 (p=0.001) T1LV −0.60 (p<0.0001) −0.28 (p=0.02) GdLV 0.15 (p=0.2) −0.34 (p=0.004) NABT MTR 0.46 (p=0.0001) 0.29 (p=0.02) Abbreviations: EDSS = Kurtzke Expanded Disability Status Scale; MSFC = MS Functional Composite; PASAT = Paced Auditory Serial Addition Test; BPF = Brain Parenchymal Fraction; GMF = Gray Matter Fraction; WMF = White Matter Fraction; T2LV = T2 Lesion Volume; T1LV = T1 Lesion Volume; GdLV = Gadolinium Lesion Volume; NABT = Normal-Appearing Brain Tissue; MTR = Magnetization Transfer Ratio

The baseline CTh correlated with other concurrent atrophy measures (BPF, GMF and WMF). Baseline CTh also correlated with baseline T1LV and T2LV but not with

GdLV. The correlations between baseline MTR and baseline CTh were also moderate.

∆CTh was related to whole brain and GM atrophy rates (r = 0.48, p < 0.0001 and r = 0.38,

p = 0.001, respectively) and cumulative GdLV (r = 0.36, p = 0.003).

5.2.4 Disability Progression vs. Thickness and Change

When patients were classified into EDSS-progressed and stable groups, there were trends for lower CTh and more atrophy in patients who progressed. For baseline

CTh, the means were 4.17 ± 0.35 mm vs. 4.20 ± 0.41 mm (p=0.78), and ∆CTh means were −0.023 ± 0.021 mm/yr vs. −0.017 ± 0.014 mm/yr (p=0.18) for progressed and stable

145

groups, respectively. However, when progression was defined by MSFC, compared to stable patients, the patients who progressed had significantly lower baseline CTh (4.07 ±

0.41 mm vs. 4.29 ± 0.33, p=0.02) and more ∆CTh (−0.027 ± 0.019 mm/year vs. −0.013 ±

0.013 mm/year, p=0.001). The baseline characteristics for these subgroups are shown in

Table 5-3.

Table 5-3: Baseline patient characteristics based on MSFC progression. MSFC Progressed MSFC Stable P-value (n=31) (n=39) Mean (SD) age, years 44.5 (9.7) 41.8 (8.7) 0.23 Mean (SD) disease <0.001 14.1 (9.4) 5.9 (6.0) duration, years Mean (SD) EDSS 4.2 (2.1) 2.2 (1.9) <0.001 Mean (SD) MSFC −0.47 (1.3) 0.26 (0.8) 0.009 Mean (SD) T2 lesion 0.014 34.0 (25.0) 19.7 (21.6) volume, ml Mean (SD) T1 lesion 0.10 5.7 (7.9) 2.9 (5.0) volume, ml Mean (SD) BPF 0.816 (0.042) 0.841 (0.029) 0.008 Mean (SD) GMF 0.528 (0.030) 0.539 (0.019) 0.083 Mean (SD) WMF 0.288 (0.017) 0.302 (0.018) 0.001 For abbreviations, see Table 5-1. * p < 0.05, ** p <0.01

5.2.5 Predictors of Cortical Thickness and Thinning

Multivariate analysis demonstrated that age (standard regression coefficient, β =

0.31, p = 0.001), baseline T2LV (β = 0.49, p < 0.001), and baseline MTRNABT (β = 0.23, p = 0.03) were independent correlates of baseline CTh. The overall adjusted R2 for this model was 0.57. Predictors for ∆CTh were baseline T2LV (β = 0.36, p = 0.002) and cumulative GdLV (β = 0.32, p = 0.006) with an overall model R2 was 0.24.

146

5.2.6 Sample Size Estimate

Sample size estimates for a one-year study are shown in Table 5-4. For a trial in

SPMS, 92 patients per arm would be required to detect a 60% effect size with 80% power at a 5% significance level. For a trial in RRMS, 233 patients per arm would be required.

Table 5-4: Number of subjects per arm needed to detect a 60% reduction in ΔCTh, 5% significance level 80% Power 90% Power RRMS+SPMS+CIS 166 221 RRMS only 233 312 SPMS only 92 100

5.3 Discussion

This chapter characterized cortical thinning in patients with CIS, RRMS, and

SPMS. There are several important findings from this study. First, our baseline results confirmed previous cross-sectional studies that the cortex was thinner at baseline in MS patients compared to NC (Sailer et al., 2003) and thinner in more advanced patients

(Calabrese et al., 2007b). The baseline CTh correlated with lesion load as well as with all measures of disability (EDSS, MSFC, 25FTW, 9HPT, and PASAT), further emphasizing that CTh is a clinically relevant marker of MS disease progression. Surprisingly, the correlation between MSFC and CTh was driven more strongly by 25FTW and 9HPT and least by PASAT. In previous studies, correlations between PASAT and measures of global GM atrophy were higher than for cortical atrophy (Fisher et al., 2008),(Amato et al., 2004; Benedict et al., 2006; Morgen et al., 2006). Benedict et al. reported the variance in PASAT was best described by normalized GM volume and to a lesser degree by

147

normalized cortical GM volume (Benedict et al., 2006), thus, the PASAT may be more closely associated with tissue loss in deep GM structures. Second, the longitudinal analysis suggested there was greater cortical thinning in MS patients as compared to age- matched NC and greater cortical atrophy as MS stage advanced. This finding was not observed in an earlier report on longitudinal CTh in MS (Chen et al., 2004), possibly due to the smaller sample size and shorter duration in that study. The result is in concordance with our previous report on longitudinal GM atrophy, where GMF change was found to be 8-fold higher in RRMS and 14-fold higher in SPMS compared to NC (Chapter 3)

(Fisher et al., 2008). Third, when we examined the MRI predictors of CTh and cortical thinning (ΔCTh), we found that both focal and diffuse damage (T2LV and MTRNABT) were independently associated with baseline CTh, accounting for 57% of the variance.

However, only focal lesion measures (T2LV and cumulative GdLV) predicted subsequent cortical thinning, accounting for only 24% of the variance. Finally, we determined that measurement of cortical thinning would be feasible as an outcome measure in MS clinical trials, particularly for proof-of-concept studies in SPMS.

The predictors of CTh and cortical thinning suggested that cortical atrophy might be the end result of a variety of brain insults: physiologic (age), focal (T2LV) and diffuse

(MTRNABT) damage. However, a large amount of the variance was not accounted.

Additional pathologic processes that conceivably contribute to cortical thinning include cortical demyelination, synaptic loss, and glial loss (Wegner et al., 2006). However, cMRI cannot detect such damage. The relationship between focal WM damage and cortical pathology and its dependence on the phase of MS have been studied within both the histopathological (Frischer et al., 2009; Kutzelnigg et al., 2005) and imaging

148

communities (Bendfeldt et al., 2009; Calabrese et al., 2007a; Ceccarelli et al., 2008;

Charil et al., 2007; Sailer et al., 2003; Sepulcre et al., 2006). Our evidence showed that both focal (T2LV, β = 0.50) and diffuse damage (MTRNABT, β = 0.23) predicted the baseline CTh, but only focal damage (T2LV, β = 0.33, and cumulative GdLV, β = 0.31) remained in the model of ∆CTh, suggesting, at least in our MS patients, that focal WM lesions played a significant role in the development of cortical atrophy.

Another question concerned the relationship between volumetric GM atrophy measures and CTh measures (Winkler et al., 2009), and which approach might be more sensitive and robust for detection of irreversible neuronal loss and ultimately MS clinical progression. There have been many previous studies investigating the clinical and biological relevance of GMF. GMF has been shown to be correlated with lesion volumes,

(Chard et al., 2002a; Dalton et al., 2004; Fisher et al., 2008; Fisniku et al., 2008; Ge et al.,

2001a; Quarantelli et al., 2003; Sanfilipo et al., 2005), disability scores (Chard et al.,

2002a; Chard et al., 2004; Ge et al., 2001a; Sanfilipo et al., 2005; Tiberio et al., 2005), and cognitive impairment (Amato et al., 2004; Amato et al., 2007; Benedict et al., 2006;

De Stefano et al., 2003). The strength of the correlations varied with study population and length of follow-up, but generally GMF correlated more strongly with clinical measures than other conventional MRI measurements. CTh was also correlated to disability (Charil et al., 2007; Sailer et al., 2003) and cognitive impairment (Calabrese et al., 2010c).

Interestingly CTh appeared to have regional specificity: thinner primary motor cortex was found in patients with pyramidal onset, and thinner visual cortex was found in patients with optical neuritis onset when compared to patients with other types of symptoms at first presentation (Calabrese et al., 2007b). Fatigue correlated to CTh in

149

parietal lobes (Pellicano et al., 2010). These results suggested that cortical atrophy was significantly involved in MS and was clinically relevant.

In this study, a moderate correlation was found between baseline CTh and concurrent GMF (r = 0.55, p < 0.001). This weaker than expected relationship might be attributable to deep GM and cerebellar atrophy, which were known to be significant in

MS (Audoin et al., 2006; Cifelli et al., 2002; Ramasamy et al., 2009) but not included in

CTh measurements. The correlation between annualized GMF percent change during the study and annualized change in global CTh was low (r = 0.38, p = 0.001), and might indicate variability in regional GM atrophy (Audoin et al., 2006). Thus, cortical and deep

GM atrophy might evolve differently and further longitudinal studies for each GM component as well as regional cortical atrophy are needed to determine the specific relationship and clinical deficit.

Some important technical issues related to the study are worth mentioning. First, we measured CTh from T1-weighted spin echo images (T1SE) with 5mm slice thickness, which are typical images acquired in prior MS clinical trials, but not ideal for estimating

CTh. These are the images available from the standardized MRI protocol used for our long-term observational study of atrophy in MS, which began in 1999. Most of the previous CTh studies in MS used FreeSurfer, but it requires high-resolution, high- contrast MRI and is not applicable to T1SE image. The analysis software, CLADA, was designed to reliably measure ∆CTh from these images, but higher-resolution and higher- contrast 3D images can potentially improve the sensitivity. Despite this limitation, CTh differences were detected between MS and control subjects in the present study. Several factors contributed to the detection of cortical thinning: (1) the longitudinal nature of

150

CLADA which uses registration and deformable models and was sensitive to subtle changes and (2) the use of multiple time-points (9 time-points during 4-year semi-annual

MRI in MS and 5 time-points in controls) to calculate the linear regressions instead of a simple percent difference from 2 time-points. Second, there was a potential influence of

MS lesions on CTh measurements. Most MS lesions were periventricular and did not significantly influence the cortical surface detection except during segmentation. In

CLADA, a lesion mask was created on the subject-specific average image to eliminate systematic segmentation errors. However, juxtacortical lesions might still affect CTh measurements, and the use of T1SE was advantageous over MPRAGE because juxtacortical lesions are rarely visible on T1SE whereas MPRAGE is sensitive to these lesions. In other surface-based cortical surface detection algorithms, these hypointense juxtacortical lesions might require manual correction and result in higher variability because of inconsistent mesh constructions. Intracortical lesions are not detectable in

T1SE images and rarely visible even with MPRAGE (Nelson et al., 2008). Some intracortical lesions may be visible (Bagnato et al., 2009) but most of the cortical lesions cannot be detected by cMRI, even by advanced MRI methods such as double inversion recovery (Geurts et al., 2005a) and phase-sensitive inversion recovery (Nelson et al.,

2007). Thus, cortical lesions were unlikely to affect CTh measurements in our study.

Our results indicate that cortical atrophy is clinically relevant in MS, and therefore, cortical atrophy is a potential marker of MS progression. However, the source of cortical GM tissue loss remains unknown, and such tissue loss may be caused by a combination of both WM and GM damage. Further studies are needed to investigate the effect of cortical and WM lesions on cortical thinning. The lower degree of inflammation

151

in the cortex and the likely association between CTh and neuronal density make CTh a potentially valuable outcome measure for MS clinical trials testing neuroprotection.

152

CHAPTER 6 SUMMARY AND FUTURE WORK

6.1 Summary

In this work, new methods to measure GMF and CTh from brain MRI were developed and validated. Using these techniques and longitudinally acquired MRI of MS patients and healthy normal controls, we determined the rates of GM and cortical atrophy.

The rates of GM and cortical atrophy in MS patients were greater than those in healthy controls, correlated with clinical disability, and were different among the patients who worsened and remained stable. The main results and conclusion are summarized here, followed by possible future direction.

In Chapter 2, a new automated GM segmentation method to measure normalized

GM volume (GMF) was developed and validated. The method was designed to be applicable for brain MRI of MS patients where typical MS features such as focal lesions and significant brain atrophy were taken into account. We validated this method in terms of reproducibility, accuracy, and sensitivity to lesion size. The average scan-rescan reproducibility using T1SE images from 9 MS patients, acquired 3 times over 2 weeks, was 1.1% for GMF. The accuracy, as measured by the similarity index, was 0.938 and

0.836 in comparison to gold standard in simulated MRIs and in comparison to manual tracing in real MRIs, respectively. The accuracy was better than commonly used software, particularly when analyzing low-resolution clinical MRIs. We determined that the

153

method could be reliably used for clinical MRIs of MS patients and for longitudinal study of duration 2 years or longer. We also discovered that the size of lesions systematically affected GM volume, and this bias could be statistically corrected.

In Chapter 3, we applied the segmentation algorithm to a 4-year longitudinal study of MS patients and age-matched healthy normal controls and measured GM atrophy. GM atrophy was calculated by measuring GMF at baseline and at year 4, subtracting the two measurements, and dividing by the interval time (4 years). There was a stage-dependent acceleration in the annualized rates of GM atrophy in MS patients while WM atrophy remained constant among MS patients. There were moderately strong correlations between GM atrophy and clinical measures of disability. The rates of GM atrophy differed when the MS group was subdivided into those who worsened and those remained stable. These findings demonstrated the clinical relevance of GM atrophy in

MS. We also found that predictors of GM atrophy differed between RRMS and SPMS groups, which suggested different mechanisms of neurodegeneration in RRMS and

SPMS.

In Chapter 4, we developed and validated CLADA, which measured cortical atrophy in terms of change in CTh. This algorithm was a registration and deformable model-based longitudinal method. CLADA used images from all time-points to create a subject-specific template image, generate a deformable cortical surface model, and deform the model to fit each image from different time-points. The validation tests for

CLADA included various reproducibility measures, accuracy, sensitivity to change, comparison of image sequence, bias in estimating atrophy, comparison of full and cross- sectional versions of CLADA, and systematic effect of WM lesions. The validation tests

154

showed that CLADA was reliable, more reproducible than cross-sectional method, could measure CTh change from low-resolution MRI, and had small bias in estimating atrophy.

In Chapter 5, CLADA was applied to the full longitudinal data from Chapter 3

(nine MRIs acquired semi-annually for patients and five MRIs acquired annually for controls) to study cortical atrophy in MS. The results showed that CLADA detected cortical atrophy in MS patients. Similar to GM atrophy measured by GMF, cortical atrophy was greater in patients with more advanced disease and correlated with clinical disability. Cortical atrophy distinguished stable and worsening patients with more significance than GM atrophy and was predicted by baseline T2 lesion volume and cumulative Gd-enhancing lesion volume in the MS group.

Overall this work demonstrated the clinical importance of GM and cortical pathology and showed that GM and cortical atrophy could be used as a reliable imaging marker of clinical progression, in particular for assessing potential neuroprotective therapies in MS. It also suggested that mechanisms of RRMS and SPMS differed which might imply different therapeutic strategies are necessary for neuroprotection in different

MS stages.

6.2 Limitations

The limitations of each study were described in the individual chapters, and here we summarize the limitations that affected all studies.

• The GM lesions were not studied in this project even though these lesions were

believed to be relevant (Calabrese et al., 2007a; Kutzelnigg and Lassmann, 2005).

Current conventional MRI techniques are insensitive to cortical lesions (Geurts et al.,

155

2005b; Kidd et al., 1999; Newcombe et al., 1991), and the use of new MRI techniques

such as double inversion recovery (DIR) or phase sensitive inversion recovery (PSIR)

might be advantageous in investigating the relationship between GM damage and

clinical correlation. However, DIR could only detect about 30% of cortical lesions

compared to the histopathological standard despite its improved sensitivity (Seewann

et al., 2010). Furthermore, there were no standard pulse sequences to acquire DIR,

and the inter-rater variability is high (Geurts et al., 2011).

• The studied patients were allowed to be on DMTs over the course of the longitudinal

study, which might have reduced the rate of atrophy. Patients could also be switching

drugs, which further complicated the interpretation of our results. However, the

medication is recorded and considered in the analysis. Furthermore, the use of

untreated patients would have ethical concerns and potential bias as untreated patients

might have the benign course due to their genetic disposition. The use of

corticosteroids also affected MRI measurements, and thus the patient visits were

delayed to minimize the possible artifact.

• During the course of the study, the scanner was changed from Siemens Vision to

Siemens Symphony. We acquired images from 9 MS patients with varying degrees of

atrophy and lesions from both scanners on the same day and used them for calibrating

the measurements. The calibration involved a linear regression of GMF and CTh

measurements and CLADA weighting parameter (α = 1.2). The consistency of

longitudinal data was also monitored by the measurements from healthy normal

controls as seen in Figure 3-1 and Figure 5-3(b).

156

6.3 Future Work

There are numerous potential studies that can build on the methods developed here. We provide five examples of such potential studies. Two of them (6.3.2 and 6.3.3) have already started with promising results (Fisher et al., 2010; Nakamura et al., 2010).

6.3.1 Higher Field MRI

Current clinical study used images from 1.5T MRI scanners. Higher field MRIs including 3T and 7T have become increasingly accessible, and the images from high field scanners can provide smaller voxel size, shorter scanning time, and/or higher signal-to- noise ratio (SNR). Segmentation algorithms benefit from high field MRI because (1) high resolution MRI reduces partial volume effects, (2) shorter scanning time reduces the risk of patient motion, and (3) high SNR clarifies the borders between different tissues. Some disadvantages also exist: (1) there are greater susceptibility artifacts, (2) geometric distortions also increase due to both object-induced susceptibility and non-uniform main magnetic field, (3) the specific absorption rate (SAR) is increased and may prevent using long sequences with numerous RF pulses, and (4) variations in phase cause signal decrease (signal drop-off) and increase (hot spot). Therefore, the developed methods will likely benefit from higher field MRI but may need to be updated with advancing MRI technologies.

6.3.2 Treatment Effect on Gray Matter Atrophy

This future project investigates the treatment effect on GM and cortical atrophy.

Although it is well established that whole brain atrophy responds to therapies (Barkhof et

157

al., 2009), the treatment effect on GM atrophy is not well established (Bendfeldt et al.,

2010; Zivadinov et al., 2007). The response to treatment is essential for GM atrophy to be a surrogate marker and to be applied in clinical trials. The image analysis software developed in Chapter 2 and Chapter 4 can be directly used in retrospective analysis of previous longitudinal clinical trials. If the treatment effect is confirmed, GM atrophy measures can be used to evaluate future therapeutic candidates.

6.3.3 Investigation on Pseudoatrophy Effect

Another potential application of GM atrophy is to study the pseudoatrophy effect.

It is believed that the resolution of inflammatory edema causes apparent whole brain atrophy in the treated patients during the first year of anti-inflammatory therapy and that reduced whole brain atrophy in the second year reflects the treatment effect (Section 1.5).

Since inflammation is absent or significantly reduced in GM (Kidd et al., 1999; Peterson et al., 2001), the observed acceleration of whole brain atrophy is likely caused by significant WM atrophy in the first year, and the observed treatment effect on whole brain atrophy is likely due to reduced GM atrophy in the treated patients in the second year. The tools we developed here can be used to test this hypothesis. This longitudinal study will be similar to the first future study but require at least one intermediate time- point to study the temporal evolution of whole brain, GM, and WM atrophy. If no pseudoatrophy effect is found in GM after anti-inflammatory therapies, GM atrophy becomes even more attractive marker of disease progression in MS because it implies that GM or cortical atrophy are more specific to irreversible damage than whole brain atrophy.

158

6.3.4 Regional Correlation of GM Atrophy and WM Damage

Another potential future work is to study the relationship between GM atrophy and WM damage. The relationship between GM and WM damage is uncertain and warrants further studies. In the case of secondary neurodegeneration, initial WM damage causes axonal transection, which leads to Wallerian, retrograde, and trans-synaptic degeneration, and ultimately GM atrophy. This hypothesis can be supported by a high correlation between GM atrophy and WM damage. Previous studies have been inconclusive using voxel-based morphometry methods (Battaglini et al., 2009; Bendfeldt et al., 2009; Bendfeldt et al., 2009; Bendfeldt et al., 2010; Sepulcre et al., 2009), which may not be sensitive enough for small change (Anderson et al., 2006; Hutton et al., 2009).

In this research, global measures of GM tissue loss were used (GMF and global CTh), and we found that the relationship between GM atrophy and focal (T2LV) WM and diffuse whole brain (MTRNABT) damage changed from RRMS to SPMS. In the future study, regional measures can be used to confirm this relationship. Since global measures do not account for spatial correspondence between GM and WM, regional comparisons can provide more specific relationship. For this study, frequent MRI is advantageous because delay effects can be explored.

6.3.5 Regional Correlation of GM Atrophy and Local Damage

If the cause of GM atrophy is not WM pathology, either diffuse or focal, the presumed cause is local GM damage. Since there is no reliable method to visualize entire

GM damage in MS, there is currently no in vivo MRI study that has examined this

159

relationship. However, in previous post-mortem and animal studies, there has been no correlation between GM atrophy and local GM demyelination (Pomeroy et al., 2008;

Pomeroy et al., 2010; Wegner et al., 2006). However, these studies were restricted to 2D histological slides when studying cortical atrophy, which is known to be inaccurate

(Fischl and Dale, 2000). Therefore, if applied to the autopsy cases, CLADA may provide more accurate and definitive insight into the relationship between GM atrophy and GM demyelination.

6.4 Conclusion

In conclusion, I have characterized the evolution of GM and cortical atrophy, demonstrated that these atrophy measures were clinically relevant in MS, and provided in vivo insights on possible MS mechanisms. The developed algorithms have been validated and shown to be reliable. The use of GM atrophy and cortical atrophy appear to be highly promising as a potential imaging marker in clinical trials, and further elucidation of GM and cortical atrophy will provide even better understanding of MS mechanisms and aid in the development of effective therapies in MS.

160

BIBLIOGRAPHY

Acosta O, Bourgeat P, Zuluaga MA, Fripp J, Salvado O, Ourselin S, Alzheimer's Disease Neuroimaging Initiative. 2009. Automated voxel-based 3D cortical thickness measurement in a combined lagrangian-eulerian PDE approach using partial volume maps. Med.Image Anal. 13(5):730-743.

Aganj I, Sapiro G, Parikshak N, Madsen SK, Thompson PM. 2009. Measurement of cortical thickness from MRI by minimum line integrals on soft-classified tissue. Hum.Brain Mapp. 30(10):3188-3199.

Ahmed MN, Yamany SM, Mohamed N, Farag AA, Moriarty T. 2002. A modified fuzzy c-means algorithm for bias field estimation and segmentation of MRI data. IEEE Trans.Med.Imaging 21(3):193-199.

Albert M, Antel J, Bruck W, Stadelmann C. 2007. Extensive cortical remyelination in patients with chronic multiple sclerosis. Brain Pathol. 17(2):129-138.

Allen IV, McQuaid S, Mirakhur M, Nevin G. 2001. Pathological abnormalities in the normal-appearing white matter in multiple sclerosis. Neurol.Sci. 22(2):141-144.

Altmann DR, Jasperse B, Barkhof F, Beckmann K, Filippi M, Kappos LD, Molyneux P, Polman CH, Pozzilli C, Thompson AJ, Wagner K, Yousry TA, Miller DH. 2009. Sample sizes for brain atrophy outcomes in trials for secondary progressive multiple sclerosis. Neurology 72(7):595-601.

Amato MP, Bartolozzi ML, Zipoli V, Portaccio E, Mortilla M, Guidi L, Siracusa G, Sorbi S, Federico A, De Stefano N. 2004. Neocortical volume decrease in relapsing- remitting MS patients with mild cognitive impairment. Neurology 63(1):89-93.

Amato MP, Portaccio E, Goretti B, Zipoli V, Battaglini M, Bartolozzi ML, Stromillo ML, Guidi L, Siracusa G, Sorbi S, Federico A, De Stefano N. 2007. Association of neocortical volume changes with cognitive deterioration in relapsing-remitting multiple sclerosis. Arch Neurol 64(8):1157-1161.

Amato U, Larobina M, Antoniadis A, Alfano B. 2003. Segmentation of magnetic resonance brain images through discriminant analysis. J.Neurosci.Methods 131(1- 2):65-74.

161

Andersen AH, Zhang Z, Avison MJ, Gash DM. 2002. Automated segmentation of multispectral brain MR images. J.Neurosci.Methods 122:13-23.

Anderson DW, Ellenberg JH, Leventhal CM, Reingold SC, Rodriguez M, Silberberg DH. 1992. Revised estimate of the prevalence of multiple sclerosis in the United States. Ann.Neurol. 31(3):333-336.

Anderson VM, Fox NC, Miller DH. 2006. Magnetic resonance imaging measures of brain atrophy in multiple sclerosis. J.Magn.Reson.Imaging 23(5):605-618.

Ardekani BA, Guckemus S, Bachman A, Hoptman MJ, Wojtaszek M, Nierenberg J. 2005. Quantitative comparison of algorithms for inter-subject registration of 3D volumetric brain MRI scans. J.Neurosci.Methods 142(1):67-76.

Arnold D. 2007. Evidence for neuroprotection and remyelination using imaging techniques. Neurology 68(22 suppl 3):S83-90.

Ashburner J and Friston KJ. 2005. Unified segmentation. NeuroImage 26(3):839-851.

Ashburner J and Friston KJ. 2000. Voxel-based morphometry--the methods. NeuroImage 11(6 Pt 1):805-821.

Audoin B, Davies GR, Finisku L, Chard DT, Thompson AJ, Miller DH. 2006. Localization of grey matter atrophy in early RRMS : A longitudinal study. J.Neurol. 253(11):1495-1501.

Avants BB, Epstein CL, Grossman M, Gee JC. 2008. Symmetric diffeomorphic image registration with cross-correlation: Evaluating automated labeling of elderly and neurodegenerative brain. Med.Image Anal. 12(1):26-41.

Azevedo FA, Carvalho LR, Grinberg LT, Farfel JM, Ferretti RE, Leite RE, Jacob Filho W, Lent R, Herculano-Houzel S. 2009. Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain. J.Comp.Neurol. 513(5):532-541.

Bagnato F, Yao B, Cantor F, Merkle H, Condon E, Montequin M, Moore S, Quezado M, Tkaczyk D, McFarland H. 2009. Multisequence-imaging protocols to detect cortical lesions of patients with multiple sclerosis: Observations from a post-mortem 3 tesla imaging study. J.Neurol.Sci. 282(1-2):80-85.

Bagnato F, Butman JA, Gupta S, Calabrese M, Pezawas L, Ohayon JM, Tovar-Moll F, Riva M, Cao MM, Talagala SL, McFarland HF. 2006. In vivo detection of cortical plaques by MR imaging in patients with multiple sclerosis. AJNR Am.J.Neuroradiol. 27(10):2161-2167.

162

Bakshi R, Thompson AJ, Rocca MA, Pelletier D, Dousset V, Barkhof F, Inglese M, Guttmann CR, Horsfield MA, Filippi M. 2008. MRI in multiple sclerosis: Current status and future prospects. Lancet Neurol. 7(7):615-625.

Bammer R, Augustin M, Strasser-Fuchs S, Seifert T, Kapeller P, Stollberger R, Ebner F, Hartung HP, Fazekas F. 2000. Magnetic resonance diffusion tensor imaging for characterizing diffuse and focal white matter abnormalities in multiple sclerosis. Magn.Reson.Med. 44(4):583-591.

Barkhof F. 1999. MRI in multiple sclerosis: Correlation with expanded disability status scale (EDSS). Mult.Scler. 5(4):283-286.

Barkhof F, Calabresi PA, Miller DH, Reingold SC. 2009. Imaging outcomes for neuroprotection and repair in multiple sclerosis trials. Nat.Rev.Neurol. 5(5):256-266.

Barkhof F, Bruck W, De Groot CJ, Bergers E, Hulshof S, Geurts J, Polman CH, van der Valk P. 2003. Remyelinated lesions in multiple sclerosis: Magnetic resonance image appearance. Arch.Neurol. 60(8):1073-1081.

Basser PJ, Mattiello J, LeBihan D. 1994. MR diffusion tensor spectroscopy and imaging. Biophys.J. 66(1):259-267.

Battaglini M, Giorgio A, Stromillo ML, Bartolozzi ML, Guidi L, Federico A, De Stefano N. 2009. Voxel-wise assessment of progression of regional brain atrophy in relapsing-remitting multiple sclerosis. J.Neurol.Sci. 282(1-2):55-60.

Bendfeldt K, Kappos L, Radue EW, Borgwardt SJ. 2009. Progression of gray matter atrophy and its association with white matter lesions in relapsing-remitting multiple sclerosis. J.Neurol.Sci. 285(1-2):268-9; author reply 269.

Bendfeldt K, Egger H, Nichols TE, Loetscher P, Denier N, Kuster P, Traud S, Mueller- Lenke N, Naegelin Y, Gass A, Kappos L, Radue EW, Borgwardt SJ. 2010. Effect of immunomodulatory medication on regional gray matter loss in relapsing-remitting multiple sclerosis--a longitudinal MRI study. Brain Res. 1325:174-182.

Bendfeldt K, Kuster P, Traud S, Egger H, Winklhofer S, Mueller-Lenke N, Naegelin Y, Gass A, Kappos L, Matthews PM, Nichols TE, Radue EW, Borgwardt SJ. 2009. Association of regional gray matter volume loss and progression of white matter lesions in multiple sclerosis - A longitudinal voxel-based morphometry study. NeuroImage 45(1):60-67.

Bendfeldt K, Blumhagen JO, Egger H, Loetscher P, Denier N, Kuster P, Traud S, Mueller-Lenke N, Naegelin Y, Gass A, Hirsch J, Kappos L, Nichols TE, Radue EW, Borgwardt SJ. 2010. Spatiotemporal distribution pattern of white matter lesion volumes and their association with regional grey matter volume reductions in relapsing-remitting multiple sclerosis. Hum.Brain Mapp. .

163

Benedict RHB, Bruce JM, Dwyer MG, Abdelrahman N, Hussein S, Weinstock-Guttman B, Garg N, Munschauer F, Zivadinov R. 2006. Neocortical atrophy, third ventricular width, and cognitive dysfunction in multiple sclerosis. Arch Neurol 63(9):1301-1306.

Besag J. 1986. On the statistical analysis of dirty pictures. J. R. Statist. Soc. B 48:259-302.

Bezdek JC, Hall LO, Clarke LP. 1993. Review of MR image segmentation techniques using pattern recognition. Med.Phys. 20(4):1033-1048.

Bjartmar C, Wujek JR, Trapp BD. 2003. Axonal loss in the pathology of MS: Consequences for understanding the progressive phase of the disease. J.Neurol.Sci. 206(2):165-171.

Bjartmar C, Kinkel RP, Kidd G, Rudick RA, Trapp BD. 2001. Axonal loss in normal- appearing white matter in a patient with acute MS. Neurology 57(7):1248-1252.

Bjartmar C, Kidd G, Mork S, Rudick R, Trapp BD. 2000. Neurological disability correlates with spinal cord axonal loss and reduced N-acetyl aspartate in chronic multiple sclerosis patients. Ann.Neurol. 48(6):893-901.

Bo L, Vedeler CA, Nyland HI, Trapp BD, Mork SJ. 2003. Subpial demyelination in the cerebral cortex of multiple sclerosis patients. J.Neuropathol.Exp.Neurol. 62(7):723- 732.

Bo L, Dawson TM, Wesselingh S, Mork S, Choi S, Kong PA, Hanley D, Trapp BD. 1994. Induction of nitric oxide synthase in demyelinating regions of multiple sclerosis brains. Ann.Neurol. 36(5):778-786.

Botsch M and Kobbelt L. 2004. A remeshing approach to multiresolution modeling. Proc. Eurographics/ACM SIGGRAPH Symp. Geometry Processing :185-192.

Brink BP, Veerhuis R, Breij EC, van der Valk P, Dijkstra CD, Bo L. 2005. The pathology of multiple sclerosis is location-dependent: No significant complement activation is detected in purely cortical lesions. J.Neuropathol.Exp.Neurol. 64(2):147-155.

Bruck W, Kuhlmann T, Stadelmann C. 2003. Remyelination in multiple sclerosis. J.Neurol.Sci. 206(2):181-185.

Calabrese M, Rocca MA, Atzori M, Mattisi I, Favaretto A, Perini P, Gallo P, Filippi M. 2010a. A 3-year magnetic resonance imaging study of cortical lesions in relapse-onset multiple sclerosis. Ann.Neurol. 67(3):376-383.

Calabrese M, Battaglini M, Giorgio A, Atzori M, Bernardi V, Mattisi I, Gallo P, De Stefano N. 2010b. Imaging distribution and frequency of cortical lesions in patients with multiple sclerosis. Neurology 75(14):1234-1240.

164

Calabrese M, Rinaldi F, Mattisi I, Grossi P, Favaretto A, Atzori M, Bernardi V, Barachino L, Romualdi C, Rinaldi L, Perini P, Gallo P. 2010c. Widespread cortical thinning characterizes patients with MS with mild cognitive impairment. Neurology 74(4):321-328.

Calabrese M, Rocca MA, Atzori M, Mattisi I, Bernardi V, Favaretto A, Barachino L, Romualdi C, Rinaldi L, Perini P, Gallo P, Filippi M. 2009. Cortical lesions in primary progressive multiple sclerosis: A 2-year longitudinal MR study. Neurology 72(15):1330-1336.

Calabrese M, DeStefano N, Atzori M, Bernardi V, Mattisi I, Barachino L, Morra A, Rinaldi L, Romualdi C, Perini P, Battistin L, Gallo P. 2007a. Detection of cortical inflammatory lesions by double inversion recovery magnetic resonance imaging in patients with multiple sclerosis. Arch Neurol 64(10):1416-1422.

Calabrese M, Atzori M, Bernardi V, Morra A, Romualdi C, Rinaldi L, McAuliffe MJM, Barachino L, Perini P, Fischl B, Battistin L, Gallo P. 2007b. Cortical atrophy is relevant in multiple sclerosis at clinical onset. J.Neurol. 254(9):1212-1220.

Camara O, Schweiger M, Scahill RI, Crum WR, Sneller BI, Schnabel JA, Ridgway GR, Cash DM, Hill DL, Fox NC. 2006. Phenomenological model of diffuse global and regional atrophy using finite-element methods. IEEE Trans.Med.Imaging 25(11):1417-1430.

Ceccarelli A, Rocca MA, Pagani E, Colombo B, Martinelli V, Comi G, Filippi M. 2008. A voxel-based morphometry study of grey matter loss in MS patients with different clinical phenotypes. NeuroImage 42(1):315-322.

Ceccarelli A, Rocca MA, Falini A, Tortorella P, Pagani E, Rodegher M, Comi G, Scotti G, Filippi M. 2007. Normal-appearing white and grey matter damage in MS. J.Neurol. 254(4):513-518.

Cercignani M, Bozzali M, Iannucci G, Comi G, Filippi M. 2001. Magnetisation transfer ratio and mean diffusivity of normal appearing white and grey matter from patients with multiple sclerosis. J.Neurol.Neurosurg.Psychiatry. 70(3):311-317.

Chalana V, Ng L, Rystrom LR, Gee JC, Haynor DR. 2001. Validation of brain segmentation and tissue classification algorithm for T1-weighted MR images. Medical Imaging 2001: Image Processing 4322(1):1873-1882.

Chard DT, Griffin CM, Parker GJ, Kapoor R, Thompson AJ, Miller DH. 2002a. Brain atrophy in clinically early relapsing-remitting multiple sclerosis. Brain 125(Pt 2):327- 337.

165

Chard DT, Griffin CM, McLean MA, Kapeller P, Kapoor R, Thompson AJ, Miller DH. 2002b. Brain metabolite changes in cortical grey and normal-appearing white matter in clinically early relapsing-remitting multiple sclerosis. Brain 125(Pt 10):2342-2352.

Chard DT, Griffin CM, Rashid W, Davies GR, Altmann DR, Kapoor R, Barker GJ, Thompson AJ, Miller DH. 2004. Progressive grey matter atrophy in clinically early relapsing-remitting multiple sclerosis. Mult.Scler. 10(4):387-391.

Charil A, Dagher A, Lerch JP, Zijdenbos AP, Worsley KJ, Evans AC. 2007. Focal cortical atrophy in multiple sclerosis: Relation to lesion load and disability. NeuroImage 34:509-517.

Chen JT, Schneider C, Nakamura K, Easley K, Kidd GJ, Chang A, Staugaitis SM, Fox RJl, Fisher E, Arnold DL, Trapp BD. 2010. Validation of MRI-based measurements of subpial cortical demyelination in an MS brain. ECTRIMS 2010 .

Chen J, Collins D, Derakhshan M, Arnold D. 2008. Quantifying MRI with increased specificity for MS pathology: A longitudinal method for obtaining whole-brain metrics of 3D maps derived from non-conventional MRI. Proc. Int. Soc. Magn. Reson. Med. :2136.

Chen JT, Narayanan S, Collins DL, Smith SM, Matthews PM, Arnold DL. 2004. Relating neocortical pathology to disability progression in multiple sclerosis using MRI. NeuroImage 23(3):1168-1175.

Chen JT, Kuhlmann T, Jansen GH, Collins DL, Atkins HL, Freedman MS, O'Connor PW, Arnold DL, Canadian MS/BMT Study Group. 2007. Voxel-based analysis of the evolution of magnetization transfer ratio to quantify remyelination and demyelination with histopathological validation in a multiple sclerosis lesion. NeuroImage 36(4):1152-1158.

Chiaravalloti ND and DeLuca J. 2008. Cognitive impairment in multiple sclerosis. Lancet Neurol. 7(12):1139-1151.

Ciccarelli O, Werring DJ, Wheeler-Kingshott CA, Barker GJ, Parker GJ, Thompson AJ, Miller DH. 2001. Investigation of MS normal-appearing brain using diffusion tensor MRI with clinical correlations. Neurology 56(7):926-933.

Cifelli A, Arridge M, Jezzard P, Esiri MM, Palace J, Matthews PM. 2002. Thalamic neurodegeneration in multiple sclerosis. Ann.Neurol. 52(5):650-653.

Clark MC, Hall LO, Goldgof DB, Clarke LP, Velthuizen RP, Silbiger MS. 1994. MRI segmentation using fuzzy clustering techniques. IEEE Engineering in Medicine and Biology Magazine 13(5):730-742.

166

Cohen JA, Calabresi PA, Chakraborty S, Edwards KR, Eickenhorst T, Felton WL,3rd, Fisher E, Fox RJ, Goodman AD, Hara-Cleaver C, Hutton GJ, Imrey PB, Ivancic DM, Mandell BF, Perryman JE, Scott TF, Skaramagas TT, Zhang H, ACT Investigators. 2008. Avonex combination trial in relapsing--remitting MS: Rationale, design and baseline data. Mult.Scler. 14(3):370-382.

Collins DL, Zijdenbos AP, Kollokian V, Sled JG, Kabani NJ, Holmes CJ, Evans AC. 1998. Design and construction of a realistic digital brain phantom. IEEE Trans.Med.Imaging 17(3):463-468.

Compston A and Coles A. 2008. Multiple sclerosis. Lancet 372(9648):1502-1517.

Confavreux C and Vukusic S. 2008. The clinical epidemiology of multiple sclerosis. Neuroimaging Clin.N.Am. 18(4):589-622, ix-x.

Confavreux C and Vukusic S. 2006. Natural history of multiple sclerosis: A unifying concept. Brain 129(Pt 3):606-616.

Confavreux C, Vukusic S, Moreau T, Adeleine P. 2000. Relapses and progression of disability in multiple sclerosis. N.Engl.J.Med. 343(20):1430-1438.

Cotton F, Weiner HL, Jolesz FA, Guttmann CR. 2003. MRI contrast uptake in new lesions in relapsing-remitting MS followed at weekly intervals. Neurology 60(4):640- 646.

Dale AM, Fischl B, Sereno MI. 1999. Cortical surface-based analysis. I. segmentation and surface reconstruction. NeuroImage 9(2):179-194.

Dalton CM, Chard DT, Davies GR, Miszkiel KA, Altmann DR, Fernando K, Plant GT, Thompson AJ, Miller DH. 2004. Early development of multiple sclerosis is associated with progressive grey matter atrophy in patients presenting with clinically isolated syndromes. Brain 127(Pt 5):1101-1107.

Das SR, Avants BB, Grossman M, Gee JC. 2009. Registration based cortical thickness measurement. NeuroImage 45(3):867-879.

Davatzikos CA and Prince JL. 1995. An active contour model for mapping the cortex. IEEE Trans.Med.Imaging 14(1):65-80.

Davie CA, Barker GJ, Webb S, Tofts PS, Thompson AJ, Harding AE, McDonald WI, Miller DH. 1995. Persistent functional deficit in multiple sclerosis and autosomal dominant cerebellar ataxia is associated with axon loss. Brain 118 ( Pt 6)(Pt 6):1583- 1592.

De Groot CJ, Bergers E, Kamphorst W, Ravid R, Polman CH, Barkhof F, van der Valk P. 2001. Post-mortem MRI-guided sampling of multiple sclerosis brain lesions:

167

Increased yield of active demyelinating and (p)reactive lesions. Brain 124(Pt 8):1635- 1645.

De Stefano N, Battaglini M, Stromillo ML, Zipoli V, Bartolozzi ML, Guidi L, Siracusa G, Portaccio E, Giorgio A, Sorbi S, Federico A, Amato MP. 2006. Brain damage as detected by magnetization transfer imaging is less pronounced in benign than in early relapsing multiple sclerosis. Brain 129(8):2008-2016.

De Stefano N, Matthews PM, Arnold DL. 1995a. Reversible decreases in N- acetylaspartate after acute brain injury. Magn.Reson.Med. 34(5):721-727.

De Stefano N, Matthews PM, Antel JP, Preul M, Francis G, Arnold DL. 1995b. Chemical pathology of acute demyelinating lesions and its correlation with disability. Ann.Neurol. 38(6):901-909.

De Stefano N, Matthews PM, Filippi M, Agosta F, De Luca M, Bartolozzi ML, Guidi L, Ghezzi A, Montanari E, Cifelli A, Federico A, Smith SM. 2003. Evidence of early cortical atrophy in MS: Relevance to white matter changes and disability. Neurology 60(7):1157-1162.

Dousset V, Grossman RI, Ramer KN, Schnall MD, Young LH, Gonzalez-Scarano F, Lavi E, Cohen JA. 1992. Experimental allergic encephalomyelitis and multiple sclerosis: Lesion characterization with magnetization transfer imaging. Radiology 182(2):483- 491.

Duning T, Kloska S, Steinstrater O, Kugel H, Heindel W, Knecht S. 2005. Dehydration confounds the assessment of brain atrophy. Neurology 64(3):548-550.

Dutta R and Trapp BD. 2007. Pathogenesis of axonal and neuronal damage in multiple sclerosis. Neurology 68(22 Suppl 3):S22-31; discussion S43-54.

Ebers GC, Heigenhauser L, Daumer M, Lederer C, Noseworthy JH. 2008. Disability as an outcome in MS clinical trials. Neurology 71(9):624-631.

Evangelou N, Esiri MM, Smith S, Palace J, Matthews PM. 2000. Quantitative pathological evidence for axonal loss in normal appearing white matter in multiple sclerosis. Ann.Neurol. 47(3):391-395.

Evans MC, Barnes J, Nielsen C, Kim LG, Clegg SL, Blair M, Leung KK, Douiri A, Boyes RG, Ourselin S, Fox NC, Alzheimer's Disease Neuroimaging Initiative. 2010. Volume changes in Alzheimer's disease and mild cognitive impairment: Cognitive associations. Eur.Radiol. 20(3):674-682.

Ferreira da Silva AR. 2007. A dirichlet process mixture model for brain MRI tissue classification. Med.Image Anal. 11(2):169-182.

168

Filippi M and Rocca MA. 2007. Conventional MRI in multiple sclerosis. J.Neuroimaging 17 Suppl 1:3S-9S.

Filippi M, Cercignani M, Inglese M, Horsfield MA, Comi G. 2001. Diffusion tensor magnetic resonance imaging in multiple sclerosis. Neurology 56(3):304-311.

Filippi M, Rocca MA, Martino G, Horsfield MA, Comi G. 1998. Magnetization transfer changes in the normal appearing white matter precede the appearance of enhancing lesions in patients with multiple sclerosis. Ann.Neurol. 43(6):809-814.

Filippi M, Paty DW, Kappos L, Barkhof F, Compston DA, Thompson AJ, Zhao GJ, Wiles CM, McDonald WI, Miller DH. 1995. Correlations between changes in disability and T2-weighted brain MRI activity in multiple sclerosis: A follow-up study. Neurology 45(2):255-260.

Fischer JS, Rudick RA, Cutter GR, Reingold SC. 1999. The multiple sclerosis functional composite measure (MSFC): An integrated approach to MS clinical outcome assessment. national MS society clinical outcomes assessment task force. Mult.Scler. 5(4):244-250.

Fischl B, Liu A, Dale AM. 2001. Automated manifold surgery: Constructing geometrically accurate and topologically correct models of the human cerebral cortex. IEEE Trans.Med.Imaging 20(1):70-80.

Fischl B and Dale AM. 2000. Measuring the thickness of the human cerebral cortex from magnetic resonance images. Proc. Natl. Acad. Sci. 97(20):11050-11055.

Fisher E, Nakamura K, Pace A, Lee JC, Foulds P, Rudick RA. 2010. Associations between grey matter atrophy and disability in a placebo-controlled trial of interferon beta-1a in relapsing-remitting multiple sclerosis. ECTRIMS .

Fisher, E. and Rudick, R.A. 2003. Measurement of central nervous system atrophy. In: J. A. Cohen and R. A. Rudick (Eds.), Multiple Sclerosis Therapeutics, 2nd ed. Marin Dunitz Ltd., pp. 173-191.

Fisher E, Cothren RM, Tkach JA, Masaryk TJ, Cornhill JF. 1997. Knowledge-based 3D segmentation of the brain in MR images for quantitative multiple sclerosis lesion tracking. SPIE Medical Imaging 3034:599-610.

Fisher E, Lee JC, Nakamura K, Rudick RA. 2008. Gray matter atrophy in multiple sclerosis: A longitudinal study. Ann.Neurol. 64(3):255-265.

Fisher E, Chang A, Fox RJ, Tkach JA, Svarovsky T, Nakamura K, Rudick RA, Trapp BD. 2007. Imaging correlates of axonal swelling in chronic multiple sclerosis brains. Ann.Neurol. 62(3):219-228.

169

Fisniku LK, Chard DT, Jackson JS, Anderson VM, Altmann DR, Miszkiel KA, Thompson AJ, Miller DH. 2008. Gray matter atrophy is related to long-term disability in multiple sclerosis. Ann.Neurol. 64(3):247-254.

Fox NC and Schott JM. 2004. Imaging cerebral atrophy: Normal ageing to Alzheimer's disease. Lancet 363(9406):392-394.

Frank JA, Stone LA, Smith ME, Albert PS, Maloni H, McFarland HF. 1994. Serial contrast-enhanced magnetic resonance imaging in patients with early relapsing- remitting multiple sclerosis: Implications for treatment trials. Ann.Neurol. 36 Suppl:S86-90.

Frischer JM, Bramow S, Dal-Bianco A, Lucchinetti CF, Rauschka H, Schmidbauer M, Laursen H, Sorensen PS, Lassmann H. 2009. The relation between inflammation and neurodegeneration in multiple sclerosis brains. Brain; Brain .

Frohman EM, Shah A, Eggenberger E, Metz L, Zivadinov R, Stuve O. 2007. Corticosteroids for multiple sclerosis: I. application for treating exacerbations. Neurotherapeutics 4(4):618-626.

Furby J, Hayton T, Altmann D, Brenner R, Chataway J, Smith KJ, Miller DH, Kapoor R. 2010. A longitudinal study of MRI-detected atrophy in secondary progressive multiple sclerosis. J.Neurol. .

Ganter P, Prince C, Esiri MM. 1999. Spinal cord axonal loss in multiple sclerosis: A post-mortem study. Neuropathol.Appl.Neurobiol. 25(6):459-467.

Ge Y, Grossman RI, Udupa JK, Wei L, Mannon LJ, Polansky M, Kolson DL. 2000. Brain atrophy in relapsing-remitting multiple sclerosis and secondary progressive multiple sclerosis: Longitudinal quantitative analysis. Radiology 214(3):665-670.

Ge Y, Grossman RI, Babb JS, Rabin ML, Mannon LJ, Kolson DL. 2002. Age-related total gray matter and white matter changes in normal adult brain. part I: Volumetric MR imaging analysis. AJNR Am J Neuroradiol 23(8):1327-1333.

Ge Y, Grossman RI, Udupa JK, Babb JS, Nyul LG, Kolson DL. 2001a. Brain atrophy in relapsing-remitting multiple sclerosis: Fractional volumetric analysis of gray matter and white matter. Radiology 220(3):606-610.

Ge Y, Grossman R, Udupa J, Babb J, Kolson D, McGowan J. 2001b. Magnetization transfer ratio histogram analysis of gray matter in relapsing-remitting multiple sclerosis. AJNR Am J Neuroradiol 22(3):470-475.

Geurts JJ, Pouwels PJ, Uitdehaag BM, Polman CH, Barkhof F, Castelijns JA. 2005a. Intracortical lesions in multiple sclerosis: Improved detection with 3D double inversion-recovery MR imaging. Radiology 236(1):254-260.

170

Geurts JJ, Bo L, Pouwels PJ, Castelijns JA, Polman CH, Barkhof F. 2005b. Cortical lesions in multiple sclerosis: Combined postmortem MR imaging and histopathology. AJNR Am.J.Neuroradiol. 26(3):572-577.

Geurts JJ, Barkhof F, Castelijns JA, Uitdehaag BM, Polman CH, Pouwels PJ. 2004. Quantitative 1H-MRS of healthy human cortex, hippocampus, and thalamus: Metabolite concentrations, quantification precision, and reproducibility. J.Magn.Reson.Imaging 20(3):366-371.

Geurts JJ, Reuling IE, Vrenken H, Uitdehaag BM, Polman CH, Castelijns JA, Barkhof F, Pouwels PJ. 2006. MR spectroscopic evidence for thalamic and hippocampal, but not cortical, damage in multiple sclerosis. Magn.Reson.Med. 55(3):478-483.

Geurts JJ, Blezer EL, Vrenken H, van der Toorn A, Castelijns JA, Polman CH, Pouwels PJ, Bo L, Barkhof F. 2008. Does high-field MR imaging improve cortical lesion detection in multiple sclerosis? J.Neurol. 255(2):183-191.

Geurts JJ, Bo L, Roosendaal SD, Hazes T, Daniels R, Barkhof F, Witter MP, Huitinga I, van der Valk P. 2007. Extensive hippocampal demyelination in multiple sclerosis. J.Neuropathol.Exp.Neurol. 66(9):819-827.

Geurts JJ, Roosendaal SD, Calabrese M, Ciccarelli O, Agosta F, Chard DT, Gass A, Huerga E, Moraal B, Pareto D, Rocca MA, Wattjes MP, Yousry TA, Uitdehaag BM, Barkhof F, MAGNIMS Study Group. 2011. Consensus recommendations for MS cortical lesion scoring using double inversion recovery MRI. Neurology 76(5):418- 424.

Geurts JJG, Stys PK, Minagar A, Amor S, Zivadinov R. 2009. Gray matter pathology in (chronic) MS: Modern views on an early observation. J.Neurol.Sci. 282(1-2):12-20.

Geurts JJ and Barkhof F. 2008. Grey matter pathology in multiple sclerosis. Lancet Neurology 7(9):841-851.

Gilmore CP, Donaldson I, Bo L, Owens T, Lowe J, Evangelou N. 2009. Regional variations in the extent and pattern of grey matter demyelination in multiple sclerosis: A comparison between the cerebral cortex, cerebellar cortex, deep grey matter nuclei and the spinal cord. J.Neurol.Neurosurg.Psychiatry. 80(2):182-187.

Gilmore CP, Bo L, Owens T, Lowe J, Esiri MM, Evangelou N. 2006. Spinal cord gray matter demyelination in multiple sclerosis-a novel pattern of residual plaque morphology. Brain Pathol. 16(3):202-208.

Gilmore CP, Geurts JJG, Evangelou N, Bot JCJ, van Schijndel RA, Pouwels PJW, Barkhof F, Bö L. 2009. Spinal cord grey matter lesions in multiple sclerosis detected by post-mortem high field MR imaging. Multiple Sclerosis 15(2):180-188.

171

Goldenberg R, Kimmel R, Rivlin E, Rudzsky M. 2002. Cortex segmentation: A fast variational geometric approach. IEEE Trans.Med.Imaging 21(12):1544-1551.

Guimond A, Meunier J, Thirion J. 2000. Average brain models: A convergence study. Comput.Vision Image Understanding 77(2):192-210.

Haidar H, Egorova S, Soul JS. 2005. New numerical solution of the laplace equation for tissue thickness measurement in three-dimensional MRI. JMMA 4(1):83-97.

Haidar H and Soul JS. 2006. Measurement of cortical thickness in 3D brain MRI data: Validation of the laplacian method. J.Neuroimaging 16(2):146-153.

Han X, Pham DL, Tosun D, Rettmann ME, Xu C, Prince JL. 2004. CRUISE: Cortical reconstruction using implicit surface evolution. NeuroImage 23(3):997-1012.

Han X, Jovicich J, Salat D, van der Kouwe A, Quinn B, Czanner S, Busa E, Pacheco J, Albert M, Killiany R, Maguire P, Rosas D, Makris N, Dale A, Dickerson B, Fischl B. 2006. Reliability of MRI-derived measurements of human cerebral cortical thickness: The effects of field strength, scanner upgrade and manufacturer. NeuroImage 32(1):180-194.

Han X, Xu C, Prince JL. 2003. A topology preserving level set method for geometric deformable models. IEEE Trans.Pattern Anal.Mach.Intell. 25(6):755-768.

Henry RG, Shieh M, Okuda DT, Evangelista A, Gorno-Tempini ML, Pelletier D. 2008. Regional grey matter atrophy in clinically isolated syndromes at presentation. J.Neurol.Neurosurg.Psychiatry. 79(11):1236-1244.

Horakova D, Cox JL, Havrdova E, Hussein S, Dolezal O, Cookfair D, Dwyer MG, Seidl Z, Bergsland N, Vaneckova M, Zivadinov R. 2008. Evolution of different MRI measures in patients with active relapsing-remitting multiple sclerosis over 2 and 5 years: A case-control study. J.Neurol.Neurosurg.Psychiatry. 79(4):407-414.

Huitinga I, De Groot CJ, Van der Valk P, Kamphorst W, Tilders FJ, Swaab DF. 2001. Hypothalamic lesions in multiple sclerosis. J.Neuropathol.Exp.Neurol. 60(12):1208- 1218.

Hutton C, Draganski B, Ashburner J, Weiskopf N. 2009. A comparison between voxel- based cortical thickness and voxel-based morphometry in normal aging. NeuroImage 48(2):371-380.

Hutton C, De Vita E, Ashburner J, Deichmann R, Turner R. 2008. Voxel-based cortical thickness measurements in MRI. NeuroImage 40(4):1701-1710.

Im K, Lee JM, Lyttelton O, Kim SH, Evans AC, Kim SI. 2008. Brain size and cortical structure in the adult human brain. Cereb.Cortex 18(9):2181-2191.

172

Inglese M, Ge Y, Filippi M, Falini A, Grossman RI, Gonen O. 2004. Indirect evidence for early widespread gray matter involvement in relapsing-remitting multiple sclerosis. NeuroImage 21(4):1825-1829.

Jenkins TM, Ciccarelli O, Atzori M, Wheeler-Kingshott CA, Miller DH, Thompson AJ, Toosy AT. 2011. Early pericalcarine atrophy in acute optic neuritis is associated with conversion to multiple sclerosis. J.Neurol.Neurosurg.Psychiatry. .

Jones SE, Buchbinder BR, Aharon I. 2000. Three-dimensional mapping of cortical thickness using laplace's equation. Hum.Brain Mapp. 11:12-32.

Joshi S, Davis B, Jomier M, Gerig G. 2004. Unbiased diffeomorphic atlas construction for computational anatomy. NeuroImage 23:S151-S160.

Kandel, E.R., Schwartz, J.H., Jessell, T.M. 2000. Principles of neural science, 4th ed. McGraw-Hill, Nyew York.

Kangarlu A, Bourekas EC, Ray-Chaudhury A, Rammohan KW. 2007. Cerebral cortical lesions in multiple sclerosis detected by MR imaging at 8 tesla. AJNR Am.J.Neuroradiol. 28(2):262-266.

Kapeller P, McLean MA, Griffin CM, Chard D, Parker GJ, Barker GJ, Thompson AJ, Miller DH. 2001. Preliminary evidence for neuronal damage in cortical grey matter and normal appearing white matter in short duration relapsing-remitting multiple sclerosis: A quantitative MR spectroscopic imaging study. J.Neurol. 248(2):131-138.

Karacali B and Davatzikos C. 2006. Simulation of tissue atrophy using a topology preserving transformation model. IEEE Trans.Med.Imaging 25(5):649-652.

Kempton MJ, Ettinger U, Schmechtig A, Winter EM, Smith L, McMorris T, Wilkinson ID, Williams SC, Smith MS. 2009. Effects of acute dehydration on brain morphology in healthy humans. Hum.Brain Mapp. 30(1):291-298.

Kermode AG, Thompson AJ, Tofts P, MacManus DG, Kendall BE, Kingsley DP, Moseley IF, Rudge P, McDonald WI. 1990. Breakdown of the blood-brain barrier precedes symptoms and other MRI signs of new lesions in multiple sclerosis. pathogenetic and clinical implications. Brain 113 ( Pt 5)(Pt 5):1477-1489.

Khalil M, Enzinger C, Langkammer C, Petrovic K, Loitfelder M, Tscherner M, Jehna M, Bachmaier G, Wallner-Blazek M, Ropele S, Schmidt R, Fuchs S, Fazekas F. 2011. Cognitive impairment in relation to MRI metrics in patients with clinically isolated syndrome. Mult.Scler. 17(2):173-180.

Kidd D, Barkhof F, McConnell R, Algra P, Allen I, Revesz T. 1999. Cortical lesions in multiple sclerosis. Brain 122:17-26.

173

Kikinis R, Shenton ME, Iosifescu DV, McCarley RW, Saiviroonporn P, Hokama HH, Robatino A, Metcalf D, Wible CG, Portas CM, Donnino RM, Jolesz FA. 1996. A digital brain atlas for surgical planning, model-driven segmentation, and teaching. IEEE Trans. Vis. Comput. Graph. 2(3):232-241.

Kooi EJ, Geurts JJ, van Horssen J, Bo L, van der Valk P. 2009. Meningeal inflammation is not associated with cortical demyelination in chronic multiple sclerosis. J.Neuropathol.Exp.Neurol. 68(9):1021-1028.

Kremenchutzky M, Cottrell D, Rice G, Hader W, Baskerville J, Koopman W, Ebers GC. 1999. The natural history of multiple sclerosis: A geographically based study. 7. progressive-relapsing and relapsing-progressive multiple sclerosis: A re-evaluation. Brain 122 ( Pt 10)(Pt 10):1941-1950.

Kuperberg GR, Broome MR, McGuire PK, David AS, Eddy M, Ozawa F, Goff D, West WC, Williams SC, van der Kouwe AJ, Salat DH, Dale AM, Fischl B. 2003. Regionally localized thinning of the cerebral cortex in schizophrenia. Arch.Gen.Psychiatry 60(9):878-888.

Kurtzke JF. 1983. Rating neurologic impairment in multiple sclerosis: An expanded disability status scale (EDSS). Neurology 33(11):1444-1452.

Kutzelnigg A and Lassmann H. 2005. Cortical lesions and brain atrophy in MS. J.Neurol.Sci. 233(1-2):55-59.

Kutzelnigg A, Lucchinetti CF, Stadelmann C, Brück W, Rauschka H, Bergmann M, Schmidbauer M, Parisi JE, Lassmann H. 2005. Cortical demyelination and diffuse white matter injury in multiple sclerosis. Brain 128(Pt 11):2705-2712.

Kutzelnigg A, Faber-Rod JC, Bauer J, Lucchinetti CF, Sorensen PS, Laursen H, Stadelmann C, Bruck W, Rauschka H, Schmidbauer M, Lassmann H. 2007. Widespread demyelination in the cerebellar cortex in multiple sclerosis. Brain Pathol. 17(1):38-44.

Lai M, Hodgson T, Gawne-Cain M, Webb S, MacManus D, McDonald WI, Thompson AJ, Miller DH. 1996. A preliminary study into the sensitivity of disease activity detection by serial weekly magnetic resonance imaging in multiple sclerosis. J.Neurol.Neurosurg.Psychiatry. 60(3):339-341.

Lerch JP and Evans AC. 2005. Cortical thickness analysis examined through power analysis and a population simulation. NeuroImage 24(1):163-173.

Lerch JP, Pruessner JC, Zijdenbos A, Hampel H, Teipel SJ, Evans AC. 2005. Focal decline of cortical thickness in Alzheimer's disease identified by computational neuroanatomy. Cereb. Cortex 15(7):995-1001.

174

Lewis EB and Fox NC. 2004. Correction of differential intensity inhomogeneity in longitudinal MR images. NeuroImage 23(1):75-83.

Li DK, Held U, Petkau J, Daumer M, Barkhof F, Fazekas F, Frank JA, Kappos L, Miller DH, Simon JH, Wolinsky JS, Filippi M, Sylvia Lawry Centre for MS Research. 2006. MRI T2 lesion burden in multiple sclerosis: A plateauing relationship with clinical disability. Neurology 66(9):1384-1389.

Li Y, Wang Y, Xue Z, Shi F, Lin W, Shen D, Alzheimer's Disease Neuroimaging Initiative. 2010. Consistent 4D cortical thickness measurement for longitudinal neuroimaging study. MICCAI 13(Pt 2):133-142.

Liu C, Edwards S, Gong Q, Roberts N, Blumhardt LD. 1999. Three dimensional MRI estimates of brain and spinal cord atrophy in multiple sclerosis. J.Neurol.Neurosurg.Psychiatry. 66(3):323-330.

Loevner LA, Grossman RI, Cohen JA, Lexa FJ, Kessler D, Kolson DL. 1995. Microscopic disease in normal-appearing white matter on conventional MR images in patients with multiple sclerosis: Assessment with magnetization-transfer measurements. Radiology 196(2):511-515.

Lohmann G, Preul C, Hund-Georgiadis M. 2003. Morphology-based cortical thickness estimation. Inf.Process.Med.Imaging 18:89-100.

Lorensen WE and Cline HE. 1987. Marching cubes: A high resolution 3D surface construction algorithm. SIGGRAPH Comput.Graph. 21(4):163-169.

Losseff NA, Wang L, Lai HM, Yoo DS, Gawne-Cain ML, McDonald WI, Miller DH, Thompson AJ. 1996. Progressive cerebral atrophy in multiple sclerosis. A serial MRI study. Brain 119 ( Pt 6)(Pt 6):2009-2019.

Lublin FD and Reingold SC. 1996. Defining the clinical course of multiple sclerosis: Results of an international survey. national multiple sclerosis society (USA) advisory committee on clinical trials of new agents in multiple sclerosis. Neurology 46(4):907- 911.

MacDonald D, Kabani N, Avis D, Evans AC. 2000. Automated 3-D extraction of inner and outer surfaces of cerebral cortex from MRI. NeuroImage 12(3):340-356.

Magliozzi R, Howell O, Vora A, Serafini B, Nicholas R, Puopolo M, Reynolds R, Aloisi F. 2007. Meningeal B-cell follicles in secondary progressive multiple sclerosis associate with early onset of disease and severe cortical pathology. Brain 130(Pt 4):1089-1104.

175

Mammi S, Filippi M, Martinelli V, Campi A, Colombo B, Scotti G, Canal N, Comi G. 1996. Correlation between brain MRI lesion volume and disability in patients with multiple sclerosis. Acta Neurol.Scand. 94(2):93-96.

Mann DM. 1991. The topographic distribution of brain atrophy in Alzheimer's disease. Acta Neuropathol. 83(1):81-86.

Marcus DS, Wang TH, Parker J, Csernansky JG, Morris JC, Buckner RL. 2007. Open access series of imaging studies (OASIS): Cross-sectional MRI data in young, middle aged, nondemented, and demented older adults. J. Cogn. Neurosci. 19(9):1498-1507.

Marroquin JL, Vemuri BC, Botello S, Calderon F, Fernandez-Bouzas A. 2002. An accurate and efficient bayesian method for automatic segmentation of brain MRI. IEEE Trans.Med.Imaging 21(8):934-945.

Memoli F, Sapiro G, Thompson P. 2004. Implicit brain imaging. NeuroImage 23 Suppl 1:S179-88.

Miller DH and Leary SM. 2007. Primary-progressive multiple sclerosis. Lancet Neurol. 6(10):903-912.

Miller DH, Rudge P, Johnson G, Kendall BE, Macmanus DG, Moseley IF, Barnes D, McDonald WI. 1988. Serial gadolinium enhanced magnetic resonance imaging in multiple sclerosis. Brain 111 ( Pt 4)(Pt 4):927-939.

Miller DH, Soon D, Fernando KT, MacManus DG, Barker GJ, Yousry TA, Fisher E, O'Connor PW, Phillips JT, Polman CH, Kappos L, Hutchinson M, Havrdova E, Lublin FD, Giovannoni G, Wajgt A, Rudick R, Lynn F, Panzara MA, Sandrock AW, For the AFFIRM Investigators . 2007. MRI outcomes in a placebo-controlled trial of natalizumab in relapsing MS. Neurology 68(17):1390-1401.

Miller MI, Massie AB, Ratnanather JT, Botteron KN, Csernansky JG. 2000. Bayesian construction of geometrically based cortical thickness metrics. NeuroImage 12:676- 687.

Mohamed FB, Vinitski S, Faro SH, Gonzalez CF, Mack J, Iwanaga T. 1999. Optimization of tissue segmentation of brain MR images based on multispectral 3D feature maps. Magn.Reson.Imaging 17:403-409.

Moller T. 1997. A fast triangle-triangle intersection test. J.Graph.Tools 2(2):25-30.

Morgen K, Sammer G, Courtney SM, Wolters T, Melchior H, Blecker CR, Oschmann P, Kaps M, Vaitl D. 2006. Evidence for a direct association between cortical atrophy and cognitive impairment in relapsing-remitting MS. NeuroImage 30:891-898.

176

Nakamura K, Fox R, Fisher E. 2011. CLADA: Cortical longitudinal atrophy detection algorithm. NeuroImage 54(1):278-289.

Nakamura K, Rudick RA, Lee JC, Foulds P, Fisher E. 2010. Effect of intramuscular interferon beta-1a on gray matter atrophy in relapsing-remitting multiple sclerosis. Presented at the Annual Meeting of the American Academy of Neurology .

Nakamura K and Fisher E. 2009. Segmentation of brain magnetic resonance images for measurement of gray matter atrophy in multiple sclerosis patients. NeuroImage 44(3):769-776.

Narayana PA, Doyle TJ, Lai D, Wolinsky JS. 1998. Serial proton magnetic resonance spectroscopic imaging, contrast-enhanced magnetic resonance imaging, and quantitative lesion volumetry in multiple sclerosis. Ann.Neurol. 43(1):56-71.

National Multiple Sclerosis Society. 2011. About MS. .

Nelson F, Poonawalla A, Hou P, Wolinsky JS, Narayana PA. 2008. 3D MPRAGE improves classification of cortical lesions in multiple sclerosis. Mult.Scler. 14(9):1214-1219.

Nelson F, Poonawalla AH, Hou P, Huang F, Wolinsky JS, Narayana PA. 2007. Improved identification of intracortical lesions in multiple sclerosis with phase-sensitive inversion recovery in combination with fast double inversion recovery MRI. AJNR Am J Neuroradiol .

Nesbit GM, Forbes GS, Scheithauer BW, Okazaki H, Rodriguez M. 1991. Multiple sclerosis: Histopathologic and MR and/or CT correlation in 37 cases at biopsy and three cases at autopsy. Radiology 180(2):467-474.

Newcombe J, Hawkins CP, Henderson CL, Patel HA, Woodroofe MN, Hayes GM, Cuzner ML, MacManus D, Boulay EPGHD, McDonald WI. 1991. Histopathology of multiple sclerosis lesions detected by magnetic resonance imaging in unfixed postmortem central nervous system tissue. Brain 114(2):1013-1023.

Noseworthy JH, Lucchinetti C, Rodriguez M, Weinshenker BG. 2000. Multiple sclerosis. N Engl J Med 343(13):938-952.

Oreja-Guevara C, Rovaris M, Iannucci G, Valsasina P, Caputo D, Cavarretta R, Sormani MP, Ferrante P, Comi G, Filippi M. 2005. Progressive gray matter damage in patients with relapsing-remitting multiple sclerosis: A longitudinal diffusion tensor magnetic resonance imaging study. Arch.Neurol. 62(4):578-584.

Oreja-Guevara C, Charil A, Caputo D, Cavarretta R, Sormani MP, Filippi M. 2006. Magnetization transfer magnetic resonance imaging and clinical changes in patients with relapsing-remitting multiple sclerosis. Arch Neurol 63(5):736-740.

177

Papadopoulos D, Dukes S, Patel R, Nicholas R, Vora A, Reynolds R. 2009. Substantial archaeocortical atrophy and neuronal loss in multiple sclerosis. Brain Pathol. 19(2):238-253.

Pellicano C, Gallo A, Li X, Ikonomidou VN, Evangelou IE, Ohayon JM, Stern SK, Ehrmantraut M, Cantor F, McFarland HF, Bagnato F. 2010. Relationship of cortical atrophy to fatigue in patients with multiple sclerosis. Arch.Neurol. 67(4):447-453.

Perona P and Malik J. 1990. Scale-space and edge detection using anisotropic diffusion. IEEE Trans. Pattern. Anal. Mach. Intell. 12(7):629-639.

Peterson JW, Bö L, Mörk S, Chang A, Trapp BD. 2001. Transected neurites, apoptotic neurons, and reduced inflammation in cortical multiple sclerosis lesions. Ann.Neurol. 50:389-400.

Pham DLP, J.L. 1999. Adaptive fuzzy segmentation of magnetic resonance images. IEEE Trans.Med.Imaging 18(9):737-752.

Pierpaoli C, Jezzard P, Basser PJ, Barnett A, Di Chiro G. 1996. Diffusion tensor MR imaging of the human brain. Radiology 201(3):637-648.

Pirko I, Lucchinetti CF, Sriram S, Bakshi R. 2007. Gray matter involvement in multiple sclerosis. Neurology 68(9):634-642.

Pluim JP, Maintz JB, Viergever MA. 2003. Mutual-information-based registration of medical images: A survey. IEEE Trans.Med.Imaging 22(8):986-1004.

Polman CH and Rudick RA. 2010. The multiple sclerosis functional composite: A clinically meaningful measure of disability. Neurology 74 Suppl 3:S8-15.

Polman CH, Reingold SC, Edan G, Filippi M, Hartung HP, Kappos L, Lublin FD, Metz LM, McFarland HF, O'Connor PW, Sandberg-Wollheim M, Thompson AJ, Weinshenker BG, Wolinsky JS. 2005. Diagnostic criteria for multiple sclerosis: 2005 revisions to the "McDonald criteria". Ann.Neurol. 58(6):840-846.

Pomeroy IM, Jordan EK, Frank JA, Matthews PM, Esiri MM. 2010. Focal and diffuse cortical degenerative changes in a marmoset model of multiple sclerosis. Mult.Scler. 16(5):537-548.

Pomeroy IM, Jordan EK, Frank JA, Matthews PM, Esiri MM. 2008. Diffuse cortical atrophy in a marmoset model of multiple sclerosis. Neurosci.Lett. 437(2):121-124.

Portaccio E, Amato MP, Bartolozzi ML, Zipoli V, Mortilla M, Guidi L, Siracusa G, Sorbi S, Federico A, De Stefano N. 2006. Neocortical volume decrease in relapsing- remitting multiple sclerosis with mild cognitive impairment. J.Neurol.Sci. 245(1- 2):195-199.

178

Poser CM, Paty DW, Scheinberg L, McDonald WI, Davis FA, Ebers GC, Johnson KP, Sibley WA, Silberberg DH, Tourtellotte WW. 1983. New diagnostic criteria for multiple sclerosis: Guidelines for research protocols. Ann.Neurol. 13(3):227-231.

Press, W.H., Teukolsky, S.A., Vetterling, W.T., Flannery, B.P. 2002. Numerical recipes in C: The art of scientific computing, 2nd ed. Cambridge University Press, Cambridge.

Prineas JW and Connell F. 1979. Remyelination in multiple sclerosis. Ann.Neurol. 5(1):22-31.

Prinster A, Quarantelli M, Orefice G, Lanzillo R, Brunetti A, Mollica C, Salvatore E, Morra VB, Coppola G, Vacca G, Alfano B, Salvatore M. 2006. Grey matter loss in relapsing-remitting multiple sclerosis: A voxel-based morphometry study. NeuroImage 29(3):859-867.

Quarantelli M, Ciarmiello A, Morra VB, Orefice G, Larobina M, Lanzillo R, Schiavone V, Salvatore E, Alfano B, Brunetti A. 2003. Brain tissue volume changes in relapsing-remitting multiple sclerosis: Correlation with lesion load. NeuroImage 18(2):360-366.

Ramasamy DP, Benedict RH, Cox JL, Fritz D, Abdelrahman N, Hussein S, Minagar A, Dwyer MG, Zivadinov R. 2009. Extent of cerebellum, subcortical and cortical atrophy in patients with MS: A case-control study. J.Neurol.Sci. 282(1-2):47-54.

Resnick SM, Pham DL, Kraut MA, Zonderman AB, Davatzikos C. 2003. Longitudinal magnetic resonance imaging studies of older adults: A shrinking brain. The Journal of Neuroscience 23(8):3295-3301.

Reuter M and Fischl B. 2011. Avoiding asymmetry-induced bias in longitudinal image processing. NeuroImage 57(1):19-21.

Ropele S and Fazekas F. 2009. Magnetization transfer MR imaging in multiple sclerosis. Neuroimaging Clin.N.Am. 19(1):27-36.

Rosas HD, Liu AK, Hersch S, Glessner M, Ferrante RJ, Salat DH, van der Kouwe A, Jenkins BG, Dale AM, Fischl B. 2002. Regional and progressive thinning of the cortical ribbon in huntington's disease. Neurology 58(5):695-701.

Rudick R, Antel J, Confavreux C, Cutter G, Ellison G, Fischer J, Lublin F, Miller A, Petkau J, Rao S, Reingold S, Syndulko K, Thompson A, Wallenberg J, Weinshenker B, Willoughby E. 1997. Recommendations from the national multiple sclerosis society clinical outcomes assessment task force. Ann.Neurol. 42(3):379-382.

Rudick RA. 2004. Impact of disease-modifying therapies on brain and spinal cord atrophy in multiple sclerosis. J.Neuroimaging 14(3 Suppl):54S-64S.

179

Rudick RA, Lee JC, Nakamura K, Fisher E. 2009. Gray matter atrophy correlates with MS disability progression measured with MSFC but not EDSS. J.Neurol.Sci. 282(1- 2):106-111.

Rudick RA, Lee JC, Simon J, Ransohoff RM, Fisher E. 2004. Defining interferon beta response status in multiple sclerosis patients. Ann.Neurol. 56(4):548-555.

Rudick RA, Fisher E, Lee JC, Duda JT, Simon J. 2000. Brain atrophy in relapsing multiple sclerosis: Relationship to relapses, EDSS, and treatment with interferon beta- 1a. Mult.Scler. 6(6):365-372.

Rudick RA, Fisher E, Lee JC, Simon J, Jacobs L. 1999. Use of the brain parenchymal fraction to measure whole brain atrophy in relapsing-remitting MS. multiple sclerosis collaborative research group. Neurology 53(8):1698-1704.

Rudick RA, Cutter G, Baier M, Fisher E, Dougherty D, Weinstock-Guttman B, Mass MK, Miller D, Simonian NA. 2001. Use of the multiple sclerosis functional composite to predict disability in relapsing MS. Neurology 56(10):1324-1330.

Rueda A, Acosta O, Couprie M, Bourgeat P, Fripp J, Dowson N, Romero E, Salvado O. 2010. Topology-corrected segmentation and local intensity estimates for improved partial volume classification of brain cortex in MRI. J.Neurosci.Methods 188(2):305- 315.

Sailer M, Fischl B, Salat D, Tempelmann C, Schonfeld MA, Busa E, Bodammer N, Heinze HJ, Dale A. 2003. Focal thinning of the cerebral cortex in multiple sclerosis. Brain 126(Pt 8):1734-1744.

Sajja BR, Datta S, He R, Mehta M, Gupta RK, Wolinsky JS, Narayana PA. 2006. Unified approach for multiple sclerosis lesion segmentation on brain MRI. Ann.Biomed.Eng. 34(1):142-151.

Salat DH, Buckner RL, Snyder AZ, Greve DN, Desikan RSR, Busa E, Morris JC, Dale AM, Fischl B. 2004. Thinning of the cerebral cortex in aging. Cerebral Cortex 14:721-730.

Sampat M, Nguyen C, Okuda DT, Henry RG, Pelleter D. 2010. Cortical thickness in MS as long term predictor of physical impairment: Results from a longitudinal and observational large study. Presented at the Annual Meeting of the American Academy of Neurology .

Sanfilipo MP, Benedict RH, Sharma J, Weinstock-Guttman B, Bakshi R. 2005. The relationship between whole brain volume and disability in multiple sclerosis: A comparison of normalized gray vs. white matter with misclassification correction. NeuroImage 26(4):1068-1077.

180

Santago P and Gage HD. 1995. Statistical models of partial volume effect. IEEE Trans.Image Process. 4(11):1531-1540.

Sastre-Garriga J, Ingle GT, Chard DT, Ramio-Torrenta L, Miller DH, Thompson AJ. 2004. Grey and white matter atrophy in early clinical stages of primary progressive multiple sclerosis. NeuroImage 22(1):353-359.

Sastre-Garriga J, Ingle GT, Chard DT, Cercignani M, Ramio-Torrenta L, Miller DH, Thompson AJ. 2005. Grey and white matter volume changes in early primary progressive multiple sclerosis: A longitudinal study. Brain 128(Pt 6):1454-1460.

Scalfari A, Neuhaus A, Degenhardt A, Rice GP, Muraro PA, Daumer M, Ebers GC. 2010. The natural history of multiple sclerosis, a geographically based study 10: Relapses and long-term disability. Brain 133(Pt 7):1914-1929.

Schmierer K, Parkes HG, So PW, An SF, Brandner S, Ordidge RJ, Yousry TA, Miller DH. 2010. High field (9.4 tesla) magnetic resonance imaging of cortical grey matter lesions in multiple sclerosis. Brain 133(Pt 3):858-867.

Schmierer K, Scaravilli F, Altmann DR, Barker GJ, Miller DH. 2004. Magnetization transfer ratio and myelin in postmortem multiple sclerosis brain. Ann.Neurol. 56(3):407-415.

Schnack HG, Hulshoff Pol HE, Baaré WFC, Staal WG, Viergever MA, Kahn RS. 2001. Automated separation of gray and white matter from MR images of the human brain. NeuroImage 13:230-237.

Scott ML, Bromiley PA, Thacker NA, Hutchinson CE, Jackson A. 2009. A fast, model- independent method for cerebral cortical thickness estimation using MRI. Med.Image Anal. 13(2):269-285.

Seewann A, Kooi EJ, Roosendaal S, Pouwels PJW, van der Valk P, Barkhof F, Polman C, Geurts JG. 2010. How reliable is 3D double inversion recovery in the detection of cortical lesions? combined postmortem MR-imaging and histopathology. ECTRIMS 2010 .

Sepulcre J, Sastre-Garriga J, Cercignani M, Ingle GT, Miller DH, Thompson AJ. 2006. Regional gray matter atrophy in early primary progressive multiple sclerosis: A voxel-based morphometry study. Arch.Neurol. 63(8):1175-1180.

Sepulcre J, Goni J, Masdeu JC, Bejarano B, Velez de Mendizabal N, Toledo JB, Villoslada P. 2009. Contribution of white matter lesions to gray matter atrophy in multiple sclerosis: Evidence from voxel-based analysis of T1 lesions in the visual pathway. Arch.Neurol. 66(2):173-179.

181

Shah A, Eggenberger E, Zivadinov R, Stuve O, Frohman EM. 2007. Corticosteroids for multiple sclerosis: II. application for disease-modifying effects. Neurotherapeutics 4(4):627-632.

Sharma R, Narayana PA, Wolinsky JS. 2001. Grey matter abnormalities in multiple sclerosis: Proton magnetic resonance spectroscopic imaging. Multiple Sclerosis 7(4):221-226.

Shattuck DW, Sandor-Leahy SR, Schaper KA, Rottenberg DA, Leahy RM. 2001. Magnetic resonance image tissue classification using a partial volume model. NeuroImage 13:856-876.

Sled JG, Zijdenbos AP, Evans AC. 1998. A nonparametric method for automatic correction of intensitynonuniformity in MRI data. IEEE Trans.Med.Imaging 18(1):87-97.

Smith SM, Rao A, De Stefano N, Jenkinson M, Schott JM, Matthews PM, Fox NC. 2007. Longitudinal and cross-sectional analysis of atrophy in Alzheimer's disease: Cross- validation of BSI, SIENA and SIENAX. NeuroImage 36(4):1200-1206.

Smith SM, Zhang Y, Jenkinson M, Chen J, Matthews PM, Federico A, De Stefano N. 2002. Accurate, robust, and automated longitudinal and cross-sectional brain change analysis. NeuroImage 17:479-489.

Smith SM, Jenkinson M, Woolrich MW, Beckmann CF, Behrens TEJ, Johansen-Berg H, Bannister PR, De Luca M, Drobnjak I, Flitney DE, Niazy RK, Saunders J, Vickers J, Zhang Y, De Stefano N, Brady JM, Matthews PM. 2004. Advances in functional and structural MR image analysis and implementation as FSL. NeuroImage 23:S208-S219.

Srinivasan R, Sailasuta N, Hurd R, Nelson S, Pelletier D. 2005. Evidence of elevated glutamate in multiple sclerosis using magnetic resonance spectroscopy at 3 T. Brain 128(Pt 5):1016-1025.

Srivastava, S., Maes, F., Vandermeulen, D., Dupont, P., Van Paesschen, W., Suetens, P. 2003. An automated 3D algorithm for neo-cortical thickness measurement. Medical Image Computing and Computer-Assisted Intervention - MICCAI 2003 2879, 488- 495.

Stadelmann C, Albert M, Wegner C, Bruck W. 2008. Cortical pathology in multiple sclerosis. Curr.Opin.Neurol. 21(3):229-234.

Stewart WA, Hall LD, Berry K, Paty DW. 1984. Correlation between NMR scan and brain slice data in multiple sclerosis. Lancet 2(8399):412.

Styner M, Brechbuhler C, Szekely G, Gerig G. 2000. Parametric estimate of intensity inhomogeneities applied to MRI. IEEE Trans.Med.Imaging 19(3):153-165.

182

Suckling J, Sigmundsson T, Greenwood K, Bullmore ET. 1999. A modified fuzzy clustering algorithm for operator independent brain tissue classification of dual echo MR images. Magn.Reson.Imaging 17:1065-1076.

Tallantyre EC, Morgan PS, Dixon JE, Al-Radaideh A, Brookes MJ, Morris PG, Evangelou N. 2010. 3 tesla and 7 tesla MRI of multiple sclerosis cortical lesions. J.Magn.Reson.Imaging 32(4):971-977.

Tedeschi G, Dinacci D, Comerci M, Lavorgna L, Savettieri G, Quattrone A, Livrea P, Patti F, Brescia Morra V, Servillo G, Orefice G, Paciello M, Prinster A, Coniglio G, Bonavita S, Di Costanzo A, Bellacosa A, Valentino P, Quarantelli M, Brunetti A, Salemi G, D’Amelio M, Simone I, Salvatore M, Bonavita V, Alfano B. 2009. Brain atrophy evolution and lesion load accrual in multiple sclerosis: A 2-year follow-up study. Multiple Sclerosis 15(2):204-211.

Tedeschi G, Lavorgna L, Russo P, Prinster A, Dinacci D, Savettieri G, Quattrone A, Livrea P, Messina C, Reggio A, Bresciamorra V, Orefice G, Paciello M, Brunetti A, Coniglio G, Bonavita S, Di Costanzo A, Bellacosa A, Valentino P, Quarantelli M, Patti F, Salemi G, Cammarata E, Simone IL, Salvatore M, Bonavita V, Alfano B. 2005. Brain atrophy and lesion load in a large population of patients with multiple sclerosis. Neurology 65(2):280-285.

Teo PC, Sapiro G, Wandell BA. 1997. Creating connected representations of cortical gray matter for functional MRI visualization. IEEE Trans.Med.Imaging 16(6):852- 863.

Thambisetty M, Wan J, Carass A, An Y, Prince JL, Resnick SM. 2010. Longitudinal changes in cortical thickness associated with normal aging. NeuroImage 52(4):1215- 1223.

Thompson PM, Hayashi KM, de Zubicaray G, Janke AL, Rose SE, Semple J, Herman D, Hong MS, Dittmer SS, Doddrell DM, Toga AW. 2003. Dynamics of gray matter loss in Alzheimer's disease. J. Neurosci. 23(3):994-1005.

Tiberio M, Chard DT, Altmann DR, Davies G, Griffin CM, Rashid W, Sastre-Garriga J, Thompson AJ, Miller DH. 2005. Gray and white matter volume changes in early RRMS: A 2-year longitudinal study. Neurology 64(6):1001-1007.

Trapp BD and Stys PK. 2009. Virtual hypoxia and chronic necrosis of demyelinated axons in multiple sclerosis. Lancet Neurol. 8(3):280-291.

Trapp BD, Ransohoff R, Rudick R. 1999a. Axonal pathology in multiple sclerosis: Relationship to neurologic disability. Curr.Opin.Neurol. 12(3):295-302.

Trapp BD, Ransohoff RM, Fisher E, Rudick RA. 1999b. Neurodegeneration in multiple sclerosis: Relationship to neurological disability. The Neuroscientist 5(1):48-57.

183

Trapp BD, Peterson J, Ransohoff RM, Rudick R, Mork S, Bo L. 1998. Axonal transection in the lesions of multiple sclerosis. N Engl J Med 338(5):278-285.

Valsasina P, Benedetti B, Rovaris M, Sormani MP, Comi G, Filippi M. 2005. Evidence for progressive gray matter loss in patients with relapsing-remitting MS. Neurology 65(7):1126-1128. van den Elskamp IJ, Knol DL, Vrenken H, Karas G, Meijerman A, Filippi M, Kappos L, Fazekas F, Wagner K, Pohl C, Sandbrink R, Polman CH, Uitdehaag BM, Barkhof F. 2010. Lesional magnetization transfer ratio: A feasible outcome for remyelinating treatment trials in multiple sclerosis. Mult.Scler. 16(6):660-669. van der Kouwe AJ, Benner T, Salat DH, Fischl B. 2008. Brain morphometry with multiecho MPRAGE. NeuroImage 40(2):559-569. van Horssen J, Brink BP, de Vries HE, van der Valk P, Bo L. 2007. The blood-brain barrier in cortical multiple sclerosis lesions. J.Neuropathol.Exp.Neurol. 66(4):321- 328.

Van Leemput K, Maes F, Vandermeulen D, Suetens P. 1999. Automated model-based tissue classification of MR images of the brain. IEEE Trans.Med.Imaging 18(10):897-908. van Walderveen MA, Kamphorst W, Scheltens P, van Waesberghe JH, Ravid R, Valk J, Polman CH, Barkhof F. 1998. Histopathologic correlate of hypointense lesions on T1-weighted spin-echo MRI in multiple sclerosis. Neurology 50(5):1282-1288.

Vercellino M, Plano F, Votta B, Mutani R, Giordana MT, Cavalla P. 2005. Grey matter pathology in multiple sclerosis. J.Neuropathol.Exp.Neurol. 64(12):1101-1107.

Wattjes MP, Harzheim M, Lutterbey GG, Klotz L, Schild HH, Traber F. 2007. Axonal damage but no increased glial cell activity in the normal-appearing white matter of patients with clinically isolated syndromes suggestive of multiple sclerosis using high-field magnetic resonance spectroscopy. AJNR Am.J.Neuroradiol. 28(8):1517- 1522.

Wegner C, Esiri MM, Chance SA, Palace J, Matthews PM. 2006. Neocortical neuronal, synaptic, and glial loss in multiple sclerosis. Neurology 67(6):960-967.

Weinshenker BG. 1996. Epidemiology of multiple sclerosis. Neurol.Clin. 14(2):291-308.

Weinshenker BG, Bass B, Rice GP, Noseworthy J, Carriere W, Baskerville J, Ebers GC. 1989. The natural history of multiple sclerosis: A geographically based study. I. clinical course and disability. Brain 112 ( Pt 1)(Pt 1):133-146.

184

Whetten-Goldstein K, Sloan FA, Goldstein LB, Kulas ED. 1998. A comprehensive assessment of the cost of multiple sclerosis in the United States. Mult.Scler. 4(5):419- 425.

Wiley CA, Masliah E, Morey M, Lemere C, DeTeresa R, Grafe M, Hansen L, Terry R. 1991. Neocortical damage during HIV infection. Ann.Neurol. 29(6):651-657.

Winkler AM, Kochunov P, Blangero J, Almasy L, Zilles K, Fox PT, Duggirala R, Glahn DC. 2009. Cortical thickness or grey matter volume? the importance of selecting the phenotype for imaging genetics studies. NeuroImage .

Wolff SD and Balaban RS. 1994. Magnetization transfer imaging: Practical aspects and clinical applications. Radiology 192(3):593-599.

Xu C and Prince JL. 1998. Snakes, shapes, and gradient vector flow. IEEE Transactions on Image Processing 7(3):359-369.

Xu C, Pham DL, Rettmann ME, Yu DN, Prince JL. 1999. Reconstruction of the human cerebral cortex from magnetic resonance images. IEEE Trans.Med.Imaging 18(6):467-480.

Yezzi AJ,Jr and Prince JL. 2003. An eulerian PDE approach for computing tissue thickness. IEEE Trans.Med.Imaging 22(10):1332-1339.

Yushkevich PA, Avants BB, Das SR, Pluta J, Altinay M, Craige C, Alzheimer's Disease Neuroimaging Initiative. 2010. Bias in estimation of hippocampal atrophy using deformation-based morphometry arises from asymmetric global normalization: An illustration in ADNI 3 T MRI data. NeuroImage 50(2):434-445.

Zeng X, Staib LH, Schultz RT, Duncan JS. 1999. Segmentation and measurement of the cortex from 3-D MR images using coupled-surfaces propagation. IEEE Trans.Med.Imaging 18(10):927-937.

Zhang Y, Brady M, Smith S. 2001. Segmentation of brain MR images through a hidden markov random field model and the expectation-maximization algorithm. IEEE Trans.Med.Imaging 20(1):45-57.

Zhou Y and Bai J. 2007. Atlas-based fuzzy connectedness segmentation and intensity nonuniformity correction applied to brain MRI. Biomedical Engineering, IEEE Transactions on 54(1):122-129.

Zhu C and Jiang T. 2003. Multicontext fuzzy clustering for separation of brain tissues in magnetic resonance images. NeuroImage 18(3):685-696.

Ziemssen T. 2009. Multiple sclerosis beyond EDSS: Depression and fatigue. J.Neurol.Sci. 277 Suppl 1:S37-41.

185

Zijdenbos AP, Dawant BM, Margolin RA, Palmer AC. 1994. Morphometric analysis of white matter lesions in MR images: Method and validation. IEEE Trans.Med.Imaging 13(4):716-724.

Zivadinov R, Locatelli L, Cookfair D, Srinivasaraghavan B, Bertolotto A, Ukmar M, Bratina A, Maggiore C, Bosco A, Grop A, Catalan M, Zorzon M. 2007. I nterferon beta-1a slows progression of brain atrophy in relapsing-remitting multiple sclerosis predominantly by reducing gray matter atrophy. Multiple Sclerosis 13(4):490-501.

Zomorodian A and Edelsbrunner H. 2000. Fast software for box intersections. Proceedings of the 16th Annual Symposium on Computational Geometry :129-138.

186