UNIVERSITY OF CALGARY. A New Method for Assessing Tissue Alignment using Clinical MRI in Multiple Sclerosis. Shrushrita Sharma A THESIS

Transcription

1 UNIVERSITY OF CALGARY A New Method for Assessing Tissue Alignment using Clinical MRI in Multiple Sclerosis by Shrushrita Sharma A THESIS SUBMITTED TO THE FACULTY OF GRADUATE STUDIES IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE GRADUATE PROGRAM IN BIOMEDICAL ENGINEERING CALGARY, ALBERTA JUNE, 2017 Shrushrita Sharma 2017

2 Abstract Changes in the alignment of white matter tracts are common in many neurological disorders such as multiple sclerosis (MS). Currently advanced MRI methods including diffusion-weighted imaging is the mainstay in assessing tissue coherency and anisotropy. In this thesis, I have implemented and verified a novel image-processing method for this purpose using conventional MRI. This is done based on Fourier transform power spectrum. Outcomes were evaluated in 3 steps: 1) testing feasibility using brain areas with highly aligned nerve fiber tracks in T2- weighted MRI; 2) confirming pathological relevance using postmortem brain sample; and 3) assessing utility by comparing with diffusion tensor imaging. To improve the accuracy of comparison with pathology, I have also conducted quantitative histology besides traditional analysis of the staining density of myelin and axons. The results suggest that advanced analysis of clinical MRI may provide valuable information as powerful as advanced MRI to enhance the measurement of tissue property. ii

3 Acknowledgement Firstly, I want to express my sincere gratitude to Dr. Yunyan Zhang for taking me under supervision to pursue my masters in Biomedical Engineering. I have found continuous guidance, supervision, motivation, support, and strong encouragement throughout my graduate study. Thank you so much for your time, patience, and effort which has helped me to handle my graduate education efficiently. I would like to thank my supervisory committee Dr. Bruce Pike, Dr. Katayoun Alikhani and Dr. Jeff Dunn for your time and support. I am also thankful to Biomedical Engineering Graduate program for providing me a chance to learn and grow. Thank you to my lab members Peng Zhai and Glenn Pridham for helping me with my research. Finally, I would like to thank my family and friends who have made my journey at University of Calgary such an enjoyable experience. - Shrushrita Sharma iii

4 Table of Contents Abstract... ii Acknowledgement... iii List of Tables... vii List of Figures... viii List of Abbreviations... xii CHAPTER 1: INTRODUCTION Overview of research Hypothesis and specific aims Thesis organization... 3 CHAPTER 2: LITERATURE REVIEW Overview of MS MRI of MS Conventional MRI methodologies Advanced MRI methodologies Image post-processing and analysis methodologies Spatial frequency analysis Structure Tensor Analysis Statistical-based image analysis Summary CHAPTER 3: FOURIER TRANSFORM POWER SPECTRUM IS A POTENTIAL MEASURE OF TISSUE ALIGNMENT IN STANDARD MRI Introduction Materials and Methods Subjects MRI acquisition Selection of image and regions of interest Analysis of tissue alignment iv

5 3.2.5 Analysis of alignment complexity Statistical Analysis Results Sample Characteristics Correspondence of dominant tissue alignment between quantified and observed angles Increased tissue angular entropy in patients as compared to controls Discussion Conclusions Acknowledgement CHAPTER 4: VALIDATION OF FT POWER SPECTRUM-BASED METHOD WITH STRUCTURE TENSOR ANALYSIS Introduction Materials and methods Sample MR imaging Protocol Directional measurement in MRI Histological preparation Directional measurement in histology Statistical analysis Results Sample characteristics Differences in orientation metrics between tissue types in MRI and histology Correlation between MRI and histology in dominant orientation Correlation between MRI and histology in angular entropy Discussion Conclusion Acknowledgement CHAPTER 5: COMPARISON OF FT POWER SPECTRUM-BASED METHOD WITH DIFFUSION TENSOR IMAGING v

6 5.1 Introduction Materials and methods Subjects and samples MR acquisition DTI analysis Tissue integrity measurement in conventional images Statistical analysis Results Sample characteristics Correspondence between FT power spectrum angular entropy and FA Correspondence between dominant orientation from FT power spectrum and principle direction of DTI Correspondence between FT-based orientation strength and dominant diffusivity Discussion Conclusion Acknowledgement CHAPTER 6: SUMMARY, CONCLUSION AND FUTURE DIRECTION Summary of the Research Limitations of the thesis Future Work Conclusions and significance References vi

7 List of Tables Table 3.1 : The number of regions of interest examined in the corpus callosum per group Table 4.1 : The number of ROIs examined in each tissue type per sample Table 4.2 : The mean (standard error) of orientation outcomes per tissue type in MRI and histology Table 4.3 : Correlation coefficients in dominant orientation between MRI and histology Table 5.1 : Number of ROIs analyzed in the corpus callosum Table 5.2 : Mean (± standard error) FA in DTI and angular entropy in T2-weighted MRI Table 5.3 : Mean (± standard error) highest diffusivity L1 and orientation strength vii

8 List of Figures Fig. 2.1: Demyelination in MS. The top panel shows a healthy neuron with intact myelin sheath and the bottom panel shows a neuron with disrupted myelin sheath Fig. 2.2: Lesion presentation in different MR images in a MS patient. Axial proton-density weighted (a,d), T2-weighted (b,e) and FLAIR (c,f) consecutive images of a patient with MS demonstrate multiple lesions around lateral ventricles as shown by arrows. Figure derived from Sahraian M. et al., MRI Atlas of MS lesions, 2008, p Fig. 2.3: Different outcomes from DTI. Left column shows non-diffusion weighted scan followed by FA, MD in color coded form. Figure derived from Winston GP., Quant Imaging Med Surg., 2012, p Fig. 2.4: (a) Three different coherent anti-stokes Raman scattering images of mouse spinal cord surface in the longitudinal orientation. From top to bottom, Experimental Autoimmune Encephalomyelitis lesions progressively afflict the tissue. (b) Processed 2D-FT of the domains. (c) Segmented objects with superposed equivalent ellipse and main axes. (d) Average fiber orientation overlaid on images with aspect ratio (see color code on the right). Figure derived from Begin et al., Biomed Opt Express, 2013, p Fig. 2.5: Flowchart of structure tensor analysis with associated outcomes. Left column shows an example natural image. Middle column shows the orientation, coherency and energy maps for studying the directional properties of the image. Color-coded map on the top right shows the orientation features of the original image and the circular histogram on the bottom right is another way of representating orientation information. Figure derived from Rezakhaniha et al., Biomech Model Mecahnobiol., 2012, p Fig. 3.1: Method demonstration. Shown is an example ROI in the right genu of corpus callosum (A, small box); Zoomed view of the ROI (B); and FT of the ROI (C). FT power spectrum of the ROI is shown after normalization (D) and thresholding (E). Based on E, the orientation profile of the ROI is calculated (F), from which corresponding angular entropy can be computed. Figure derived from Sharma S. and Zhang Y., PLoS one, 2017, p Fig. 3.2: Representative ROIs along with predicted major aligning orientations based on the anatomy of corpus callosum. Panel A shows an example T2-weighted MR image from a control subject, where 6 ROIs are highlighted, located respectively in the left (1), central viii

9 (2), and right (3) aspects of the genu and splenium (4, 5, 6). The predicted major orientations of these ROIs are: ROI 1 = ROI 6 = 45 ; ROI 2 = ROI 5 = 0 ; and ROI 3 = ROI 4 = 135, similar to the trajectory direction of specific fiber tracks. Note that the same angle between the paired ROIs (1 versus 6; 2 versus 5; 3 versus 4) reflect a near parallel trajectory direction of fiber tracks going through correspondent ROIs. Figure derived from Sharma S. and Zhang Y., PLoS one, 2017, p Fig. 3.3: Examples of calculated orientation profiles using our method from chosen corpus callosum ROIs. Panel A shows the same T2-weighted image and ROIs as seen in Fig. 2. Panel B shows the major (the highest peak) and all aligning directions in each ROI in the frequency domain, which are perpendicular to the angles shown in Fig Here, ROI 1 = ROI 6 = 135 ; ROI 2 = ROI 5 = 90 ; and ROI 3 = ROI 4 = 45, 90 from the direction of fiber tracts. Figure derived from Sharma S. and Zhang Y., PLoS one, 2017, p Fig. 3.4: Summarized dominant directions per ROI and subject group based on T2-weighted MRI of corpus callosum. Top row shows example images of corpus callosum from a control subject (A), and from patients with relapsing-remitting (B) and secondary progressive (C) MS, suggesting increasing degrees of brain atrophy. Bottom left plot shows the dominant aligning angles of each ROI from the 3 groups. Bottom right plot shows the correlation between predicted and calculated dominant orientations at each aligning angle. Data shown are mean and standard error in plots (L: left, C: centre and R: right). Figure derived from Sharma S. and Zhang Y., PLoS one, 2017, p Fig. 3.5: Angular entropy in each ROI of the 3 subject groups. Panel A shows the mean and standard error of angular entropy summarized by group and ROIs, and panel B demonstrates the distribution of angular entropy of all ROIs per group according to the severity of angular entropy. Note that large value refer high angular entropy, thus high tissue complexity. In panel B, the peak location of the distribution curves shifts toward lower values of angular entropy (less negative) from control to RRMS and then to SPMS, representing greater tissue complexity and injury in patients, particularly in those with SPMS (L: left, C: centre and R: right). Figure derived from Sharma S. and Zhang Y., PLoS one, 2017, p Fig. 4.1: Example ROIs in high-resolution images. Upper panel shows examples of T2-weighted MR image, myelin-stained image and axon-stained image from a post-mortem brain sample ix

10 of MS patient, with ROIs in focal lesion (maroon), diffusively abnormal white matter (DAWM, blue) and normal appearing white matter (NAWM, orange). Lower panel shows the zoomed view of the ROIs Fig. 4.2: Methodology based on MR images. (A) Example ROI in NAWM in a brain slice with zoomed view of the selected ROI; (B) Fourier transform of the selected ROI; (C) Normalized FT power spectrum of the ROI; (D) Thresholded FT power spectrum of the ROI. Thresholded image shows one dominant orientation at around 135 in NAWM Fig. 4.3: Methodology based on histology images. Method demonstration in myelin- and axonstained images based on structure tensor analysis Fig. 4.4: Calculation of dominant orientation and aspect ratio in ROIs. Left panel shows the original images with ROI in focal lesions (red), DAWM (blue) and NAWM (orange) in myelin- and axon-stained images, the central column shows the greyscale converted images and the right column shows the dominant orientation and aspect ratio for each ROI calculated by structure tensor analysis. The ellipticity of each of the ROI represents the degree of anisotropy Fig. 4.5: Example results of orientation profiles of tissue ROIs. Examples of orientation profile of the FT power spectrum of focal lesions (red), DAWM (blue) and NAWM(orange) in T2- weighted image (left panel), myelin-stained image (middle panel) and axon-stained image (right panel) Fig. 4.6: Angular entropy outcomes in both MRI and histological images. Angular entropy results calculated from three tissue types in (A) MR, (B) myelin- and (C) axon-stained images. These results show that the angular entropy in all 3 images was significantly higher in focal lesions than in DAWM and NAWM Fig. 4.7: Relationship in dominant orientation between MRI and histology. Panel A shows correlation between FT power spectrum-based dominant orientation in T2-weighted MRI and structure tensor-based orientation in myelin-stained images along with correlation coefficient (p < 0.05). Panel B shows similar plot between MRI and axon-stained images. Both plots show significant correlation between MR and histology images Fig. 4.8: Relationship in angular entropy between MRI and histology images. Panel A shows correlation between FT power spectrum-based angular entropy in T2-weighted MRI and structure tensor-based angular entropy in myelin-stained images along with correlation x

11 coefficient (p < 0.05). Panel B shows similar plot between MRI and axon-stained images. Both plots show significant correlation between MR and histology images Fig. 5.1: Different ellipticity representing different FA. Images from left to right represent increasing ellipticity and increasing FA. The image on the left end has the lowest FA and the image on the right end has the highest Fig. 5.2: ROIs selection in T2-weighted image and corresponding maps. Top left panel shows six ROIs in left, center and right genu and splenium in T2-weighted image. Top right, bottom left and bottom right panel shows the corresponding ROIs on the FA, L1 and V1 maps Fig. 5.3: Distribution of FA in 6 sets of ROIs. Bar graphs shows the distribution of FA values (mean ± standard error) in 6 ROIs in patients (RRMS, SPMS) and controls Fig. 5.4: Correspondence of FA with angular entropy. Scatterplot shows correlation between FT power spectrum-based angular entropy and average FA in (A) patients (RRMS, SPMS) and (B) controls Fig. 5.5: Correspondence of orientation strength with highest diffusivity. Panel (A) shows correlation between FT power spectrum based-orientation strength and highest diffusivity in patients and panel (B) shows similar correlation in controls xi

12 List of Abbreviations Abbreviation ADC DAWM DTI DWI FA FLAIR FT MD MRI MS NAWM PPMS PRMS QSM RD ROI RRMS SPMS SWI Definition Apparent diffusion coefficient Diffusively abnormal white matter Diffusion tensor imaging Diffusion -weighted imaging Fractional anisotropy Fluid attenuated inversion recovery Fourier transform Mean diffusivity Magnetic resonance imaging Multiple sclerosis Normal appearing white matter Primary-progressive multiple sclerosis Primary-relapsing multiple sclerosis Quantitative susceptibility mapping Radial diffusivity Region of interest Relapsing-remitting multiple sclerosis Secondary-progressive multiple sclerosis Susceptibility-weighted imaging xii

13 CHAPTER 1: INTRODUCTION 1.1 Overview of research Magnetic resonance imaging (MRI) has become a pivotal tool in the diagnosis and management of many neurological diseases such as multiple sclerosis (MS) (1). MS is an inflammatory demyelinating disorder of the central nervous system that primarily affects the young adults at 20 to 40 years of age (2). Approximately 2.5 million people are suffering from this disease throughout the world, with females 2 to 3 times more than males (3). There is still no cure. While the exact mechanisms remain unclear (4,5), MS has an extremely high occurrence rate in places with extreme climates such as North America and Northern Europe, and Canada remains at the top of the list (6 8). The hallmark of MS pathology includes demyelination, axonal damage, and remyelination, besides inflammation, gliosis and other associated repair processes (9). These changes are mostly manifested as focal lesions in MRI; however, they can present as diffusely abnormal white matter (DAWM) or invisible as part of the normal appearing white matter (NAWM). Collective consequences of such MS pathology are believed to play a major role in the progression of patient disability. Traditionally in a clinical setting, MRI has been mainly used to evaluate the number and volume of MS lesions. However, based on the current understanding of the neuropathophysiology of MS, not all MS lesions are the same, and not all lesions are associated with the same consequences (10). Except for the impact of location to neurological functions, the fundamental pattern of tissue structure of these lesions may play an important role. Advances in MRI technology have led to the development of several new imaging methods for characterizing tissue structure, 1

14 particularly those for assessing the alignment, and anisotropy of nerve fiber tracks, represented by diffusion-tensor imaging (DTI) (11). My purpose in this thesis is to determine if tissue alignment characteristics can be obtained using standard MRI, with the assistance of advanced image processing and analysis techniques. Development of new, diffusion-like orientation metrics of tissue structure in clinical MRI can help improve our disease managing ability and thereby enhance the quality of patient care. 1.2 Hypothesis and specific aims The overall goal of this research is to develop and validate an image post-processing method based on Fourier transform (FT) power spectrum. The FT is the cornerstone of both CT and MRI systems. Every MR image is acquired in a spatial frequency domain originally. Therefore, studying the frequency spectrum of MRI provide an intuitive way of understanding tissue property. The anisotropic feature of myelinated nerve fibers is essential for maintaining the structure and function of white matter. With either myelin or axonal disturbance, these unique white matter properties can be lost. In this thesis, I hypothesize that alterations in the content and organization of tissue microstructure are associated with changes in the coherency and anisotropy of MR signal intensity. My overall objectives include: 1) Development of a method for assessing tissue alignment; 2) Validation of the method using postmortem brain samples with MS; and 3) Testing the utility of this method in a clinical setting along with comparisons with advanced diffusion MRI. Specific aim 1: Develop an imaging analysis method for detecting tissue orientation in T2- weighted MRI. This is achieved based on the FT power spectrum, with testing using MRI of 2

15 corpus callosum, a highly aligned white matter structure of the brain. This aim is described in Chapter 3 of this thesis. In this Chapter, the methodology has been explained and outcome presented using images from both MS patients and healthy controls. Specific aim 2: Validate the developed method with MR images from postmortem brain samples with MS. To ensure accurate comparison, orientation outcomes in histology is derived using a new tissue directionality-assessing method known as structure tensor analysis from the same brain sample. Findings of this aim have been presented in Chapter 4 of this thesis. This includes comparison of MRI outcomes with those from both myelin- and axon-stained images. Specific aim 3: Test the utility of the developed method in a clinical setting through comparison with diffusion tensor imaging (DTI). This aim is explained in Chapter 5. Archived images are used for this aim, which includes both conventional and DTI images from the same subject acquired using a clinical MRI scanner. Corresponding outcome measures derived using my method with T2-weighted MRI are compared with that using DTI. 1.3 Thesis organization The overall objective of this thesis is to develop a FT power spectrum-based image processing method for quantitatively assessing tissue alignment and integrity using clinical MRI. This thesis is organized as follows. Chapter 1 gives a brief overview of the research and the structure of the thesis. Chapter 2 is literature review, which describes the general facts about MS, current MR imaging methods for studying tissue alignment, and relevant image analysis techniques. This also highlights the different pre-processing approaches applied in the research including the 3

16 selection of greyscale conversion and the technique of upscaling the size of regions of interest. Chapter 3 describes the development details for the FT power spectrum-based method. The procedures as well as applications of the method are presented in this chapter. Chapter 4 and 5 are based on the discussions of the methodology developed in Chapter 3. Specifically, Chapter 4 presents further validation of the method based on post-mortem images. This is facilitated by the use of the new method, structure tensor analysis, in myelin and axonstained images. Tissue alignment and complexity are measured from identical regions of interest in both MRI and histology. Chapter 5 describes the feasibility of my new method for use in standard clinical MR images, in comparison with DTI. The focus was on the corpus callosum, with six regions of interest selected per subject. Chapter 6 presents the summary and conclusion of this thesis. It also includes a brief discussion about future work that can be done to enhance the utility and potential of this method. 4

17 CHAPTER 2: LITERATURE REVIEW 2.1 Overview of MS MS is one of the most common neurological disorders that affects young adults in Canada, with over 100,000 people impacted (12). MS is diagnosed based on the McDonald criteria, in which MRI plays a pivotal role through detection of lesion number, location, size, and evolution (13,14). Four subtypes of MS have been commonly reported: [1] relapsing-remitting MS (RRMS) subtype has been the commonest, where at least 70% of the patients start with this phenomenon; [2] primary- (PPMS) and [3] secondary-progressive (SPMS) subtypes are fewer and [4] progressive-relapsing (PRMS) with disease progression from the onset with acute relapses, but are more challenging to treat based on the current evidences (15). However, early detection and intervention of MS may have the potential to improve patient outcomes in all types of MS (16). For the early detection of MS, MRI becomes an optimal imaging modality, as the pathological processes that cause ultra-structural changes in the nanometer to micrometer scale can lead to detectable changes at a millimeter scale in MR images (17,18). The hallmark of MS is multifocal plaques that include: inflammation, demyelination, gliosis, edema, and remyelination. Pathologically MS patients usually develop 4-5 new lesions per year along with reactivation or enlargement of pre-existing lesions (19). The myelin sheath is an extremely organized structure in the healthy brain along with the underlying axons. The main function of myelin is to protect the underlying nerve cells from damage and to support the efficient conduction of nerve impulses (20). The anisotropic orientation of myelinated nerve fibers appears to be mainly due to the uniform alignment of the myelin sheath (21). Loss of myelin not only causes disruption of nerve signal transmission: which can interrupt many 5

18 neurological functions of the body (22), but also compromises the structural integrity of myelin and axons, leading to alterations in measurement of tissue alignment and anisotropy (Fig. 2.1) (23). Following remyelination and repair, white matter anisotropy may be restored, likely not as integrated as the normal myelin (24). Fig. 1.1: Demyelination in MS. The top panel shows a healthy neuron with intact myelin sheath and the bottom panel shows a neuron with disrupted myelin sheath. 2.2 MRI of MS MRI is revolutionary in the investigation of MS and at present, it is regarded as the best imaging modality for MS (25,26). MRI creates unique tissue contrast, allowing for a sensitive detection of tissue pathology. The flexibility in the choice of imaging sequence and parameters has made MRI a highly effective method for the detection of tissue abnormalities. The benefit of MRI for assessing MS has been shown repeatedly in the literature (27). For example, within the first year 6

19 of a single attack, MRI has 94% sensitivity and 83% specificity in diagnosing MS (25). In the following paragraphs, I will review some conventional and advanced MRI techniques relevant to this research Conventional MRI methodologies Common conventional MRI techniques used in the diagnosis of MS include pre- and postcontrast T1-weighted imaging, T2-weighted imaging, and sometimes proton density and FLAIR imaging (28). Lesion image intensity increases with increased echo-time on T2-weighted images and proton density images using conventional spin echo techniques (29). Each MR technique has its unique advantages over the others (27,30), as shown below T1-weighted imaging T1-weighted imaging uses a short echo time and short repetition time to differentiate anatomical structures by their longitudinal relaxation time. Tissues with high-fat content appear brighter in contrast to water-filled tissues which appear darker. T1 hypointense lesions are more common in MS than in other cortical pathologies (31). T1 hypointensity is caused by extracellular space expansion due to the accumulation of water and loss of structural components (32). Without chronic tissue injury or edema, MS lesions appear isointense compared to healthy tissue on T1- weighted imaging (33). Post-contrast T1-weighted imaging is mainly used for the identification of new lesion activity. Gadolinium-enhancement help visualize newly formed lesions and locate the breakage of the blood-brain barrier that leads to the infiltration of immune cells (34). Gadolinium-enhanced T1-weighted imaging can also help determine the dissemination of lesions in space and time in patients (20). 7

20 T2-weighted imaging T2-weighted imaging has a long repetition time and long echo time and is often used to demonstrate lesion burden as most of the lesions are filled with water, which appears hyperintense. Only a fraction of T2 hypointense lesions are evident as compared to T1 hypointense lesions (35,36). In T2-weighted imaging, multiple focal hyperintense lesions, usually round or oval, are observed in MS (Fig. 2.2). Such lesions may be present in any part of the brain or spinal cord in patients, and they are visible longer than post-contrast T1 lesions (23). It is difficult to differentiate between chronic and acute lesions using T2-weighted MRI alone (32,37). Nonetheless, there is evidence showing that there is a close correlation between lesions seen in pathology and lesions seen on the T2-weighted MRI (38). Fig. 2.2: Lesion presentation in different MR images in a MS patient. Axial proton-density weighted (a,d), T2-weighted (b,e) and FLAIR (c,f) consecutive images of a patient with MS demonstrate multiple lesions around lateral ventricles as shown by arrows. Figure derived from Sahraian M. et al., MRI Atlas of MS lesions, 2008, p

21 Proton density weighted imaging In proton density-weighted imaging, the anatomical structures are differentiated based on proton density by controlling the scan parameters (short echo time and long relaxation time) to minimize the effects of T1 and T2 relaxations. The tissues filled with water appear bright in proton density weighted imaging whereas fatty tissues appear dark. Proton density weighted imaging is particularly sensitive to plaques in posterior fossa in MS (39). However, lesions do not appear prominent. Multi-slice mapping of myelin water fraction has been shown to be feasible using T2* decay (40) FLAIR imaging FLAIR (Fluid Attenuated Inversion Recovery) MR imaging is a very sensitive sequence in the detection of MS lesions (29). FLAIR produces images weighted strongly by T2 time due to cerebrospinal fluid suppression using inversion recovery sequence. Inversion recovery sequence produces signal representing longitudinal magnetization after the application of a 180 pulse which can help suppress the fluid. Thus, lesions near the ventricles or cortex can be better identified using FLAIR than other sequences (39). In particular, sagittal thin-section fast FLAIR sequencing demonstrates superior sensitivity in identifying MS plaques in the brain as compared to axial proton-density or T2-weighted spin echo pulse sequences (41). Nonetheless, as for most of the standard MR sequences, FLAIR imaging is limited in its capability to specifically quantify the lesion pathology. 9

22 2.2.2 Advanced MRI methodologies The continuous advance of MR technology has led to an increased capacity to detect subtle changes in MS pathology (42,43). In assessing fiber tract alignment, representative technologies are DTI and susceptibility-weighted imaging (SWI) Diffusion-weighted imaging (DWI) DWI is based on measuring the Brownian motion of water molecules. (19,44). This motion depends on structural orientation of the white matter and is often anisotropic in the central nervous system (45,46). DWI is obtained by adding symmetric pairs of diffusion sensitizing gradient coils around 180 pulse to standard T2-weighted spin-echo sequence. Hence, DWI is actually T2-weighting so changes in T2 time in a tissue will affect the appearance of diffusion images, but those changes do not actually reflect the diffusivity within a tissue (44). Measurement of diffusion gives information about the orientation, structure, and integrity of the tissues (47,48). The diffusion of biological tissue is more restricted than that of free water, and is often measured as apparent diffusion coefficient (ADC) in DWI: it is the measure of the magnitude of overall diffusion (49). DTI calculates the magnitude and direction of water diffusion by means of diffusion tensor; therefore, the existence of tissue anisotropy is captured by the tensor (50). Elevated radial diffusivity (RD) within active MS lesions has been shown to be indicative of severe tissue injury (51). The diffusion tensor also enables the calculation of mean diffusivity (MD) and fractional anisotropy (FA) at each voxel (Fig. 2.3) (45). Mean diffusivity measures diffusion independent of the structural orientation (52). Fractional anisotropy provides the degree of anisotropy of diffusion in different directions and may reflect the fiber density, axonal diameter, myelination, and demyelination. MS plaques have shown 10

23 increased diffusivity in DWI and then increased ADC (46,53). DTI fiber tracking (tractography) uses diffusion tensor to track fibers along their pathway by following the continuity of the main diffusion of adjacent voxels (54,55). Regions with complex architectures of crossing myelin fibers can also be evaluated using DWI fiber orientation distribution (FOD) (56). Diffusion imaging is used in a variety of neurological disorders but it has not been implemented in routine clinical studies in MS. Standardization is one of the limitations. Fig. 2.3: Different outcomes from DTI. Left column shows non-diffusion weighted scan followed by FA, MD in color coded form. Figure derived from Winston GP., Quant Imaging Med Surg., 2012, p Susceptibility-weighted imaging Magnetic susceptibility is a measure of the magnetization response of a material in the presence of an external magnetic field which illustrates whether a material is repelled from or attracted to 11

24 a magnetic field. This method of imaging uses the difference in tissue magnetic susceptibility to generate image contrast. Susceptibility-weighted imaging (SWI) uses the magnitude and phase to separate MR signals from tissues with different susceptibilities, such as: paramagnetic, diamagnetic or ferromagnetic. The main application of SWI has been the imaging of veins and microhemorrhages (57,58). Quantitative susceptibility mapping (QSM) is one of the useful outcomes that can be generated from the phase images of SWI, assuming that the phase shift is due to changes in bulk tissue magnetic susceptibility (59 61). Phase contrast is an integral part of MR images and is readily available with every MR scan without additional scan time. Phase imaging can provide superior contrast and resolution for small anatomical structures that are not clearly visible in the respective magnitude images. The heterogeneity of magnetic susceptibility within brain tissues creates a unique contrast between gray and white matter in MR phase images obtained by the gradient echo sequence (62,63). Phase images have shown differences in MS lesion pathology, indicating that QSM may be applicable to MS imaging. (64). Susceptibility tensor imaging has been investigated to study anisotropic susceptibility by using tensors in the quantitative susceptibility mapping. Structures at both cellular and molecular level affect both the frequency and amplitude of MRI signal in white matter, causing orientationdependent disturbance in atomic structures, resulting in anisotropic susceptibility (65). The magnetic properties of brain tissue are dominated by: myelin, iron and deoxyhemoglobin. The distribution of these materials is highly heterogeneous, leading to fluctuations in the overall susceptibility mapping (66). Further, experimental and theoretical explanations have shown that magnetic susceptibility itself is dependent upon white matter fiber orientation. The measurement of bulk susceptibilities will not only be affected by the presence of tissue structure but also by 12

25 the components of the paramagnetic susceptibility tensor parallel to the fiber direction (67). Magnetic anisotropy is the strongest in white matter, mostly due to the anisotropically organized myelin as compared to gray matter (68,69). The observed consistency in large fiber bundles between diffusion- and susceptibility tensor-measured orientation indicates the existence of common biology mainly related to the structural support from axons and myelin, the primary source of susceptibility anisotropy (67). 2.3 Image post-processing and analysis methodologies Besides image-acquisition based methods, tissue structural properties following injury and repair may be obtained through image analysis strategies. Demyelination causes alterations in the content and organization of tissue microstructure that may lead to changes in the distribution of MR signal intensities (also known as MRI texture). These changes are detectable using mathematical analysis of MRI. Some of the image post-processing techniques, their procedures and applications are described below Spatial frequency analysis Specific image has a unique Fourier transform (70). The FT represents an image as a summation of sinusoidal images: a natural basis for periodic structures. The FT thus can decompose the information in an image into frequency distributions at different orientations (angles) in polar coordinates. Directional anisotropy and fiber orientation can be extracted on the basis of 2- dimensional (2D) FT. Correlations between the orientations of neighboring fibers have been previously shown in animal model of MS, experimental autoimmune encephalitis (70). Moreover, the 2D FT has been used as a method for studying local tissue architecture of collagen 13

26 (71). On this basis, the directional anisotropy of the myelin can be inferred. Fourier spectrum preserves the rotational information and it s orientation has been used to represent the general orientation of the nerve fibers in histology images. Begin et al segmented images using thresholding methods and assessed myelin integrity at local areas of interest based on directional anisotropy (Fig. 2.4). Average fiber orientation was inferred from the direction of the minor axis of image spectra-formed ellipses, where the minor axis corresponded to the main axis of the original image (18). A number of other FT-based approaches have also been introduced for the determination of tissue directionality in an image including FT power spectrum and dominant orientation analysis. Fig. 2.4: An image processing example using Fourier transform (FT) in histology. (a) Three different coherent anti-stokes Raman scattering (CARS) images of mouse spinal cord 14

27 surface shown in longitudinal orientation. From top to bottom, Experimental Autoimmune Encephalomyelitis lesions progressively afflict the tissue. (b) Processed 2D-FT of the CARS images in (a). (c) Segmented objects with superimposed equivalent ellipse and main axes. (d) Average fiber orientation overlaid on images with aspect ratio (AR, see color code on the right). Figure derived from Begin et al., Biomed Opt Express, 2013, p The squared modulus of the FT is known as the power spectrum, integration of which yields what is known as the energy. The power spectrum carries enhanced information regarding the intensity and distribution of frequencies in an FT Image, which corresponds with spatial information. In comparison to conventional Fourier analysis, the power spectrum has a simple pattern, and is sensitive to slight differences in the degree of orientation. With enhanced energy content,, the FT power spectrum can be used to determine the periodicity of spatial frequency and textural features (72). Due to the large dynamic range of the FT power spectrum, the log scale is generally used for visualization and analysis purposes. It is worth noting that despite the many advantages, the FT does not always demonstrate tissue specific frequency information in a sensible manner. The wavelet transform is a better choice than the FT for extracting frequency and space variations, but it is also sensitive to noise (73). The short-time FT is a method for determining the frequency and phase content of local signals by dividing the original time signal into short equal segments and computing the FT by a technique called windowing. The short-time FT has an inherent resolution problem such that narrow windows have poor frequency resolution and wide windows have poor time resolution. The wavelet transform can split the signal into different frequency bands that give the time 15

28 interval of the specified frequency. Another FT-related method, the Stockwell transform, overcomes this limitation by combining aspects of the short-time FT and wavelet transforms. The Stockwell transform facilitates time-frequency decomposition showing phase and frequency information at each pixel (74 76). Due to the use of a self-adjustable Gaussian window, the Stockwell transform can provide good resolution at relatively high frequencies (77). A polar form of the Stockwell transform has also been devised. Prior studies using polar Stockwell transform have shown that pathological changes in an animal model of MS lead to significant changes in the intermediate ranges of frequency energy, suggesting that this method could be used as a tool for understanding lesion pathology (78,79) Structure Tensor Analysis Structure tensor analysis has been recently introduced as another image post-processing strategy for assessing tissue alignment. Studies have been mostly focused on animal models currently. Structure tensor, also known as the second-moment matrix, is derived from the gradient of mathematical functions. The calculation of structure tensor is related to the eigenvectors of an image, which in turn relates to the image directional gradient along the dominant orientations. The strength of eigenvectors is given by eigenvalues. The directionality outcomes of structure tensor analysis can be visualized as an ellipsoid. This ellipsoid has semi-axes lengths equal to the eigenvalues in the direction of the respective eigenvectors. This information is relevant to DTI calculations as well. Structure tensor analysis has been implemented in many image processing algorithms (80). Recently, it has been applied to light microscopic images obtained from histological samples for the validation of diffusion MRI (81). 16

29 Fig. 2.5: Flowchart of structure tensor analysis with associated outcomes. Left column shows an example natural image. Middle column shows the orientation, coherency and energy maps for studying the directional properties of the image. The color-coded map on top right shows the orientation features of the original image. The circular histogram on bottom right is another way of representing orientation information, where the radius of points in the blue curve shows the orientation strength of a feature at a particular angle. Figure derived from Rezakhaniha et al., Biomech Model Mecahnobiol., 2012, p Structure tensor analysis has been used in multiple applications through assessment of tissue coherency and anisotropy. It shows success in the study of microstructural integrity in collagen 17

30 and elastin fibers (82,83), and recently the orientation and coherency of nerve fibers (Fig. 2.5) (84). Furthermore, using an animal model of MS, structure tensor analysis has also shown the promise for assessing the directionality of myelin in the white matter of spinal cord (46) Statistical-based image analysis Advanced image analysis using statistical methods has shown to be another approach to enhance the power of MRI. These methods have shown promise in detecting subtle changes in MRI signal intensity, or MRI texture. The textural features in MRI reflect microstructural information of the underlying tissue structures (85,86). The characteristic distribution and variations of MRI pixel intensity calculated mathematically include: contrast, entropy, and angular second moment, among multiple other first- and second-order parameters (87,88). Statistical analysis of T2- weighted MRI shows the ability to separate active and inactive lesions, where the detection sensitivity increased from 55.6% to 88.9% (89); the ability to separate MS patients from healthy controls (90); and the potential to identify tissue integrity not directly visible in earlier stages of MS (46). In a cuprizone mouse model of demyelination, statistical analysis has also shown the sensitivity to detect remyelination several weeks after demyelination (78), although this still needs confirmation in human studies. 2.4 Summary Clearly, there are many standard and advanced imaging methods under development that may help advance the assessment of MS pathology. The focus of my research is on the development and verification of a new image post-processing method, based on FT power spectrum. Specifically, I will develop outcome measures related to changes in tissue directionality in 18

31 subjects with MS. This will be based on standard MRI, and compared with existing approaches including DTI and structure tensor analysis. This research may help improve our disease monitoring ability in clinical practice. Ultimately, it may help enhance patient management with MS and similar diseases. 19

32 CHAPTER 3: FOURIER TRANSFORM POWER SPECTRUM IS A POTENTIAL MEASURE OF TISSUE ALIGNMENT IN STANDARD MRI 3.1 Introduction The alignment and integrity of nerve fibers are associated with the conducting efficiency of nerve signals in the brain. In many neurological diseases such as MS, the usual coherency of nerve fibers is disrupted following tissue injury. Coherency, in general, means how well the nerve fibers are aligned in one specific direction. This disruption is shown to occur not only in focal plaques, but also in the NAWM of MS (91), leading to paramount functional impairments in patients (92). A number of studies have attempted to measure white matter coherency using MRI; however, there are few methods using clinical MRI protocols. The availability of a method clinically has important implications both in the evaluation and treatment of nerve pathology. Current MRI methods for assessing white matter coherency are mainly based on advanced MRI techniques, such as DTI (93,94). Diffusion-based MRI evaluates the random movement activity of water molecules in a tissue. Through assessment of both the magnitude and direction of water diffusion, DTI including FA and directional diffusivity can detect the organizational property of specific fiber tracks that particularly dominates the white matter (45). Recent studies have also found that the potential of DTI can be significantly enhanced with combination with image postprocessing strategies (95,96). Indeed, Stamile et al show that histogram analysis of DTI fractional anisotropy is more sensitive to subtle tissue changes in MS lesions than traditional regions-of-interest (ROI) approaches and can differentiate nerve fibers with or without changes within the same bundle (96). On the other hand, a new MRI method named susceptibility tensor imaging has also shown promise for detecting white matter alignment (69). This method uses 20

33 specific MRI pulse sequences that are sensitive to magnetic susceptibility generated in a tissue, where a 3-dimensional susceptibility tensor is modeled. Using both animal and human brains, Li and colleagues demonstrate that the susceptibility anisotropy relates to the orientation of myelin in white matter and is significantly decreased in mouse with dysmyelination. However, while with great potential, most of the aforementioned techniques require advanced image acquisition and assessing skills, are not routinely obtained in clinical practice in many diseases, and their utility are subject to further validation. An alternative approach for assessing tissue anisotropy is to use image post-processing methods. In this regard, spatial frequency-based methods play an important role such as the FT. The FT identifies the total content of frequencies enclosed in an image. In the frequency domain, the pixels tend to group along the direction of tissue alignment and form shape-specific clusters; circular cluster represents isotropic tissue and elliptical cluster refers to anisotropic tissue. Based on this theory, Begin et al. (70) have successfully detected the orientation of myelin using histological images from demyelinated spinal cord of mouse. Similarly, using fixed hippocampal specimens from healthy human brain, Nazaran et al. (97) demonstrate that the dominant orientation of an angular histogram based on two-dimensional FT corresponds to the alignment of myelinated fibers in histology. In addition, multiple other studies have indicated the enhanced potential of orientation metrics derived from the power spectrum of FT. Some studies have demonstrated the ability of Fourier power spectrum for identifying the weave pattern and aligning stripes in natural fabrics (98,99). Others (100,101) show that the density and orientation of FT power spectra are highly correlated with the organization of collagen fibers in mice with 21

34 fibroblasts. These studies highlight the potential of FT-based image analysis methods, although demonstration of their utility in clinical images is limited. In this study, based on the power spectrum of 2D FT, we developed a new approach for assessing tissue alignment using T2-weighted MRI. This method is validated using a highly aligned brain white matter structure, corpus callosum. To test the feasibility of this method for clinical use, we also compared results from healthy controls with those from patients with mild and advanced MS. 3.2 Materials and Methods Subjects This is a retrospective study based on MR images acquired from 19 female MS patients and 19 age- and gender-matched controls (PI: Dr. Lenora N Brown, University of Calgary). Of the 19 patients, 10 had RRMS and 9 had SPMS. The mean (standard deviation) age was 37.5 (12.5) years for RRMS and 59.5 (10.5) years for SPMS. The focus of the original clinical study was to assess the structure and function of corpus callosum, a white matter structure coordinating the conduction of numerous inter-hemispheric functions (102). To understand the difference between cohorts, the recruitment was done in a way such that only patients with very mild RRMS (disability score 3/10) or very advanced SPMS (disability score 6/10) were included. This study was approved by the Institutional Health Research Ethics Board. Written informed consent was obtained from all participants. 22

35 3.2.2 MRI acquisition All images were acquired at a single 3T MR system (GE Healthcare, DISCOVERY MR750, Milwaukee, USA). Whole brain MRI protocols included both T1- and T2-weighted sequences based on clinical standard. In this study, we focused on T2-weighted MRI because that can sensitively detect tissue abnormalities without the need for contrast injection. These images were obtained with a spin-echo sequence with field of view = 240 x 240 mm 2 ; matrix size = 256 x 256; repetition time (TR)/echo time (TE) = 6035/83 ms; and slice thickness = 3 mm, without gap. Acquisition of T2-weighted MRI took 2 minutes per subject Selection of image and regions of interest In both patients and controls, we examined T2-weighted MR images that demonstrated the best delineation and the largest area of corpus callosum at equivalent regions, both for the genu and splenium of the structure. To include image areas with different aligning directions, six ROIs were selected from the axial images of the brain, located at the left, central, and right aspects of genu and splenium respectively (Fig. 3.1). The dominant aligning angle of nerve fibers in these ROIs can be visually observed and were estimated at 45, 0 and 135 at left, central and right genu and 135, 0, 45 at left, central and right splenium respectively (Fig. 3.2). These angles are based on the anatomical trajectory of the nerve fiber tracts. Therefore, these degrees are relative to the anatomy of the brain rather than physical degrees. So, the change in patient position relative to scanner will not affect those measurements. The size of these ROIs ranged from 6x6 to 8x8 pixels, equivalent to mm 2 to mm 2 ; large ROIs were located in the central regions of the genu and splenium where showed the largest cross-sectional area of the corpus callosum. Based on our preliminary tests, ROIs smaller than the proposed size could not provide 23

36 sufficient content of directional information whereas ROIs larger than that could not fit into the boundary of the structure. To keep consistency, MR images that involved focal lesions within or crossing the corpus callosum were excluded. In addition, patients with the corpus callosum being too atrophic to fit any ROIs were also excluded. Fig. 63.1: Method demonstration. Shown is an example ROI in the right genu of corpus callosum (A, small box) and its zoomed view (B). Then the FT (C) of the ROI, with normalization (D) and thresholding (E). Based on E, the orientation profile of the ROI is calculated (F), from which corresponding angular entropy can be computed. Figure derived from Sharma S. and Zhang Y., PLoS one, 2017, p

37 Fig. 73.2: Representative ROIs along with predicted major aligning orientations based on the anatomy of corpus callosum. Panel A shows an example T2-weighted MR image from a control subject, where 6 ROIs are highlighted, located respectively in the left (1), central (2), and right (3) aspects of the genu and splenium (4, 5, 6). The predicted major orientations of these ROIs are: ROI 1 = ROI 6 = 45 ; ROI 2 = ROI 5 = 0 ; and ROI 3 = ROI 4 = 135, similar to the trajectory direction of specific fiber tracks. Note that the same angle between the paired ROIs (1 versus 6; 2 versus 5; 3 versus 4) reflect a near parallel trajectory direction of fiber tracks going through correspondent ROIs. Figure derived from Sharma S. and Zhang Y., PLoS one, 2017, p Analysis of tissue alignment Our orientation analysis method involved several image processing steps. These included: 1) convert identified images into the frequency domain; 2) calculate power spectrum; 3) extract angular distributions of the power spectrum; and 4) compute the dominant orientation and 25

38 directional complexity in a ROI. We used a fast FT algorithm (ImageJ, NIH, version 1.8, USA) to obtain the frequency content in each ROI. Further analysis of these images was done using algorithms built in-house. To enhance frequency energy, we calculated the power spectrum of the FT, the power spectrum was then normalized through computation of its logarithmic value and thresholded to eliminate noise and edges. Eventually, the upper twenty percentiles (80-100) of the intensity of normalized power spectrum was preserved. Please note that the application of thresholding may exclude information outside the selected area and which may not align in the dominant orientation. This mostly removes information due to artifacts and the information represented by only few number of pixels. Therefore, the major alignment of the nerve fiber tracts doesn t significantly change. Next, polar conversion was applied to the thresholded power spectrum, from which the distribution of the orientation angles was obtained for each ROI; the higher the peak, the greater aligning strength at that angle. The angular location of the highest peak in an orientation distribution was considered the dominant direction of the tissue contained in a ROI (see plots in Figs 3.1 and 3.3). 26

39 Fig. 83.3: Example orientation profiles calculated using our method from chosen corpus callosum ROIs. Panel A shows the same T2-weighted image and ROIs as seen in Fig. 2. Panel B shows the major (the highest peak) and all aligning directions in each ROI in the frequency domain, which are perpendicular to the angles shown in Fig Here, ROI 1 = ROI 6 = 135 ; ROI 2 = ROI 5 = 90 ; and ROI 3 = ROI 4 = 45, 90 from the direction of fiber tracts. Figure derived from Sharma S. and Zhang Y., PLoS one, 2017, p Analysis of alignment complexity Based on the angular distribution profile of a ROI, we also calculated another outcome named angular entropy. It is a measure of angular scattering of tissue orientations and thus reflects the complexity of tissue alignment. Complexity, in this study, means the amount of tissue heterogeneity following tissue damage. Angular entropy (ε) was calculated as ε = -Σ pθ*{log pθ}, where pθ referred to the probability of alignment at a certain angle θ. Based on this equation, the angular entropy gave rise to a negative value, with zero being the maximum. 27

40 3.2.6 Statistical Analysis We used a mixed-effect modeling method to evaluate differences in tissue complexity using a statistical package Stata (StataCorp, Texas, USA; version 12). This method allowed us to consider both intra-subject variances between different ROIs and inter-subject variances between control and patient groups. Tissue angular entropy was defined as the dependent variable. In the analysis of dominant orientations of a tissue, we compared the derived angles using our method with angles estimated based on anatomy within individual ROIs. This correlation was also done using mixed-effect modeling in Stata. For all of the comparisons, P-value <= 0.05 was defined as significance. 3.3 Results Sample Characteristics We examined 204 corpus callosum ROIs in total, 90 from patients and 114 from controls. Of the 90 patients ROIs, 35 were in SPMS from 9 patients, and 55 in RRMS from 10 patients. Anatomically, there were 14, 15, and 15 patient ROIs located in the left, central, and right aspects of genu, and 15, 16, 15 in the left, central, and right aspect of splenium. Please see Table 1 for details. In controls, 6 ROIs per subject were identified in the corpus callosum, totaling 19 subjects. Overall, 24 ROIs were excluded from patients due to excessive tissue atrophy; most of these regions were located in the left (4) and right (4) genu of SPMS patients. No lesions were involved in any of the remaining ROIs, and all control ROIs appeared normal in MRI. 28

41 Table 3.1 : The number of regions of interest examined in the corpus callosum per group. Control RRMS SPMS Patients Total Left Genu Central Genu Right Genu Left Splenium Central Splenium Right Splenium Correspondence of dominant tissue alignment between quantified and observed angles From analysis of the polar-converted FT power spectrum, we detected a spectrum of orientation peaks per ROI, in which a dominant peak was persistently identified in each ROI (Fig. 3.3). The mean (± standard error) angles of the dominant peaks were (0.684 ), (4.58 ) and (0.977 ) in the left, central and right genu, and 45 (0 ), (3.41 ) and (1.1 ) in the left, central and right splenium respectively in the frequency domain. Based on the anatomical location of the fiber tracks of the corpus callosum, the dominant orientation of the 6 ROIs were paired with each other in 3 groups: left genu showed similar angles to right splenium ( versus ); right genu with left splenium (46.42 versus 45 ), and both central ROIs of genu and splenium were parallel with each other ( versus ). Following perpendicular conversion of these angles based on the reciprocal Theorem of the FT, these dominant angles corresponded to 44.32, and in the left, central and right genu, and 135, 7.73 and in the left, central and right splenium in image domain. Such converted orientations were significantly correlated with the observed directions based on the 29

42 anatomical location of the corpus callosum, with an overall correlation coefficient of (p = 0.008), where for controls; for RRMS patients and for SPMS patients (Fig. 3.4). Fig. 93.4: Summarized dominant directions per ROI and subject group based on T2- weighted MRI of corpus callosum. Top row shows example images of corpus callosum from a control subject (A), and from patients with relapsing-remitting (B) and secondary progressive (C) MS, suggesting increasing degrees of brain atrophy. Bottom left plot shows the dominant aligning angles of each ROI from the 3 groups. Bottom right plot shows the correlation between predicted and calculated dominant orientations at each aligning angle. 30

43 Data shown are mean and standard error in plots (L: left, C: centre and R: right). Figure derived from Sharma S. and Zhang Y., PLoS one, 2017, p Increased tissue angular entropy in patients as compared to controls Based on mixed-effect modeling, we found that the angular entropy in MS patients was significantly higher (less negative) than in control subjects (mean ± standard error = ± versus ± , p = 0.013), and tended to be higher in SPMS than in RRMS patients (mean ± standard error = ± versus ± , p = 0.069). This pattern was consistent across all 6 ROI locations (Fig. 3.5A). To understand the heterogeneity of the tissue structure within a group, we also evaluated the distribution of angular entropy from all ROIs (Fig. 3.5B). As shown by individual fitting curves, the peak location was found to be in controls, significantly different than and in RRMS and SPMS groups. In fact, the lowest angular entropy was measured at from a ROI in controls, and the highest at 0 from a ROI in SPMS. 31

44 Fig : Angular entropy in each ROI of the 3 subject groups. Panel A shows the mean and standard error of angular entropy summarized by group and ROIs, and panel B demonstrates the distribution of angular entropy of all ROIs per group according to the severity of angular entropy. Note that large value refer high angular entropy, thus high tissue complexity. In panel B, the peak location of the distribution curves shifts toward lower values of angular entropy (less negative) from control to RRMS and then to SPMS, representing greater tissue complexity and injury in patients, particularly in those with SPMS (L: left, C: centre and R: right). Figure derived from Sharma S. and Zhang Y., PLoS one, 2017, p

45 3.4 Discussion In this study, we presented a new image processing method for assessing tissue alignment and anisotropy based on FT power spectrum. Using the highly oriented corpus callosum as a validation model, we show that tissue alignment derived from our method is correlated with the anatomical orientations of this structure. Furthermore, through quantitative analysis of the aligning complexity of corpus callosum, we further demonstrate that the current method may have the potential to detect subtle structure disruptions in a lesion-free area of patients using conventional MRI. It is well known that the power spectrum of FT provides intensified information of an image. Mathematically, the power spectrum calculates the square of the spectral energy contained in an image (100). This property facilitates a dramatic increase in both the contrast and detectability of power spectrum. Indeed, using a mouse model of MS (experimental autoimmune encephalomyelitis), a recent study has shown the ability of FT power spectrum to detect the anisotropy of demyelinated axons in mouse spinal cord (70). However, that study was done using histological images that have ultra-high resolution and the alignment of nerve fibres are already visible using bare eyes. In the current study, we took advantage of the benefit of several image pre-processing procedures including spectral normalization and thresholding besides calculation of the power spectrum. Together with polar conversion of corresponding power spectra, unique alignment characterizes of a ROI were generated using standard MRI. This could be useful in future clinical studies when assessment of focal tissue structure is required without the acquisition of additional advanced imaging sequences. 33

46 To test the validity of our method, we used corpus callosum of the brain as an evaluation model. Corpus callosum is a unique interhemispheric structure that contains highly organized white matter fibers. According to the literature, different regions of corpus callosum follow different white matter trajectories, facilitating the conduction of specific brain functions (103). Specifically, the genu connects anterior aspects of the brain between hemispheres allowing for sensory and motor functions, and the splenium links primarily the occipital lobes to coordinate visual function (104). In anatomical MRI, these fiber bundles appear projecting at different angles as predicted in Fig When focusing on a local ROI, however, these small corpus callosum areas are simply a group of pixels with similar signal intensity, where the directional property of its originating tract is no longer visible (see Fig. 3.2B). With the assistance of image analysis using our power spectrum method, we detected distinct alignment profiles in each ROI. This alignment profile refers to the aligning information of the nerve fibers in the given ROI. Following polar conversion, these ROIs showed clear dominance of an orientation peak that dictated the dominant aligning angle of specific white matter tracts passing through correspondent ROIs. Moreover, while this study is targeting the NAWM, this method can be used to assess different tissue structures such as the gray matter, which is expected to show increased complexity of angular profile (data not shown) resulting from increased combination of nerve structures. Using the corpus callosum, we show that the dominant orientation of each ROI calculated using FT power spectrum is consistent with the predicted directions of the corpus callosum at individual ROI regions. This is similar for both patients and controls. In this study, the angular values calculated directly from the ROI were frequency-domain orientations. After reciprocal 34

47 conversion based on the FT theorem, the new orientation values were used to compare with projected directions. We found that the right genu and splenium show the best consistency, while the central region of genu and splenium showed the largest variability (Fig. 3.4). The observation of similarity in dominant aligning angles between patients and controls is reasonable because both cohorts share a similar pattern of fiber projection in the corpus callosum. Moreover, the exclusion of lesions involving the corpus callosum may also play a role in the uniformity of this sample. Nonetheless, as MS patients are associated with greater NAWM damage than controls (105), suggesting the complexity of MS tissue is higher and so its coherency is lower. The degree of tissue damage in this study is evaluated using a new quantitative outcome, angular entropy. Measurement of entropy has been shown to be a powerful tool for assessing tissue injury, although most previous studies have been focusing on the traditional calculation of entropy that is based on the complexity rather than aligning directions of image features. Using a statistical assessing method, Theocharakis et al have shown that sum of entropy in signal intensity is one of the best variables to differentiate MS lesions from microangiopathy in the brain using T2-weighted MRI (106). Based on wavelet transform, Zhang et al (107) demonstrate that the entropy of wavelet sub-bands can help enhance the detectability of MS tissue from healthy controls. Currently, with routine MRI, assessment of entropy based on the aligning property of a tissue structure is scarce, making angular entropy a valuable complementary index. In contrast to traditional calculations, angular entropy takes account of the number, distribution, and dominance of the aligning directions contained in a tissue. This makes it particularly useful to detect subtle changes in white matter because nerve tracts are typically highly aligned. Indeed, in a recent histological study, researchers demonstrate that demyelination is associated with 35

48 significantly greater angular entropy than the intact myelin in mouse spinal cord (46). In the present study, we showed that angular entropy is also significantly elevated in MS patients as compared with controls using T2-weighted MRI, that of considerably lower resolution than histological images. These findings suggest that higher angular entropy (less negative values in this study) is associated with greater tissue complexity, and therefore more severe nerve damage (46). This is further reflected in the SPMS patients who tended to show the highest angular entropy as compared to RRMS and control subjects in corpus callosum. Despite the encouraging findings, we note some limitations in this study. This was a proof-ofconcept study, and therefore we were mainly focusing on a highly organized structure, corpus callosum. As the largest inter-hemispheric white matter structure, corpus callosum is a frequent site of MS pathology (108). Moreover, the anatomical trajectory of fiber tracks in the corpus callosum has already been validated previously in post-mortem studies (103), making this structure an ideal model for evaluating fiber track outcomes (109). This method could be used to assess white matter areas in other parts of the brain and the spinal cord areas besides corpus callosum based on coherency of nerve fibers. In this study, we positioned the ROIs in areas of corpus callosum with visually predictable orientation of fiber tracks. This may raise the suspicion of reliability of the ROIs. However, it is worth noting that although the alignment of white matter tracks in the corpus callosum when visualized as a whole is predictable in brain MRI, the pattern of alignment of individual MRI pixels is not identifiable visually in any arbitrary ROIs, either in the image (Fig. 3.1B) or in the pre-processed FT domains (Fig. 3.1C). In fact, in the process of ROI determination, the only confounder was the ROI size, for which we made sure that each ROI was located within the boundary of corpus callosum to avoid partial volume effect. In 36

49 addition, the sample size of this study was small, which may contribute to the variability between ROIs and limit the degree of significance of the results. Furthermore, while the outcome measures are sensitive to tissue alignment, the contribution of specific pathologies to these measurements is subject to further confirmation. This method could be applied in other MR image types including conventional and advanced images. It is also possible to create parametric maps by analyzing ROIs of similar size by shifting them throughout the whole image. In the future, we seek to investigate how tissue orientation measures help understand disease progression, and how our method compares with other image assessing methods such as machine learning (110) and with advanced MRI techniques including diffusion-tensor imaging (45) and structure tensor analysis (84). 3.5 Conclusions We have shown that it is possible to characterize the alignment of white matter tracks using clinical MRI through quantitative analysis of the FT power spectrum. This would be critical for improving the evaluation of disease activity in normal appearing tissues such as the corpus callosum because NAWM pathology is shown to play a significant role in the progression of disability in MS patients. With further confirmation, this method may also be used to evaluate tissue repair following injury in patients with or without treatment. Moreover, the ROIs we assessed in this study have a similar size to that of typical MS lesions in MRI. Therefore, assessing lesion injury and repair may be also possible using this method. In the clinical management of MS patients, we can compute the dominant orientation and angular entropy of select areas such as corpus callosum, or of focal lesions that are new or that show changes over time to detect invisible characteristics of tissue coherency. This information can be obtained at a 37

50 secondary workstation and serve as valuable add-ons to the routine clinical reports that focuses mostly on lesion size, number, or volume. Finally, while this study is focused on MS, given its image post-processing nature, this method is applicable to other white matter diseases that involve disruption. 38

51 3.6 Acknowledgement We thank the patient and healthy volunteers for their support to this study, the clinical MS research team for recruiting and administrating the study, and Dr Lenora N Brown for initiating this study with the support of the MS Society of Canada. We are very grateful for the funding support for the imaging program from the MS Society of Canada (ID: 2750), Natural Sciences and Engineering Council of Canada (ID: ), and Alberta Innovates Health Solutions. In addition, we thank the scholarship supports for S Sharma from the University of Calgary and NSERC CREATE I3T Program. 39

52 CHAPTER 4: VALIDATION OF FT POWER SPECTRUM-BASED METHOD WITH STRUCTURE TENSOR ANALYSIS 4.1 Introduction MS is a complex inflammatory demyelinating disease of the central nervous system that is associated with severe physical disability in patients. While the fundamental mechanisms of dysfunction are not fully understood, loss of signal-transmitting axons is believed to play a major role (91). Demyelination and axonal loss are both considered key characteristics of MS pathology; however, the severity of these changes is difficult to evaluate in patients, partially due to the high variability between subjects, and even within a subject between different regions. Therefore, validating new outcome measures using post-mortem samples is critical (111). In MRI, MS abnormalities can be divided visually into 3 types: focal lesions, DAWM and NAWM. While each tissue type is associated with different degrees of tissue injury (112), the integrity of myelin and axons is not detectable using conventional MRI methods. Quantitative assessment of the severity of MS pathology using MRI has important implications both for the treatment and prognosis of MS patients. In the literature, numerous reports have demonstrated the potential of quantitative MRI methods: magnetization transfer imaging, myelin water imaging, diffusion-weighted imaging, and magnetic resonance spectroscopy (113). Based on the relaxation property of hydrogen molecules, calculations of magnetization transfer ratio and myelin water fraction have shown the sensitivity to changes in myelin content, although correlations are also found between these measures and axonal degeneration, myelin fragments, and microglial activation (114,115). Diffusion-based measures demonstrate the ability to detect the integrity of tissue microstructure by assessing the diffusing activity of water molecules (116). 40

53 Increases in the mean diffusivity of DTI in MRI-visible lesions and NAWM of MS patients suggest loss of axonal integrity (117,118). Similarly, the reduction in DTI fractional anisotropy in white matter indicates disease progression in MS patients (119). Using magnetic resonance spectroscopy, reduced N-acetyl aspartate was identified in the NAWM of MS patients with high disability, suggesting significant loss of neuronal integrity (120,121). Despite these promising results, these methods require additional acquisition protocols and are not used routinely in clinical practice in MS and many other diseases. Given the sensitivity and clinical relevance of conventional MRI, advanced analysis of such images may provide an alternative approach. Based on prior research, unique signal intensity patterns in MRI create unique spatial frequencies and are associated with specific tissue structure (88,122). The FT is an efficient image analysis method for assessing spatial frequencies, particularly regarding the alignment of a tissue. After polar conversion of the FT, image pixels of an anisotropic structure can cluster into directions perpendicular to the tissue alignment, and the extent of ellipticity of cluster of those pixels is determined by the degree of tissue anisotropy. According to prior evidence, evaluating Fourier frequencies is simpler and much more informative than directly evaluating the image of tissue structure (123). Using histological images stained with myelin, Begin et al (70) have demonstrated that FT-based method is able to detect the orientation and integrity of both healthy and degraded myelin in mouse spinal cord. In addition, Bayan et al (71) have successfully used FT power spectral to estimate the directionality and local organization of collagen fibers obtained with second harmonic generation imaging, similar to what is shown by Kim A et al (100) using this approach. Moreover, a new study by Nazaran et al (97) has shown that a semi-automatic technique based on the FT can help identify 41

54 the directionality of nerve fibers in human brain stained in histology. However, studies on direction comparisons between FT-associated outcomes in MRI and quantitative histology are scarce, and this can be an important step towards the development of new imaging outcome measures for patients. In this study, we used post-mortem MR images to evaluate the validity of a FT power spectrumbased method for assessing the severity of myelin and axonal injury in the brain of MS patients. This was done by assessing the alignment and complexity of a structure using clinical T2- weighted MRI in 3 tissue types: focal lesions, DAWM, and NAWM. The spectral results from MRI were compared with the directional outcomes from correspondent histological images quantified using structure tensor analysis. 4.2 Materials and methods Sample Post-mortem brain samples were obtained from 3 subjects with progressive MS. These samples were previously approved for use in another study in collaboration with UBC (PI: Dr Wayne Moore, neuropathologist) (124). Written informed consents were obtained from each individual prior to death for use of their tissues for medical research. The brain samples were first sectioned in a coronal plane to 10-mm-thick slices for lesion screening using MRI. Chosen areas of interest were cut, fixed in formalin solutions, and then embedded in paraffin. This resulted in 10 fixed brain sections for further study using high-resolution MRI, image analysis, and histology. Detailed information regarding the sample can be found in a published report for a different study (124). 42

55 4.2.2 MR imaging Protocol All images were acquired using a high-resolution MRI system from the fixed brain sections. MRI protocols included a multi-echo T2 sequence with field of view = 6 cm 2 ; matrix size = 256 x 256; TR/TE = 1500/7 ms for 32 echoes, and slice thickness = 1 mm (124). For subsequent image analysis, we focused on the 10 th echo (TE = 67 ms) that was equivalent to typical T2-weighted images. Regions of interest (ROIs) were identified using selected T2 images and then confirmed with correspondent histology for both location and tissue property. ROIs included 3 types: focal lesions, DAWM, and NAWM (Fig. 4.1). DAWM in MRI was defined as areas with the signal intensity higher than NAWM but lower than focal lesions. Fig : Example ROIs in high-resolution images. Upper panel shows examples of T2- weighted MR image, myelin-stained image and axon-stained image from a post-mortem brain sample of MS patient, with ROIs in focal lesion (maroon), diffusively abnormal white 43

56 matter (DAWM, blue) and normal appearing white matter (NAWM, orange). Lower panel shows the zoomed view of the ROIs Directional measurement in MRI Image analysis was done for all ROIs associated with each tissue type. Briefly, image processing steps include the following: 1) calculation of FT and the corresponding power spectrum; 2) Spectral thresholding and normalization; 3) generation of the orientation profile per ROI; and 4) determination of the dominant orientation and calculation of angular entropy, a measure of tissue complexity. Computation of FT and power spectrum were done using ImageJ (NIH, USA). This required a square ROI that was placed within the territory of the original, histology-determined region for each tissue type (Fig. 4.1). Follow-up steps were conducted using an in-house software written in Java (Oracle Corporation, USA). The FT power spectrum was normalized logarithmically to minimize noise and to enhance the reliability of the thresholding that allowed to preserve the upper twenty percentile of the spectra (125). After subsequent polar conversion, an orientation profile showing the distribution of all aligning angles of the structure in a ROI was derived from the thresholded spectrum. From this spectrum, further outcome measures were obtained. The peak height in the distribution represented the strength of alignment at an individual angle (Fig. 4.2), and the location of the highest peak was defined as the angle of dominant orientation for a ROI. Notably, the dominant orientation assessed in the frequency domain was 90 off from the tissue alignment shown in the image domain based on the theory of FT. For easy comparison, the frequency domain orientations were flipped by 90 to generate the alignment profile of a tissue structure. Angular 44

57 entropy was then calculated from the orientation profile using the following equation: ε = -Σ pθ * {log pθ}, where pθ referred to the probability of the alignment at a certain angle θ (46). Outcomes of angular entropy were inversely proportional to the coherency of a tissue, so higher angular entropy was expected in focal lesions than in NAWM. Fig : Methodology based on MR images. (A) Example ROI in NAWM in a brain slice with zoomed view of the selected ROI; (B) Fourier transform of the selected ROI; (C) Normalized FT power spectrum of the ROI; (D) Thresholded FT power spectrum of the ROI. Thresholded image shows one dominant orientation at around 135 in NAWM. 45

58 4.2.4 Histological preparation After MRI, the brain samples were prepared for histological analysis. Guided by the imaging plane of these tissue samples, a series of 10-µm-thick sections were cut for each sample, and then stained with luxol fast blue (LFB) for assessing myelin and Bielschowsky for axons (see Fig. 4.1). Stained histological images were then digitized and registered with the correspondent MR images to ensure evaluation of the same tissue across imaging modalities (124). The location of ROIs in histology were determined by a neuropathologist with reference to the matched MR images. Based on the staining density of myelin and axons, the type of ROIs was defined as following: complete loss of myelin and axonal staining (lesions); partial loss of such staining (DAWM), and preservation of each staining (NAWM). Fig : Methodology based on histology images. Method demonstration in myelin- and axon-stained images based on structure tensor analysis. 46

59 4.2.5 Directional measurement in histology Structure tensor analysis was performed to obtain the directional metrics from histological images to validate MRI findings. This is a new method for assessing tissue directionality with proven evidences in histology (84). Prior to further processing, the original color histological images were converted to 8-bit grey scale images using a weighted color conversion procedure, which took the weighted sum of red, green, and blue channels at a ratio of : : (126). Structure tensor analysis was applied to these gray scale images with the following orientation outcomes: coherency, energy, and orientation maps, and the histogram distribution of alignment angles. The directional maps were first calculated from the high-resolution histological images and were then downsized to 256 x 256 pixels to match the resolution of MRI using a B-spline-based downsampling method. ROI values were then extracted from the downsized maps. The angular histogram was computed at a ROI basis, and that became the orientation profile of the ROI (Fig. 4.3). Angular entropy was calculated from the orientation profile using the same equation as that used for MRI and also represent the tissue heterogeneity (Fig 4.4). To correspond with the MRI measures, we focused on 2 outcomes from histology: orientation and angular entropy. The outcomes of structure tensor analysis such as dominant orientation and coherency can be visualized as an ellipse. The dominant orientation is represented by direction of major axis and coherency is represented by its aspect ratio (Fig. 4.5). 47

60 Fig : Calculation of dominant orientation and aspect ratio in ROIs. Left panel shows the original images with ROI in focal lesions (red), DAWM (blue) and NAWM (orange) in myelin- and axon-stained images, the central column shows the greyscale converted images and the right column shows the dominant orientation and aspect ratio for each ROI calculated by structure tensor analysis. The e llipticity of each of the ROI represents the degree of anisotropy. 48

61 Fig : Example results of orientation profiles of tissue ROIs. Examples of orientation profile of the FT power spectrum of focal lesions (red), DAWM (blue) and NAWM(orange) in T2-weighted image (left panel), myelin-stained image (middle panel) and axon-stained image (right panel). 49

62 4.2.6 Statistical analysis A mixed effect modeling was used to evaluate differences in tissue alignment and angular entropy in both MRI and histology, and the relationship between MRI and histology measures. To evaluate the effects of myelin and axons on the variance of MRI angular entropy, preliminary regression analysis was also done. The adjusted R 2 values associated with myelin and axons were considered the percentage of variances explained by respective pathology. All of the assessments were conducted using a statistical package Stata (StataCorp, Texas, USA; Version: 12), where p-value < 0.05 was defined as significance. 4.3 Results Sample characteristics A total of 88 ROIs were examined from 10 brain samples, 34 from focal lesions, 16 from DAWM, and 38 from NAWM. In each sample, 6 to 15 ROIs were examined, which included 1 to 5 lesion ROIs, 1 to 4 DAWM ROIs and 2 to 6 NAWM ROIs. All 3 types of ROIs were identified in 7/10 brain samples; no DAWM regions were present in the rest 3 samples. (Table 1) Overall, the size of examined ROIs ranged from 7 x 5 to 27 x x47 voxels, and all of them were located within the border of each tissue type. 50

63 Table 4.1 : The number of ROIs examined in each tissue type per sample. Sample Lesion DAWM NAWM Total Total Differences in orientation metrics between tissue types in MRI and histology Based on the mixed-effect modeling analysis, we found that the angular entropy was significantly different between tissue types in both MRI and histological images (Table 2). In MRI, the angular entropy was higher in lesions (p = 0.03) and DAWM (p = 0.03) than in the NAWM. There was a trend for lesion angular entropy to be higher than DAWM, but the difference did not reach significance (p = 0.07). Similarly, in histology, the angular entropy was also higher in lesions (p = 0.02 for myelin and p = 0.01 for axonal images) and DAWM (p = 0.02 for myelin and p = 0.02 for axon images) than NAWM, and the lesions tended to be higher than DAWM (p = 0.08 for myelin and p = 0.07 for axonal images) (Fig. 4.6). For measures of 51

64 orientation, the orientating direction of each ROI types was shown in Table 2, but we primarily focused on the presence of dominant angles because they represented more of tissue coherency than angle directions. For instance, as shown in Fig. 4.4, the direction of the major axis of the resultant ellipse represented the dominant orientation and the aspect ratio represented the coherency. The aspect ratio in the resultant ellipses for each ROI are highest in the focal lesions and the lowest in the NAWM representating the amount of tissue integrity. Fig : Angular entropy outcomes in both MRI and histological images. Angular entropy results calculated from three tissue types in (A) MR, (B) myelin- and (C) axon-stained images. These results show that the angular entropy in all 3 images was significantly higher in focal lesions than in DAWM and NAWM. 52

65 Table 4.2 : The mean (standard error) of orientation outcomes per tissue type in MRI and histology. MR images Myelin-stained images Axon-stained images Angular entropy NAWM -7.22(0.65) (2.52) (2.33) DAWM -2.81(0.23) -3.81(0.33) -2.46(0.31) Focal lesion -2.14(0.26) -2.56(0.27) -1.55(0.17) Overall -4.45(0.4) -8.77(1.3) -6.63(1.16) NAWM (6.04 ) (3.1 ) (2.99 ) Orientation DAWM 3.19 (7.12 ) (1.87 ) (2.53 ) Focal lesion (6.28 ) -8 (3.81 ) (3.23 ) Overall (0.98 ) (0.47 ) (0.33 ) Note: The orientation values are measured in degrees ( ) Correlation between MRI and histology in dominant orientation To investigate how tissue orientation derived from MRI relates to that from histology, we conducted a multi-level regression analysis based on mixed-effect modeling. After taking account of the variances between brain samples and across ROIs within a sample, we discovered a strong correlation between measures of dominant orientation (Fig. 4.6). Specifically, the correlation of MRI-based dominant orientation with that from myelin-stained images was: r = 0.81; p = and with axon-stained images was r = 0.84; p = With either myelin- or axon-stained images, the correlation with MRI dominant orientation appeared to be stronger in NAWM (r = 0.86; p = for myelin; and r = 0.86; p = for axons) and lesions (r = 0.81; 53

66 p = for myelin; and r = 0.86; p = for axons) as compared to DAWM (r = 0.75; p = for myelin; and r = 0.75; p = for axons). Fig : Relationship in dominant orientation between MRI and histology. Panel A shows correlation between FT power spectrum-based dominant orientation in T2-weighted MRI and structure tensor-based orientation in myelin-stained images along with correlation coefficient (p < 0.05). Panel B shows similar plot between MRI and axon-stained images. Both plots show significant correlation between MR and histology images Correlation between MRI and histology in angular entropy To further understand the potential of FT power spectrum-based analysis of tissue integrity, we assessed the relationship in angular entropy between MRI and histology. Overall, there were significant correlations between MRI angular entropy and that of myelin (r = 0.76; p = 0.006) and axons (r = 0.69; p = 0.005), after considering for the variances between sample and ROIs. Specifically, the correlation was the highest in NAWM (r = 0.86; p = for myelin and r = 54

67 0.86; p = for axonal images), as compared with lesions and DAWM (p = 0.06 to 0.09; Table 3). Moreover, based on the correlation analysis, the estimated contribution of demyelination and axonal loss to the variances of MRI angular entropy was about 57% and 46% respectively, when all samples were considered. Regarding each tissue type, the relationship in lesions (67% for myelin and 74% for axons) appeared to be slightly higher than in DAWM (57% for both myelin and axons) (Fig. 4.7). Fig : Relationship in angular entropy between MRI and histology images. Panel A shows correlation between FT power spectrum-based angular entropy in T2-weighted MRI and structure tensor-based angular entropy in myelin-stained images along with correlation coefficient (p < 0.05). Panel B shows similar plot between MRI and axonstained images. Both plots show significant correlation between MR and histology images. 55

68 Table 4.3 : Correlation coefficients in dominant orientation between MRI and histology. Myelin-stained images Axon-stained images Focal lesion 0.81* 0.85* DAWM 0.75* 0.75* NAWM 0.86* 0.86* Overall 0.81* 0.84* Note: The stars indicate significant values (p < 0.05) between each pair of tissue types. 4.4 Discussion In this study, we evaluated the potential of using standard T2-weighted MRI to evaluate myelin and axonal integrity as shown in post-mortem brain samples. Through direct comparison with quantitative outcomes of histology, we show that the orientation profile of MRI measured by FT power spectrum-based method is strongly correlated with that measured by structure tensor analysis in histology. Moreover, the pattern of differences between tissue types in angular entropy measured in MRI is highly consistent with the pattern quantified in histology. It appears that demyelination contributed more than axonal loss to MRI entropy of lesions and the lesionfree areas. These findings suggest that mathematical measures of tissue integrity in conventional MRI may be a useful method for estimating the severity of MS pathology. The potential of image processing techniques for assessing subtle changes in tissue structure have been repeatedly shown previously. As compared to other such methods (127), the FT power spectrum is unique in its ability for extracting relatively intuitive information from a complex image domain, beyond the performance of direct FT (128). In this way, the distribution pattern of 56

69 spatial frequencies associated with a structure can be strengthened by a power of 2, allowing for considerably enhanced characterization (see Fig. 4.2). In this study, we focused on variable sizes of ROIs, with many similar to the size of MS lesions. The orientation profile of each ROI was successfully identified. Previously, the utility of FT power spectrum-based analysis was mainly focused on high-resolution microscopic images. This includes studies using collagen fibers (123,129,130), and more recently, nerve fibers (97). In the current study, with the assistance of several image post-processing steps, we showed the possibility of characterizing tissue orientation and complexity using standard MR images. Validation of MRI outcomes is critical both for improving the accuracy of data interpretation and for clinical applications. In this study, we used a novel image analysis method, structure tensor analysis, to derive quantitative histological measures for verifying our MRI metrics. According to the literature, structure tensor analysis is specifically designed to characterize the coherency and anisotropy of a tissue, and is particularly useful for assessing histological images (131). The advantage of post-mortem study is to have direct comparison between MR outcome and histology properties. The main concern of using post-mortem samples comes from tissue shrinkage due to formalin fixation. However, the overall pattern of the tissue structure remains the same. Using similar concepts to the FT, structure tensor analysis can generate a variety of orientation indices including the dominant direction, and the distribution of enclosed orientations in a ROI. Previously, this method has demonstrated great success in evaluating nerve fiber alignment and anisotropy in the white matter of rat brain using Nil-red stained histological images (84). Then, this method has proved to be a promising approach for validating DTI outcomes such as fractional anisotropy in animals. Based on the orientation profile of a structure, 57

70 in the current study, we also calculated another new parameter, angular entropy, to further quantify the complexity of a tissue following injury in both MRI and histology. Our histology results show that angular entropy has the potential to detect increased demyelination and axonal loss as seen in focal lesions, consistent with a prior study studying myelin content in mouse spinal cord using this outcome (46). Using post-mortem brain samples of MS patients in the present study, we found that there were strong correlations between structure tensor- and MRImeasured outcomes. Measures of dominant orientation show significant correlations between MRI and histology, suggesting the feasibility of using FT power spectrum in standard MRI. Following the generation and polar conversion of the orientation profile of a ROI, the location and strength of dominant orientations become distinct in MRI. Since this measure identifies the organization pattern of image pixels, which in turn are associated with the aligning property of the underlying tissue structure, measurement of dominant orientation provides a virtual marker of tissue alignment, the alignment of nerve fibers specific to this study. As NAWM was identified in areas with intact myelin and axons, this tissue type showed the highest correlation (Fig. 4.7). The relatively low correlation in DAWM as compared to focal lesions may be due to the known heterogeneity in the pathology of this tissue (132), the smaller sample size than other tissue types, and our definition of lesions in this study. Here, all lesions were exclusively identified as complete loss of myelin and axons, likely indirectly increased their uniformity. Collectively, the dominant orientation measured by FT power spectrum corresponded strongly with the actual orientation of myelin and axons in histology. 58

71 Based on the probability of individual aligning directions characterized in an area, we further examined the angular entropy of tissue structure. As an advanced measure of tissue regularity, angular entropy provides information on the complexity of a tissue, particularly regarding the coherency and anisotropy of a structure in this study. Previously, assessment of alignment entropy was mainly attempted using diffusion-weighted MRI data that readily highlights the inequality of alignment in tissue organization (133). Indeed, in a recent study of living patients following traumatic brain injury, researchers have shown that measures of diffusion entropy are more sensitive to changes in axonal density than fractional anisotropy in DTI (134). In conventional MRI, the inherent anisotropic feature of nervous system provides a structural basis for assessing tissue entropy. In the current study, we show that angular entropy calculated using T2-weighted MRI is significantly correlated with that derived from histology. Myelin and axonal entropy measures the amount of myelin and axonal heterogeneity respectively. Moreover, it seems that myelin entropy explains more of the variance to MRI entropy than axonal entropy. This observation is consistent with what has been shown in the literature (46). In an animal study, it has been found that angular entropy is strongly associated with the integrity of myelin, with the highest entropy in tissues with complete demyelination. These evidences suggest that the degree of angular entropy reflects the integrity of myelin and axonal pathology. There are a few limitations worth to mention in this study. In assessing the relationship between MRI-measured tissue orientations and those in histology, we focused only on myelin and axonal property. While this may have omitted the interference of other pathology such as inflammation, prior studies have shown that inflammation explained the least variance of MRI pixel regularity compared to myelin and axons (124), which in combination seemed to have explained over 90% 59

72 of the variance of MRI angular entropy based on our regression analysis. In addition, the measure of tissue alignment is based on the neighbouring relationship of MRI pixels. Therefore, the MRI alignment of a tissue structure is a measure of the average directionality of nerve fiber tracts, instead of individual axons. This mismatch in pixel sizes has been a known limitation in MRI-histology studies (135). Additionally, the minimum size of ROIs in MRI was limited to about 35 pixels. However, the size range of ROIs used in this study represented the various sizes of MS lesions, and the pathological changes in DAWM were clearly detected in both MRI and histology. Moreover, the directional information measured by structure tensor analysis is based on the assessment of eigenvalues and eigenvectors of an image (136). The eigenvalues, however, can be subject to noise, affecting the consistency of orientations (137). Hence, the ROIs with extremely low coherency need to be carefully considered. In the future, we attempt to compare the tissue alignment measurement method with other quantitative MR approaches such as diffusion tensor imaging and further validate the contribution of other tissue pathologies involved related to MRI signal intensity. 4.5 Conclusion Using post-mortem brain samples from MS patients, we evaluated the validity of our new imaging measures of tissue coherency. Findings in this study demonstrate that myelin and axonal integrity of MS lesions can be identified using conventional MRI using simple steps of mathematical calculation. Therefore, assessing lesion severity using this method may help enhance the characterization of disease progression and thereby optimize the clinical management of patients including the selection of targeted therapy. With further validation, the utility of these new MRI metrics can also be extended to other neurological diseases beyond MS. 60

73 4.6 Acknowledgement This research was supported by funding from the Natural Science and Engineering Council of Canada (NSERC), MS Society of Canada, and Alberta Innovates- Health Solutions. This investigation was supported in part by a studentship from University of Calgary (Queen Elizabeth II Graduate Scholarship) and NSERC I3T CREATE program awarded to S. Sharma. 61

74 CHAPTER 5: COMPARISON OF FT POWER SPECTRUM-BASED METHOD WITH DIFFUSION TENSOR IMAGING 5.1 Introduction Diffusion-weighted imaging is one of the main MRI methods used to evaluate tissue integrity through assessment of the alignment and anisotropy of tissue structure. Represented by DTI, diffusion-weighted MRI has shown great promise for detecting invisible structural changes in brain white matter and has been used in many neurological diseases including MS. To further validate the potential of my new method, this chapter aims to compare the FT power spectrumbased outcomes with those from DTI. Diffusion-weighted images are based on the theory of Brownian motion of water molecules. Brownian motion, or thermal molecular diffusion, is any type of random movement of the fluid disturbed by the thermal energy. In a homogenous fluid, the diffusion is equal in all directions i.e. isotropic. However, in a restricted medium, such as in certain biological tissues the diffusion is anisotropic (138). The diffusion coefficient characterizes the rate at which fluid molecules spread out and the diffusion-weighted MRI provides a means of measuring these changes in biological systems. In traditional diffusion-weighted MRI, tissue diffusivity is measured as a summary termed as apparent diffusion coefficient (ADC), which is calculated by collecting a series of diffusion-weighted images. Newly developed diffusion techniques have the potential to further probe tissue characteristics. Several options exist for acquiring diffusion-weighted images in the literature. DWI can be acquired by adding two extra diffusion sensitizing gradients with opposite directions to a spin- 62

75 echo T2-weighed MR sequence (139). If the water molecules have net movement, the signal intensity should be proportional to the net displacement in the specified direction. Water molecules with a net movement of zero do not produce diffusion signal. In addition, echo planar imaging can also be used for diffusion acquisition and can help reduce imaging time and minimize motion artefacts (44). The signal intensity in diffusion-weighted images is also dependent upon the strength of the applied diffusion gradient pulses, which is parameterized as the b-value. The relationship between b value and gradient strength can be seen from the following equation (1). The b0 images are obtained when the diffusion gradient is turned off. These b0 images (S0, see equation below) are similar to T2-weighed images in signal intensity and are often used as anatomical basis when comparing with other imaging contrasts S DWI = S 0 exp b.adc (1) DTI is proposed to improve the characterization of tissue diffusion. It is modeled in a 3- dimensional (3D) space as an ellipsoid or orientation distribution function (140), or as a 3 x 3 matrix mathematically as shown in equation (1) below that reflects the covariance of diffusion displacement in the space (116). D xx D xy D xz D = [ D yx D yy D yz ] (2) D zx D zy D zz This matrix can be diagonalized to evaluate the eigenvalues ( 1, 2, 3 where 1> 2 > 3) and the corresponding eigenvectors (V1,V2,V3), providing the magnitude of ADC in each direction (141). 63

76 In particular, 1, also known as L1, represents the largest magnitude of diffusion in a spatial direction (x, y or z), and V1 represents the direction having the largest diffusion. The diffusionweighted signal has the highest attenuation in the direction of dominant fiber orientation (V1). DTI produces several parameters, namely: mean diffusivity (MD), radial diffusivity (RD), axial diffusivity (AD) and fractional anisotropy (FA), each reflecting a certain aspect of structure property as reflected in the following equations (2-5). These outcomes enables the mapping and visualization of white matter tracts and also assists in localizing structure disruption in nerve fibers (142). DTI is the simplest model that encapsulates the anisotropic diffusion based on the organization of the white matter. MD = λ 1+ λ 2 + λ 3 3 (3) RD = λ 2+ λ 3 2 (4) AD = λ 1 (5) FA = 3{(λ 1 MD) 2 +(λ 2 MD) 2 +(λ 3 MD) 2 } 2{λ 1 2 +λ 2 2 +λ 3 2 } (6) The relationship between DTI and tissue pathology has been shown in numerous studies previously including those in MS ( ). DTI can detect subtle changes in different grey and white matter regions, which might vary across MS phenotypes. MS plaques have shown increased diffusion, with acute plaques having significantly higher diffusion than the chronic 64

77 plaques, suggesting changes in the size of extracellular space due to demyelination (146). The radial diffusivity has been observed to increase with demyelination (147). FA is a measure of tissue anisotropy and has been used significantly for measuring pathological changes in MS (Fig. 5.1) (148). FA is a rotationally invariant measure of the magnitude of diffusion with values 0 to 1, higher values mean greater anisotropy in diffusion. Previous studies have demonstrated that FA is lower in MS lesions than control tissue and is also sensitive to NAWM abnormality (117, ). There are several novel diffusion-weighted imaging methodologies developed recently that may improve the performance of DTI. These include high angular resolution diffusion imaging, neurite orientation dispersion and diffusion imaging, and diffusion basis spectrum imaging ( ). However, these methods have not been tested extensively in a clinical setting. Therefore, my current study is focused on the DTI. Fig : Different ellipticity representing different FA. Images from left to right represent increasing ellipticity and increasing FA. The image on the left end has the lowest FA and the image on the right end has the highest. 65

78 The purpose of this study is to evaluate the correspondence between tissue alignment and complexity as measured by FT power spectrum-based method and DTI outcomes. Specifically, measures of dominant orientation and angular entropy derived from my method were compared with the principle orientation and FA from DTI. This was done using the corpus callosum, given its high coherent nature. It was hypothesized that the dominant orientation and angular entropy calculated in T2-weighted MRI using my method correlate with the principal direction and fractional anisotropy of DTI. 5.2 Materials and methods Subjects and samples This data is the part of the protocol used in Chapter 3 (See Acknowledgments in both Chapters). In a previously completed study of MS patients, 19 MS patients including 10 RRMS and 9 SPMS patients and 19 controls were recruited. The original purpose of that study was to evaluate the structure and function of corpus callosum in relation to neuropsychological measures. All patients provided written informed consents for image analysis. I took the advantage of that study and focused my analysis on the corpus callosum of the patients. To ensure consistency and reliability of the measurements, corpus callosum regions showing extreme atrophy or with focal lesions in the structure were excluded MR acquisition All MRI scans were conducted with a single 3T MR system (GE Healthcare, DISCOVERY MR 750, Milwaukee, USA). Axial DTI was performed using a spin-echo echo-planar imaging sequence with the following parameters: TR/TE = 10000/89.5 ms, matrix size = 256 x 256, field 66

79 of view = 240 x 240 mm 2, slice thickness = 3 mm, no gap. Diffusion-weighting was applied in 23 directions, b= 1000 s/mm 2 with 5 b0 volumes. For T2-weighted images, the parameters were: TR/ TE = 6035/83 ms, field of view = 240 x 240 mm 2, matrix size = 256 x 256 and the slice thickness = 3 mm, no gap DTI analysis The DTI data were corrected for motion and eddy-current distortions using tools provided in FSL (156). Outcome measures included FA, V1, V2, V3, and L1, L2, and L3 maps, representing the degree of anisotropy (FA), the major (V1) and minor (V2, V3) orientation, and the corresponding major (L1) and minor (L2, L3) strength of the diffusion within a voxel. For comparison of tissue alignment and integrity with the FT power spectrum-based analysis, I primarily studied FA and V1, L1 maps. Six regions of interest (ROIs) sized 6 x 6 pixels were selected from each corpus callosum, located in the left, centre and right of genu and splenium respectively. ROIs were initially selected in T2-weighted images and then matched to the corresponding maps (Fig. 5.2). Average values were calculated for each ROI and parameter. 67

80 Fig : ROIs selection in T2-weighted image and corresponding maps. (A) Top left panel shows six ROIs in left, center and right genu and splenium in T2-weighted image. (B) Top right, (C) bottom left and (D) bottom right panel shows the corresponding ROIs on the FA, L1 and V1 maps Tissue integrity measurement in conventional images Tissue integrity was measured in T2-weighted images by using the image processing algorithm developed in Chapter 3 using the same ROIs chosen in corpus callosum. To summarize: image ROIs were selected and converted to the frequency domain; the power spectrum was calculated, the angular distribution profile of the power spectrum was extracted; and finally, the dominant 68

81 orientation and angular entropy were computed. Angular entropy (ε) was calculated as shown in the equation below (6), where pθ is the probability of alignment at a certain angle θ. The ROIs used in T2-weighted MRI were manually matched to the corresponding DTI slices. The V1 and FA values in DTI were used for comparison with dominant orientation and angular entropy respectively from T2-weighted MRI. For ROIs with significant correlation between V1 and dominant orientation, the relationship between L1 and orientation strength was also evaluated. = p θ {log p θ } (6) Statistical analysis Mixed-effect modelling in Stata was done to study the correspondence of V1 maps from DTI and the FT power spectrum-based dominant orientation, and of FA maps and the angular entropy. For the ROIs having significant correspondence between FT dominant orientation and V1; similar method was used to evaluate the relationship of orientation strength with L1. All analyses were conducted using Stata (StataCorp, Texas, USA; Version 12). For the given analysis, p < 0.05 was considered statistically significance. 5.3 Results Sample characteristics A total of 114 ROIs were studied from the patients, 60 from RRMS and 54 from SPMS patients. In total, there were 16, 14 and 17 ROIs from left, centre and right genu and 15, 13 and 16 ROIs from left, centre and right splenium, respectively. The total number of corpus callosum ROIs used in this study are listed in Table 1. Overall, 23 ROIs were excluded due to extreme tissue 69

82 atrophy: 3, 5 and 2 from right, centre and left genu and 4, 6 and 3 from right, centre and left splenium. Similarly, 114 ROIs were studied from 19 controls; 6 ROIs per subject, located at corresponding areas to those in matched patients. Table 5.1 : Number of ROIs analyzed in the corpus callosum. Patient Control Total Left Genu Centre Genu Right Genu Left Splenium Centre Splenium Right Splenium Correspondence between FT power spectrum angular entropy and FA Using T2-weighted MR images, FT power spectrum-based angular entropy was calculated in six sets of ROIs in corpus callosum. A significant overall correlation was observed between angular entropy and FA in patients (r= ; p = 0.008), and in controls (r= ; p = 0.002). (Fig. 5.3). The correlation in SPMS (r= ; p = 0.009) appears to be slightly lower than in RRMS (r= ; p = 0.011). Moreover, the angular entropy was found to be significantly higher in central genu and splenium (p = 0.015) than other regions of corpus callosum, and this was correspondent to decreased FA in the same two ROI regions. Specific values of FA and angular entropy in each of the ROIs are shown in Table 2. 70

83 Fig : FA outcome in 6 sets of ROIs in corpus callosum. Bar graph shows the distribution of FA values (mean ± standard error) in each location of ROIs in patients and controls. Fig : Correspondence of FA with angular entropy. Scatterplot shows correlation between FT power spectrum-based angular entropy and average FA in (A) patients (RRMS, SPMS) and (B) controls. 71

84 Table 5.2 : Mean (± standard error) FA in DTI and angular entropy in T2-weighted MRI. Fractional Anisotropy Angular Entropy Patients Controls Patients Controls Left Genu 0.43(±0.02) 0.63(±0.02) -5.51(±0.41) (±0.41) Centre Genu 0.31(±0.02) 0.39(±0.01) -2.41(±0.49) -6.31(±0.31) Right Genu 0.50(±0.02) 0.63(±0.02) -5.38(±0.28) (±0.53) Left Splenium 0.49(±0.02) 0.67(±0.02) -6.52(±0.53) -11.5(±0.49) Centre Splenium 0.32(±0.03) 0.4(±0.01) -3.58(±0.46) -6.01(±0.20) Right Splenium 0.49(±0.02) 0.67(±0.02) -5.62(±0.28) (±0.41) Correspondence between dominant orientation from FT power spectrum and principle direction of DTI In comparing the major orientation of tissue alignment between DTI (V1) and FT power spectrum-based method (dominant orientation), Mixed-effect modelling method was used. Overall, only a trend for significant correlation was found (p = 0.079) when all ROIs were considered. The mean (± standard error) of principle direction V1 in DTI was found to be (±0.045), (±0.038) and 0.5 (±0.053) in left, central and right genu and (±0.039), (±0.031) and (±0.030) in left, central and right splenium. Similarly, the mean (± standard error) degree of dominant orientation generated from T2-weighted MRI using FT power spectrum was (±1.623 ), 3.31 (±1.583 ) and (±3.415 ) in left, central and right genu, and (±1.202 ), 6.46 (±1.674 ) and 50.5 (±1.848 ) in left, central and right splenium respectively. 72

85 5.3.4 Correspondence between FT-based orientation strength and dominant diffusivity Finally, mixed-effect modelling was used to study the relationship between orientation strength L1 of DTI and the strength of dominant orientation from FT power spectrum-based method. The mean (± standard error) orientation strength from FT power spectrum-based method and highest diffusivity L1 from DTI are shown in Table 3. Significant correlation was found between orientation strength and dominant diffusivity in patients (r=0.57; p=0.036), and in controls (r = 0.69; p=0.023). Fig : Correspondence of orientation strength with highest diffusivity. Panel (A) shows correlation between FT power spectrum based-orientation strength and highest diffusivity in patients and panel (B) shows similar correlation in controls. 73