University of Southern California Dissertations and Theses

Cerebrovascular disease of white matter in patients with chronic anemia syndrome

(USC Thesis Other)

Cerebrovascular disease of white matter in patients with chronic anemia syndrome

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Cerebrovascular Disease of White Matter in Patients with Chronic Anemia Syndrome by Yaqiong Chai A Dissertation Presented to the FACULTY OF THE USC GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (BIOMEDICAL ENGINEERING) December 2019 2 Dedication This work is dedicated to my mother and my father, for supporting and helping me selflessly always, and Eric Chen, who gives me joys and frustrations, yet makes me a better mom. 3 Acknowledgements Once we accept our limits, we go beyond them. Albert Einstein (1879-1955) No work of this large could have been completed without the support from many people who have crossed my path, but most importantly, being a part in my personal and professional development as a scientist. To John and Natasha: I am so blessed to have both of you as my advisors. You helped and encouraged me at each single step of the research and made me not only a more independent scientist, but also a more professional and skilled person. To Natasha: You first took me in when I was looking for a lab to transfer my Ph.D. program. Therein you opened a window of how to perform scientific research by simply start from knowing the data, process them and compare the results. Thank you for giving me the opportunity to start a new journey at Children’s Hospital Los Angeles (CHLA), and holding my hand and walking me step by step, from a conference abstract to a full journal paper, from a shy, goalless conference participant to a sophisticated conference presenter, from a qualifying proposal to a Ph.D. thesis. Thank you for supporting and cheering my up when I felt lost or unhappy with my life. I am lucky to have an advisor who cares not only her students’ research progress, but also their well-being. Thank you for encouraging me to overcome the difficulties with your enthusiasm, for making me feel proud of my research with your passion, and for pushing the limits and believing me more than I do. To John: As the most knowledgeable person that I have ever known personally, you imparted physiology models and pathology principles to me as a medical mentor, as well as signal processing and physics/statistical methods to me as an engineer. The most impactful lessons that you set through examples, “Learning by doing it” made me a more perseverant and more patient person. Thank you for teaching me how to be an excellent scientist by performing experiments with creativity, integrity and purpose. Thank you for pushing me to go further: think of a problem as applying for a grant, write up an essay as an independent researcher. Thank you for all those one-on-one meetings and discussions over the desk-phone, from which you enlightened me not only with research ideas but also with life strategies. Your way of balancing life and work showed me how to live a life with excellence. One cannot possibly ask for a better mentor and for which I am always being grateful. 4 To my defense and qualification exam committee. Dr. Micheal Khoo, Dr. Yonggan Shi, Dr. Danny JJ Wang, and Dr. Vidya Rajagopalan. Thank you for offering your time, support and academic advice throughout my qualifying proposal and dissertation. Thank you all for inspiring me with many questions from different perspective that I would not think of. A very special gratitude goes out to Dr. Danny JJ Wang, who generously filled the role as my indispensable qualifying committee member last minute. It could not have made possible if you would not have participated in my qualifying exam. I would also like to thank the members of my immediate research family. Yi Lao, my first lab-mate who showed me how to make the best use of one’s time. Your 4.5 year of graduation with a Ph.D. set up a hallmark of my Ph.D. timetable. Your quirky but efficient research schedule broadened my view of “working hard” and “working productively”. I enjoyed and missed your accompanying. Sinchai Tao, I thank you for being my first “debugger” and teaching me the foundation of being an engineer: patience. “Try 100 different ways till you make it.”, which I still remember and encourage myself while getting feed up with my work. I regret that we did not spend enough time when we cross paths again. Adam Bush, thank you for being the “role model” lab- mate and all the advices on my research and life. I have employed many of your words in my research efforts. I was impressed and motivated by your enthusiastic experiment time around the scanner, without any breaks at times. Botian, thank you for opening the gate of artificial intelligence in my research path. You masterly good at dissecting complex engineering models into digestible sips of information. You helped me understand complex concepts and frequently provided insights on how to make improvements. Your perfection- and detail-oriented working style will always drive me towards excellence. Your commitment to our collaborated project, all the incisive discussions and constructive suggestions have accelerated my academic pace. I enjoyed our office debates and lunch conversations and I await the time we cross paths again. academic research and how it relates to patient care and outcomes. Niharika Gajawelli, thank you for being great friends who I could always seek for advice on work-family balance. Chau Vu, thank you for making me more confident using the scanner, and I miss the times when we were roommates during the conferences. Furthermore, I would like to thank the members of my extended research family that shared all the peaks and troughs in my training: Noel Arugay, Lisa Villaneuva, Obdulio Carreras, Julia Castro, So Young Choi, Jian Shen, Julie Coloigner, Henk Jan Mutsaerts, Aart Nederveen, Sharon 5 O’Neil, Yihao Xia, Jieshen Chen, Xiaoping Qu, Yiceng Li, Chaoran Ji, Rinu Sebastian, Nathan Smith, John Sunwoo, Silvie Suriany, Bertin Valdez, Adam Walker, Toey Wanwara, and last but not least Cristina Galarza. Throughout my educational years in the U.S, my parents have been ready to jump in and help, at the cost of taking a long leave from their own work and being isolated by a different language and culture in a foreign country. They offered selfless help during my critic periods of my training: qualifying exam and my defense, for which I cannot thank you enough. Thank for taking care of Eric, my son and me when I needed you most. This thesis is dedicated to you. 6 Table of Contents Dedication ...................................................................................................................................... 2 Acknowledgements ......................................................................................................................... 3 Chapter I: Introduction ............................................................................................................... 11 Part I: Cerebrovascular disease in sickle cell disease .......................................................... 11 Section I: Sickle cell disease ............................................................................................... 11 Section II: Cerebrovascular complications ...................................................................... 12 Section III: Brain changes in SCD .................................................................................... 16 Section IV: Cerebral oxygen transport ............................................................................. 17 Part II: Magnetic Resonance Imaging in SCD ..................................................................... 24 Section I: MRI basis ........................................................................................................... 24 Section II: Arterial Spin Labeling ..................................................................................... 26 Section III: Diffusion-weighted Imaging .......................................................................... 29 Part III: Techniques: Statistics and deep learning .............................................................. 33 Section I: Automatic SCI segmentation ............................................................................ 33 Section II: Super-resolution recovery ............................................................................... 41 Chapter 2: Impaired white matter oxygen delivery in anemic patients ..................................... 45 Abstract .................................................................................................................................... 45 Introduction ............................................................................................................................. 45 Methods .................................................................................................................................... 46 Population ............................................................................................................................ 46 Data acquisition ................................................................................................................... 47 Pre-processing and WMH segmentation .......................................................................... 47 CBF quantifications ............................................................................................................ 48 Oxygen delivery maps ......................................................................................................... 49 Statistical Analysis .............................................................................................................. 49 Results ...................................................................................................................................... 50 Demographics ...................................................................................................................... 50 Discussion ................................................................................................................................ 56 Chapter 3: White matter microscopic injury is associated with neurological performance ..... 60 Abstract .................................................................................................................................... 60 Introduction ............................................................................................................................. 60 Methods .................................................................................................................................... 62 Participants .......................................................................................................................... 62 Neuroimaging ...................................................................................................................... 63 Registration and 3D representations ................................................................................. 63 Neurocognitive assessment ................................................................................................. 65 Statistical Analysis .............................................................................................................. 65 Results ...................................................................................................................................... 66 Basic clinical characteristics of participants .................................................................... 66 Tract specific analysis ......................................................................................................... 66 Discussion ................................................................................................................................ 70 7 Chapter 4: Silent cerebral infarcts stroke map in pediatric cohort: its implications and research significance ................................................................................................................... 75 Introduction ............................................................................................................................. 75 Methodology ............................................................................................................................ 76 Data Representation of LR images and Proposed Framework ...................................... 76 Experiments and Evaluation .................................................................................................. 78 Data Preparation ................................................................................................................. 78 Training Procedure ............................................................................................................. 79 Results ...................................................................................................................................... 79 Conclusion and Discussion ..................................................................................................... 80 Chapter 5: Conclusion and ongoing work .................................................................................. 81 Reference: ..................................................................................................................................... 85 8 Table of Figures Figure 1.1. The normal (A) and sickle (B) hemoglobin, red blood cell and circulation in microvasculature. Adapted from: Bioninja.com.au .............................................................. 12 Figure 1.2: A: T2-weighted MRI shows overt stroke in left anterior cerebral artery and middle cerebral artery territories; B: MR-angiogram shows occlusion of the internal carotid artery. Figure was adapted from Switzer, J. et. al 20 . ....................................................................... 13 Figure 1.3: A T2-weighted FLAIR MRI image of a patient with sickle cell disease, showing silent cerebral infarction circled in red. ................................................................................ 13 Figure 1.4: The prevalence of SCI in children with sickle cell anemia. The percentage are from four different multi-central studies and one longitudinal study 34-36 . ..................................... 15 Figure 1.5: White matter volume group comparison results, where adjusted p-values are shown on the mid-cortical surface (Positive p-values in red indicate the GM volume of SCD patients are higher than that of healthy controls. Figure adapted from Choi et al 46 . ............ 16 Figure 1.6: Oxygen hemoglobin dissociation curve relating the partial pressure of arterial oxygen (PaO 2 ) to the oxygen saturation for hemoglobin (Hb) A, shown in solid line, and for Hb S, shown in dashed line. Figure adapted from Wagner, et. al 65 . ............................................... 17 Figure 1.7: Fick principle states that the total uptake of oxygen is equal to the product of the blood flow to the arterial-venous concentration difference of oxygen. ................................ 18 Figure 1.8: A Krogh Cylinder Model can be used to model oxygen diffusion across the tissue from the vascular cylinder (red arrow). Figure credit goes to David Wootton. .................... 19 Figure 1.9: Effects of CTTH on oxygen extraction (from 1.0-0.4 shown indicated in color bar) with homogenous capillary flow velocities (B) and heterogeneous flow velocities (A). Figure was adapted from Jespersen and Østergaard 77 . ......................................................... 20 Figure 1.10: Vascular territories of the cerebral cortex in A: lateral sagittal view and B: midline sagittal view. Inferior view. Figure adapted from: Radiopaedia.org. Case courtesy of Prof Frank Gaillard. ...................................................................................................................... 21 Figure 1.11: Changes in cerebral blood flow (CBF) with respect to age. CBF values are fitted using cubic regression. Note that the horizontal axis is scaled by one-fifth for adults to better illustrate the changes in childhood. Figure is adapted from Wu, et, al 78 . ................... 23 Figure 1.12: Diagram of MR signal generation. A: polarization, generating M; B: excitation, generating M xy ; C: signal detection, yielding S(t); D: spatial encoding, generating S(k xy ); E: image reconstruction, producing p(r). Figure was adapted from Justin Haldar lecture notes (fall 2016). ............................................................................................................................. 24 Figure 1.13: PCASL labeling pulse train and gradients. In the labeling sequence (red), the magnetization is inverted during a series of slice selective flips; in the control sequence (blue), the RF profile is reversed alternatively to preserve the equilibrium magnetization. 26 Figure 1.14: Tensor shapes and their corresponding types of diffusion. Figure courtesy from: www.diffusion-imaging.com ................................................................................................ 30 Figure 1.15: Workflow of imaging preprocessing for a tissue-type prior supervised segmentation method. .................................................................................................................................. 36 Figure 1.16: The scheme of the algorithm design using k-nearest neighbor. FDA=Fisher’s discriminant analysis. ............................................................................................................ 37 Figure 1.17: A: One neuron unit and its input connections, ⍬ i is the i th neuro unit of the layer; B: Neuroal network with multiple hidden layers. ...................................................................... 37 9 Figure 1.18: Examples of activation functions. A: Linear, f(a i )=f(w T x-⍬i); B: Hard threshold, f(a i ) = -1 if x<0, f(a i ) = 1 otherwise; C: f(a i )=tanh(a i ); D: Rectified linear unit (ReLU), f(a i ) = 0 if x<0, f(a i ) = x otherwise; E: Leaky ReLU, f(a i ) = 0 if x<0, f(a i ) = 0.01x otherwise; F: Sigmoid, f(a i )=S(a i ). ............................................................................................................. 38 Figure 1.19: The architecture of CNN classification model using T2-FLAIR MR images. ........ 40 Figure 1.20: A clinical T2-weighted image from SIT Trail dataset. A: coronal view; B: sagittal view; C and D: axial view of the neighboring slices shown on A and B in yellow and blue, respectively. Resolution: 0.86 x 0.86 x 6mm 3 , dimension: 256*256*22. ............................ 41 Figure 2.21: Whole brain CBF map, oxygen delivery map and T-score maps in axial view. Top row: average CBF map of healthy controls (N=25), sickle cell patients (N=32) and T-score. The color-bar represents absolute perfusion in ml/100g/min. Bottom row: average oxygen delivery map of healthy controls (N=25), sickle cell patients (N=32) and T-score. The color- bar depicts oxygen delivery in milliliter of oxygen /100g brain tissue per minute. ............. 51 Figure 2.22: White matter lesion mask overlaid on T-score map. The scaled WM lesion mask in its axial view is overlaid on top of the T-score for CBF (top row) and O2 delivery (bottom row). The number z indicates the coordinates in the standard atlas ..................................... 54 Figure 2.23: Boxplot comparisons of PC CBF estimates (left) and relative error between PCASL and PC CBF (right) among subjects with and without venous outflow. Patients with venous outflow displayed larger absolute CBF and relative error than subjects without venous outflow (p<.05). .................................................................................................................... 55 Figure 3.1: Medial representation of the 11 white matter tracts in coronal, axial and sagittal view. Red: corpus callosum (CC), green: cortico-spinal tract (CST), orange: inferior fronto- occipital tracts (IFO), magenta: inferior longitudinal tracts (ILF), cyan: superior longitudinal tracts (SLF) and purple uncinates (UNC). ............................................................................ 64 Figure 3.3: The significant clusters of reduced FA in patients with SCD (panel A) and non-sickle anemic syndromes (panel B) compared to healthy controls (in red), overlaid on the corresponding t-statistics maps on the skeleton surfaces of 11 tracts. .................................. 68 Figure 3.3: The significant clusters of reduced FA in SCD patients with white matter hyperintensities (panel A) and patients without hyperintensities (panel B) compared to healthy controls (in red), overlaid on the corresponding t-statistics maps on the skeleton surfaces of 11 tracts. ............................................................................................................. 68 Figure 4.1: In 2D slice stacks, through-plane (axial) slices have large thickness, making the resolution much lower than in-plane resolution. Non-isotropic SR is needed to retrieve the missing slices. (b); The slices to be restored are modeled as 1-valued masks, when viewing in other two planes (coronal and sagittal), Image in-painting could reproduce fine details, especially for the complex gyrus and fissures. The edge generator hallucinates edges in the missing areas, supposing that edge recovery is easier than image completion. Then the image in-paint network combines the edges in the missing regions with texture information of the rest of the image to fill the missing areas. .................................................................. 75 Figure 4.2: An overview of the proposed method. Edge generator (in green) firstly connects the missing edges from the LR images. Then contrast generator fills the intensities based on the original contrast from LR images, guided by edges generated from the first step and supervised by HR images. Both steps follow an adversarial model. .................................... 77 Figure 4.3: An axial slice of: the ground truth, reconstructed by nearest neighbor (NN) interpolation, bicubic interpolation, low-rank total variation (LRTV), super-resolution CNN 10 (SR-CNN), super-resolution GAN (SR-GAN) and our proposed method (EGGAN) resulting image. .................................................................................................................................... 79 Table of Tables Table 1.1: common measurements used to evaluate WM lesion segmentation. Abbreviations: TP=true positives, TN=true negatives, FP=false positives, FN=false negatives ………………..33 Table 2.1: Patient Demographis. Continuous variables presented in mean ± standard deviation. Comparison of continuous variables was made using pooled, two-tailed t-tests. Comparison of categorical variables (sex) was made using a Fisher’s exact test. All the tests and p-value was the comparison between patients with SCD (N=32) and healthy controls (N=25). CTL = controls, SCD = sickle cell disease, WMH = white matter hyperintensity, Hb=hemoglobin…………………….48 Table 2.2: Global/regional CBF and oxygen delivery. Measurements presented as global (GM/WM) with mean ± standard deviation, values in bold indicates significant difference with p<0.01. CTL = controls, SCD = sickle cell disease, CBF = cerebral blood flow, ACA = anterior cerebral artery, MCA = middle cerebral artery, PCA = posterior cerebral artery………….………….…………..51 Table 2.3: Predictors of SCD brain perfusion and oxygen delivery. Bold lettering indicates retention on multivariate analysis. Abbreviations: Mean corpuscular hemoglobin concentration (MCHC), Red cell distribution width (RDW), Grey matter (GM), White matter (WM), Oxygen (O 2 ) ………….………….………….………….………….………….………….……………….52 Table 3.1: Demographical characteristics and neuropsychological performance results. Abbreviations: CTL: controls, SCD: sickle cell disease, ACTL: anemic controls, WMH: white matter hyperintensity, L/R= left hemisphere/right hemisphere. FSIQ: full-scaled IQ.†Psychometric scores are presented as scaled scores, which have a mean of 10 and a standard deviation of 3, followed by the range. ‡In group 2, neuropsychological tests are only available for 9 patients in ACTL group, therefore, they are not analyzed. *Significant difference between two groups using t-test with Bonferroni correction. ………….………….………….………….…….65 11 Chapter I: Introduction Part I: Cerebrovascular disease in sickle cell disease Section I: Sickle cell disease Sickle cell disease overview Hemoglobin (HbA), the protein that carries the oxygen in the red blood cell, comprises two beta-globin and 2 alpha-globin chains. Sickle cell disease (SCD) is defined by the sickle mutation on the beta-globin gene (HBB, Glu6Val) 1 , forming sickle hemoglobin (HbS), which produces profound alterations in the stability and solubility of the hemoglobin molecule. Genotypes of SCD are characterized by the inheritance of an abnormal gene: either the sickle cell gene from both parents (homozygosity for the HbS gene) or from one parent, along with another hemoglobin variant, such as hemoglobin C (HbC) or beta-thalassemia (compound heterozygosity). The first case, HbSS, is usually the most severe form of the disease, and is commonly called sickle cell anemia; in the second case, individuals with HbSC condition or HbS-beta thalassemia tend to have a slightly milder form of SCD. Lastly, the presence of a single healthy HbA gene, along with a HbS or HbC gene, gives rise to a benign condition, named sickle cell trait (SCT). SCD is a complex disease affected by environmental factors and genetic variability. In the U.S., more than 100,000 people have the condition, with the majority of them being of African or Hispanic American decent 2,3 . The annual health care costs are estimated at more than 1.5 billion in the U.S. 4 Polymerization When a sickle hemoglobin releases oxygen during vascular perfusion, the hemoglobin molecules (made of alpha- and beta-globin subunits) aggregate into a viscous gel composed of long multi-stranded helical polymers; this process distorts the red blood cell into a crescent, or sickle shape (Figure 1.1, B), from which the disease received its name. Upon re-oxygenation, the polymers disassemble and the red blood cell usually resume back to its normal biconcave shape 1,5 . The repeated sickling-unsickling cycles causes red cells to be more fragile and to be removed more readily from circulation. The damaged sickle cells often rupture and release their intra-cellular contents into the surrounding vasculature, and this intravascular hemolysis causes scavenge nitric oxygen and interferes with endothelial function, leading to a chronic inflammatory state 6 . The average life span of normal red blood cell is about 120 days, and while that of a sickled one is 8- 12 40 days 7,8 . The reduced numbers of red blood cells, or hemoglobin levels, is a hallmark of the disease 1,9 . Even though normal red blood cells are larger than the capillaries of the microcirculation, they are quite flexible and easily circulate through them. Polymerized HbS results in much more rigid red blood cells which can become lodged in the microvasculature, thereby increasing resistance, or worse, transiently or permanently blocking a vessel, occluding blood flow 10 . This vaso-occlusion decreases oxygen supply to the surrounding tissues. The resulting organ damage is a major cause of the morbidity and mortality of the disease, which includes pain crisis, tissue ischemia, and, subsequently, organ failure 11–13 . Cerebrovascular damage, or more concretely, stroke, is amongst the most concerning complications, due to the profound personal, professional and social costs to patients. Section II: Cerebrovascular complications Stroke In SCD, there are several forms of stroke, which, depending on the neurological consequences and severity, are categorized as either overt or silent. Overt stroke is the most severe form, and comes in two primary classes: ischemic and hemorrhagic (Figure 1.2). The former is the most common type, accounting for more than 87% of strokes in the SCD 14 . It is caused by a clot Figure 1.1. The normal (A) and sickle (B) hemoglobin, red blood cell and circulation in microvasculature. Adapted from: Bioninja.com.au 13 or blockage within a blood vessel feeding to the brain tissue. It can also happen when the oxygen delivery to the brain is insufficient even with normal blood flow 15 . Hemorrhagic stroke occurs when a large blood vessel in the brain ruptures, releasing blood into the surrounding tissues. The complications of overt stroke can be fatal, including high rates of mortality and lifelong morbidity. In the early 90s, before the development of modern intervention protocols, approximately 11% of SCD patients had an one by the age of 20, and the number grew to 24% by the age of 45 16,17 . As a result, 22% patients with SCD died of overt stroke, and half of those did so on the day it occurred 16,18 . Additionally, the risk of recurrent stroke is more than 10 times larger than that of primary stroke in patients with SCD without treatment or intervention 19 . Silent strokes, or silent cerebral infarctions (SCI), are asymptomatic clinically, but detectable through many imaging modalities. Figure 1.3 shows an example of SCIs identified by brain MRI image.They also place the patient at 50% increased risk of an overt stroke in the future 20 . It has been demonstrated that 27% of SCD patients had an SCI before 6, and 37% by 14 years old 21 . Also, patients with SCIs are 5-7 times more likely to develop new or enlarged SCIs, with or without the treatment 22,23 . SCIs differ both in their locations (deep white matter, sub-cortical, and periventricular SCI), and causes (thrombotic and embolic). A thorough discussion of this topic goes beyond the scope of this thesis, but can be found in Caplan et. al. 24 Figure 1.2: A: T2-weighted MRI shows overt stroke in left anterior cerebral artery and middle cerebral artery territories; B: MR-angiogram shows occlusion of the internal carotid artery. Figure was adapted from Switzer, J. et. al 20 . Figure 1.3: A T2-weighted FLAIR MRI image of a patient with sickle cell disease, showing silent cerebral infarction circled in red. 14 Stroke prevention therapies and studies Among the interventions that have been used to prevent stroke in sickle cell disease, chronic transfusion therapy has proven to be the most effective 25 . This breakthrough in stroke prevention drew on the success of transcranial Doppler (TCD) ultrasound screening, which non- invasively examine the blood flow velocity in the main artery using ultrasound technique. The main strength of TCD is that it non-invasively provides fast estimation of cerebral blood velocity for major arteries. To understand the efficacy of blood transfusion in preventing primary stroke, the Stroke Prevention Trial in Sickle Cell Anemia (STOP) was initiated in 1998 26 . Screening was performed with TCD, and patients with reads higher than 200 cm per second in either the internal carotid artery or the middle cerebral artery were selected for the trial based on their high risk of stroke. Half of those patients were placed in the standard of care arm, and the other half were placed in the chronic transfusion therapy arm. After 11 strokes were identified in the standard of care arm versus 1 stroke in the transfusion one, the trial was terminated 1.5 years earlier than proposed 26 . Today, patients with SCD and TCD measured MCA velocities >200 cm/s are placed on chronic transfusion therapy to prevent the primary stroke prevention 27,28 . Additionally, given the observation that recurrent strokes were reduced by tenfold in patients placed on chronic transfusion 17,26 , it has become the standard of care in clinics for secondary stroke prevention 29 . Typically, transfusion therapy takes place as either simple transfusion, in which healthy HbA red blood cells are mixed with a patient’s own red blood cell via venous infusion; or exchange transfusion, in which a patient’s red blood cells are replaced with healthy donor blood via catheter. Simple transfusions account for 88% of all transfusions, and they are capable of suppressing the percentage of HbS to less than 30% 30 . While simple transfusions are known to be complicated by alloimmunization and iron overload, exchange transfusions are iron neutral. However, the widespread adoption of exchange transfusion is limited by several considerations, including: more consumption of blood units, better venous access required, and more demanding transfusion techniques, which requires either an extracorporeal device for automated exchange, or phlebotomy for manual exchange 31,32 . Chronic blood transfusions can improve oxygen saturation and increase hemoglobin- oxygen affinity, further reducing red blood cell sickling 33 . The other emerging treatment option is the use of hydroxyl carbamide, also known as hydroxyurea. Hydroxyurea was initially prescribed as a chemotherapy agent, with the observed side effect that it increases marrow fetal hemoglobin 15 (HbF) levels in adults. In addition, hydroxyurea increases the mean corpuscular volume, together with increased HbF level, the polymerization process is delayed, which alleviate RBC hemolysis 34 . Silent stroke in children and its cognitive outcomes To investigate whether blood transfusion therapy could prevent or lower the recurrence of SCI, another multi-center study, the Silent Cerebral Infarct Transfusion (SIT) Trial, was designed, in which children with SCD related SCI but normal TCD were assigned to either receive standard of care treatment (observation group) or regular blood transfusion therapy (transfusion group) 35 . At 36-month follow-up, new or enlarged SCI were identified in 2% of the transfusion group and 4.8% of the transfusion group, a relative risk reduction of 58% 36 . Despite the dramatic reduction of the SCI risk from blood transfusion therapy, children with SCD who had an overt stroke still experienced SCIs 23 . In spite of transfusion and hydroxyurea treatments, the cumulative prevalence of SCI increases linearly with age, with a predicted prevalence of 50% by the age of 30 37,38 , as shown in Figure 1.4. In spite of their name, SCIs are not ‘silent’ from a neurocognitive perspective. Studies comparing children with SCIs and disease matched controls have shown that children with silent strokes perform statistically poorer on cognitive performance 39 . Patients with silent strokes commonly display deficits in executive function, full-scaled IQ (FSIQ), verbal IQ and performance IQ 40–43 . When compared to healthy controls, adolescent SCD patients showed a range of academic difficulties, such as poorer math performance, and impairments in visual-motor functions 44 . Figure 1.4: The prevalence of SCI in children with sickle cell anemia. The percentage are from four different multi-central studies and one longitudinal study 34-36 . 16 Additionally, there is a preponderance of evidence that individuals with SCD experience difficulties with working memory, executive functions and processing speed 45,46 . Section III: Brain changes in SCD In addition to SCIs, a number of neuroanatomical changes have been associated to SCD in recent years, including brain volume loss, abnormal white matter (WM) microstructural injury and cortical thinning 47–50 . Choi et. al examined measures of brain morphology of clinically asymptomatic, young adult patients with SCD, by implementing structural magnetic resonance imaging (MRI). WM volume in the right and left frontal, parietal and temporal lobes was found to be lower in the patient group compared to controls 49 . Figure 1.4 shows the left and right hemisphere WM volume changes from Choi et. al’s work 49 . The distribution of the WM volume loss was in line with the regional vulnerability to stroke in SCD cohorts in the literature 14,51,52 , and the degree of the volume loss was primarily predicted by hemoglobin level (the index of the severity of anemia). Previous volumetric studies also showed grey matter (GM) volume decrease and cortex thinning in patients with SCD 53–56 . However, grey matter starts to shrink from childhood, and decrease more dramatically in early adulthood, as a natural developmental process due to synaptic pruning 57 . Therefore, the lower GM volume found in children with SCD may due to delayed cortical development, and has drawn less attention compared to WM health in this cohort 58–60 . Besides macro-structural findings in patients with SCD, previous studies have explored microstructural alterations of WM in patients with SCD, using methods such as region of interest (ROI) analysis 61 , tract-based spatial statistics 62 and tract-specific analysis 48,63 , all using diffusion- weighted MRI. It was reported that the corpus callosum, the largest and most prominent WM Figure 1.5: White matter volume group comparison results, where adjusted p-values are shown on the mid-cortical surface (Positive p-values in red indicate the GM volume of SCD patients are higher than that of healthy controls. Figure adapted from Choi et al 46 . 17 structure, was affected in the region of genu and splenium, including reduced fiber counts, and axonal damage 48,61 . In addition, the cortico-spinal tract was also found to have lower anisotropy in the inferior regions 62 . These changes in major WM tracts may indicate myelin loss due to chronic hypoxia, or a change in neuronal connection architecture 50,55,64 . Section IV: Cerebral oxygen transport Oxygen Content and Delivery The human brain weights less than 5% of the whole-body weight, but it requires around 20% of the body’s oxygen supply, due to the heavy metabolic demands of the neurons. Oxygen molecules are carried throughout the body by red blood cells to fuel aerobic respiration of the tissues. Each hemoglobin binds four oxygen molecules via its four globin chains. The binding of the oxygen molecules forms oxygenated hemoglobin, and it is a reversible process. At low oxygen concentrations, oxygenated hemoglobin dissociates to deoxygenated hemoglobin and oxygen, and this process releases the oxygen molecules to surrounding tissues. Given the high hemoglobin-oxygen affinity, 98% of the oxygen is carried in the red blood cell, while 2% is transported in the plasma as dissolved gas 65 . Taken together, oxygen content is expressed as the addition of hemoglobin-combined oxygen and dissolved oxygen 65 : Oxygen Content = 1 .34 × Hemoglobin × SpO2 + 0.003 × pO2 (1), where SpO 2 is oxygen saturation and pO 2 is oxygen partial pressure, which is assumed to be 100 torr for room air at the sea level 66 . To understand the quantitative “hemoglobin affinity for oxygen”, the oxygen-hemoglobin dissociation curve 67 or oxygen dissociation curve (ODC) shows how readily hemoglobin takes up and releases oxygen into the surrounding tissues (Figure 1.5). The ODC can also be modified to ensure adequate tissue oxygenation, and several contributing factors can lead to a lateral shift in the ODC in order to aid in the retention or release of oxygen from hemoglobin. These include the Figure 1.6: Oxygen hemoglobin dissociation curve relating the partial pressure of arterial oxygen (PaO 2 ) to the oxygen saturation for hemoglobin (Hb) A, shown in solid line, and for Hb S, shown in dashed line. Figure adapted from Wagner, et. al 65 . 18 pH of the blood, carbon dioxide concentration, temperature, and concentration of carbon monoxide 67 . In patients with SCD, the ODC is right shifted (dashed line in Figure 1.5) due to the presence of HbS, leading to lower partial pressure of arterial oxygen (PaO 2 ). This condition triggers sickle hemoglobin to form polymers, as mentioned earlier, and cause hemoglobin desaturation. HbS polymerization slows down blood transit in the microvasculature, and further interrupt the retention or release of oxygen from the hemoglobin, which in turn creates a vicious circle. On the other hand, ODC can be left shifted in the presence of HbF 68 , which can be induced in adults and children by taking hydroxyurea, a hydroxyl carbamide agent to alleviate pain crisis and acute chest syndrome. Increased HbF has been proved to retard polymerization of HbS, increase red blood cell survival and hemoglobin levels, and thereby lessening the right shift of the ODC due to HbS, as mentioned in the first scenario 69 . Fick Principle In addition to oxygen content, cerebral blood flow (CBF) is the other determinant of oxygen delivery to the brain. To understand the oxygen bioavailability of the brain, a physiology model is necessary to explain the oxygen supply and demand relationship. Fick principle, developed over a century ago, states that in a closed system, oxygen consumption is the difference of cerebral oxygen inflow and outflow 70 . Based on Fick’s principle, the cerebral metabolic rate of oxygen (CMRO2) can be stated as: 𝐶𝑀𝑅𝑂 B =𝑂𝑥𝑦𝑔𝑒𝑛 𝐼𝑛𝑓𝑙𝑜𝑤−𝑂𝑥𝑦𝑔𝑒𝑛 𝑂𝑢𝑡𝑓𝑙𝑜𝑤 (2) Oxygen inflow and outflow can be expressed as the product of CBF and the concentration of oxygen content in the arterial and venous systems: 𝐶𝑀𝑅𝑂 B =𝐶𝑎𝑂 B × 𝐶𝐵𝐹 × 𝑂𝐸𝐹 (3), Figure 1.7: Fick principle states that the total uptake of oxygen is equal to the product of the blood flow to the arterial- venous concentration difference of oxygen. 19 where CaO 2 is the arterial oxygen content and OEF is oxygen extraction fraction. This states that if oxygen content remains constant and OEF = 1, then increases in metabolic demands can be satisfied by a proportionate increase in CBF. If we take equation (1) and drop the second term (contributing less than 2% of the total oxygen content), OEF can be derived as: 𝑂𝐸𝐹 = abc d e af c d abc d (4), where SaO B is arterial oxygen saturation and SvO B is venous oxygen saturation. In normal brain, OEF is about 30%. In early 1990’s, Fox and Raichle demonstrated that CBF increases disproportionately larger than the increase in oxygen utilization measured by positron emission tomography (PET). However, more recently, several studies showed that the consumption rate of oxygen increases can be supported by increases in OEF, independent of CBF changes. Krogh Cylinder The arterial blood (oxygenated) provides tissues with oxygen, and is divided into two portions along the oxygen-supplying capillaries: part of it diffuses into the tissue to be consumed and the other part outflows to the venules and unites to a vein. To more closely study how oxygen is diffused in the tissue, August Krogh (1919) modeled the situation by assuming that the oxygen diffuses from a cylindrical capillary into a concentric cylindrical tissue 71 (Figure 1.8). This Krogh cylinder model is powerful and has been used to make predictions and help explain observed physiological parameters in healthy and diseased microvasculature 72,73 . However, the model was established based on the assumptions that: 1) oxygen exchange occurs only in the capillary, not in arterioles and venules; 2) oxygen does not diffuse out of the tissue cylinder; 3) capillary radius, length and transit time is homogenous and does not depend on time or position 74 . Krogh also argued that capillaries themselves are part of the regulation of oxygen supply by means of capillary recruitment: to open previously closed capillaries so that the tissue’s capillary surface area available for diffusion can be increased. However, in the brain, capillary recruitment does not hold true after an observation of RBC transiting through the capillary bed in the mouse by Kushinsky and Paulson 75 . Figure 1.8: A Krogh Cylinder Model can be used to model oxygen diffusion across the tissue from the vascular cylinder (red arrow). Figure credit goes to David Wootton. 20 Capillary transit time heterogeneity On the contrary, neurovascular coupling mechanisms 76 , or changing arteriolar tone as a response of metabolic needs, became the cornerstone of many related work to understand the link between cerebral hemodynamics 77–79 . Unlike Krogh’s hypothesis mentioned above, Villringer and colleagues observed extreme heterogeneity of RBC transit times across cerebral capillaries 80 instead of temporarily closing and opening of capillaries, named capillary transit time heterogeneity (CTTH). More recently, Jespersen and Østergaard extended the existing compartment models: using in-vivo rat data, Jespersen took into account CTTH and tissue oxygen tension, hypothesizing that capillary flow patterns may affect the efficacy of oxygen extraction, owing to biophysical limitations of oxygen diffusion from blood to tissue 77 . As illustrated in Figure1.9 by Jespersen and Østergaard 77 , venous outflow oxygen concentration is affected by the heterogeneity of capillary flows, in spite of identical total blood flows and number of capillaries. In Figure 1.9A, better oxygen extraction (shown in capillaries with warm color) in capillaries with longer transit times dominates over poor oxygen extraction (shown in capillaries with cold color) in capillaries with short transit times, resulting an increased OEF (approximately 0.7) compared with the homogenous case ((approximately 0.7). This example shows that OEF can increase in the absence of total CBF changes, only by changing CTTH. On the other hand, Buxton and Frank further investigated the physiological relationship between CBF and CMRO 2 by tracking signal changes in functional MRI 81 . According to their ‘oxygen limitation model’ 82 , oxygen availability depends on blood flow and a rate characterizing oxygen diffusion from blood to tissue. As OEF decreases with CBF according to the Bohr-Kety-Crone-Renkin (BKCR) model 83 , large CBF elevation is needed to support even modest increase in oxygen consumption, and in order to increase the local CBF, either the transit time have to decrease, or the capillary volume have to increase, or both 82 , where the latter is still an unresolved and debatable topic. Due to this diffusion limitation, CBF increases non-linearly and disproportionately larger than the increase of oxygen consumption. Figure 1.9: Effects of CTTH on oxygen extraction (from 1.0-0.4 shown indicated in color bar) with homogenous capillary flow velocities (B) and heterogeneous flow velocities (A). Figure was adapted from Jespersen and Østergaard 77 . 21 While gas diffusion allows the entering of oxygen into the blood, adequate perfusion helps to move arterial blood to the capillary bed in brain tissue. The rapid development of quantitative imaging techniques has allowed us to reveal spatial variance in brain perfusion. In this work, brain perfusion, CBF will be introduced in the next section. Section V: Cerebral blood flow in health and disease Brain vascular structure Normal function of the brain depends upon adequate supply of oxygen and nutrients. This is made possible by the brain’s highly organized dense network of blood vessels. Blood is supplied to the brain, face, and scalp via two major sets of arteries: the right and left common carotid arteries and the right and left vertebral arteries. At jaw level, the common carotid arteries bifurcate into the internal carotid arteries, supplying blood to most of the anterior portion of the cerebrum, and the external carotid arteries, supplying blood to the face and scalp. Whereas vertebra-basilar arteries supply the cerebellum, brain stem and the posterior of the cerebrum 84 . At the base of the brain, the basilar and internal carotids form a circle of communicating arteries, known as the Circle of Willis. From this circle, the largest cerebral artery pair, the middle cerebral arteries (MCA), supplies blood to the left and right lateral surface of temporal and parietal lobes. The anterior cerebral arteries (ACA) supply blood to the frontal and parietal lobes, and posterior cerebral artery (PCA) supply blood to the occipital and part of the temporal lobes. Figure 1.10 illustrates the schematic territory of three main blood flow territories. Even though the drawings of the territorial atlases began in 1900’s, most of the drawings were based on postmortem studies with the assumption of symmetrical and identical territorial distribution 85,86 . In a review article by Van Der Zwan and Hillen, the most representative studies were compared and the variability of the cerebral vascular territories investigated 87 . Van Der Zwan Figure 1.10: Vascular territories of the cerebral cortex in A: lateral sagittal view and B: midline sagittal view. Inferior view. Figure adapted from: Radiopaedia.org. Case courtesy of Prof Frank Gaillard. A B ACA MCA PCA 22 and colleagues showed that the discrepancies of the cerebral vascular territories were significantly greater than was assumed 88 . However, their published maps of vascular variability based on autopsy series has limitations due to a few number of cases. Jan van Laar and colleagues then extended the experiments to in-vivo studies in subjects with non-variant type and variant types of the circle of Willis 89 . Their results showed that the variations in flow territories of the brain were relatively small within subgroups 89 . In this work, the templates of flow territories in standard space was used from Tatu and colleagues’ product 90 . Other than tissues that are directly supplied by major arteries, some areas of the brain simultaneously receive blood supply from the most distal branches of two arteries, so-called watershed area or border zones. Watershed locations are known with the least supply of flow, therefore, are most vulnerable to hemodynamic changes, or any reduction in flow 91 . Due to the terminal end of the vascular supply, watershed or border zone infarcts have been reported previously in normal aging and patients with neurological diseases, using either MRI or CT 92,93 . Similarly, in patients with SCD, more than 80% brain infarctions were found in arterial watershed areas 94–96 . The extension and the terminus of the large arteries are the microvasculature bed, including arterioles, capillaries and venules, which are all working together to perfuse the deep brain. Arterioles are small arteries leading to the capillaries. With a lumen of about 30 micrometers or less on average, arterioles are very important in resisting blood flow, and therefore, they regulate blood pressure. They are sometimes also referred to as resistance vessels. A capillary is a microscopic channel where perfusion happens. In this process, there is an exchanging of gasses and other substances between the blood and the surrounding tissues. Exiting from a capillary bed, capillaries converge to a venule and eventually, many venules join to a vein. Cerebral blood flow and regulation Cerebral blood flow is the volume of arterial blood delivered to capillary beds and brain tissue per minute. For instance, the CBF of the healthy human brain is typically from 45-55 mL per 100 grams of brain tissue per minute, often expressed mL/100g/min 97 . To ensure the proper delivery of basic metabolic substrates (oxygen and glucose) to brain tissues, CBF is highly regulated. This regulation is enacted by contracting or relaxing the smooth muscles of the vessel wall; these processes are also known as vasoconstriction and vasodilation, respectively. 23 CBF is known to vary across the lifespan. In healthy children, whose brain metabolic rates are high, CBF is greater than in adults. More concretely, CBF increases dramatically from 7 months to 6 years of age, declines at a moderate rate until the age of 18, and more gradually in adults until the sixth decade of life. Figure 1.11 plots the age-related changes in CBF 98–100 . By the same token, whole brain oxygen delivery is expected to follow the same fluctuation pattern with respect to age, assuming that oxygen content is not age-dependent. But no related work has reported longitudinally how oxygen delivery varies with age. CBF is modulated by mechanical, chemical and neuronal stimuli. For example, to maintain oxygen delivery to the brain, CBF increases when supply of oxygen is reduced, or in a hypoxia condition, which can be induced by carbon monoxide or decreased SaO 2 . Increases in CBF can also be induced by vasodilatory challenge, including carbon dioxide (CO 2 ) inhalation, or acetazolamide injection, which mimics CO 2 inhalation. In patients with SCD, to ensure the sufficient oxygen delivery to maintain the cerebral metabolism, CBF is elevated to compensate for the decreased oxygen content, which is proportional to hemoglobin level. In the whole brain, it was demonstrated that oxygen delivery was identical in patients with SCD as healthy controls 101 . However, regional oxygen delivery inhomogeneity requires further investigation. CBF changes in response to anemia and SCI CBF is also well regulated regionally. On top of the global CBF increases reported in patients with SCD in both pediatric and adults, studies have also demonstrated variations in CBF levels in different arterial flow territories and in distinct tissues such as WM and GM 102–104 . CBF (mL/min) Figure 1.11: Changes in cerebral blood flow (CBF) with respect to age. CBF values are fitted using cubic regression. Note that the horizontal axis is scaled by one-fifth for adults to better illustrate the changes in childhood. Figure is adapted from Wu, et, al 78 . 24 Although the CBF did not differ significantly between flow territories from previous work 105–107 , CBF in GM is significantly higher than that in WM, because GM is more metabolically active than WM 108,109 . CBF differences in the left and right hemispheres have also been studied in anemic patients; however, the asymmetries in CBF could either reflect perfusion differences due to pathologies in SCD, or be the result of technical difficulties encountered in imaging modalities 110 . This will be discussed more thoroughly in the Part II of the chapter. CBF regional abnormalities in the brain were associated with SCI. In normal aging group, measurements showed significant lower CBF in regions with SCIs than the ones with normal WM 111,112 . More recently, it was reported that CBF is lowest in the regions of highest SCI density, using the SIT Trial dataset 113 . The cerebral hemodynamic and metabolic assessments introduced above are highly dependent on the effectiveness and accuracy of measurements. In the next section, MRI will be introduced as a means to improve these measurements and strengthen related findings. This modality will be used here to identify more specific biomarkers of stroke. Part II: Magnetic Resonance Imaging in SCD Section I: MRI basis MRI is one of the most commonly used imaging modalities in clinics. It offers unique detail of brain, spinal cord and vascular anatomy, with high reproducibility. MRI is based on the magnetization properties of atomic nuclei, or “spins”. Figure 1.12 shows the diagram of how a nuclei forms imaging signal, and eventually, an MR image. When a uniform external magnetic field B 0 is applied, the nuclear spins of the tissue being examined will align with B 0 and form a non-zero equilibrium magnetization M 0 . This process is known as polarization (workflow A in Figure 1.12). At room temperature, M 0 is defined by: (5), where ρ is the proton density of water, the Larmor frequency γ is 42. 6 MHz/T, k is the Boltzmann constant, T is the temperature in Kelvin, and ħ is Planck’s constant 114 . Figure 1.12: Diagram of MR signal generation. A: polarization, generating M; B: excitation, generating M xy ; C: signal detection, yielding S(t); D: spatial encoding, generating S(k xy ); E: image reconstruction, producing p(r). Figure was adapted from Justin Haldar lecture notes (fall 2016). 25 To generate the actual signal, the equilibrium state is then perturbed by the introduction of an external radio frequency (RF) pulse. This process is known as excitation and is represented by workflow B in Figure 1.12. After excitation, spins return to their resting alignment. The recovery (spin-lattice relaxation) of the longitudinal magnetization, M z and of the transverse magnetization M xy (spin-spin relaxation) are defined by the following equations 115 : (6-7), where T 1 is the time constant for regrowth of M z , t is the time after excitation, and T 2* can be further separated into two components, which are the T 2 relaxation due to interactions between spins that cause random loss of phase coherence, and the T 2’ due to time-dependent field inhomogeneities. T 1 , T 2 and T 2* are measurable, characteristic time parameters depending on the internal properties of the measured tissue. In the human brain, different T 1 , T 2 and T 2* are the key to differentiating tissues and physiologic states. Workflow C in Figure 1.12 represents the signal detection following Faraday’s Law. In the classical representation, oscillating of the M xy components creates magnetic flux in the transverse plane, and an external coil loop sensitive to the flux will detect the signal, which is proportional to the rate of the change to the magnetic field (dB/dt). To sort out where each signal comes from, the detected signal, S(t), will go through a process called spatial encoding (workflow D in Figure 1.12), which generates S(k xy ). Generally, there are three strategies to “encode” the time signal to spatial signal: coil sensitivity encoding (applying small surface coils that only see a small portion of the image), RF encoding (exciting spins at certain spatial positions), and Fourier encoding (manipulating the z-component of applied magnetic field as a function of spatial position) 114 . The last step is image reconstruction. Conventionally, Fourier transformation is used to convert the frequency and spatial information S(k xy ) in the imaging plane to the corresponding intensity values, which are then visualized as images. Image reconstruction can also be performed using filtered back projection (for radial imaging) or gridding (interpolating values onto a rectilinear grid of Fourier transformation) 115 . However, the full introduction of image reconstruction techniques falls beyond the scope of this thesis. 26 Section II: Arterial Spin Labeling Arterial Spin Labeling (ASL) is a MRI technique developed about 20 years ago to measure tissue perfusion 116 . Conventionally, measuring perfusion requires a diffusible tracer that can exchange between the vascular compartment and tissue. This “tracer” can be exogenously administrated by intravenous injection prior to imaging, and then be detected by a variety of imaging modalities, such as positron emission tomography (PET) 117 , computed tomography (CT) perfusion 118 and dynamic susceptibility contrast (DSC) MRI 119 . However, the ironizing radiation and risk of contrast reaction limit the use of perfusion studies in healthy individuals. ASL, on the other hand, is radiation free and contrast free, since it utilizes magnetically labeled arterial blood water as endogenous tracer in order to create quantitative cerebral blood flow maps 120 . Labeling To be more concrete, ASL uses RF pulses to invert inflowing spins in arterial blood water; this process is called “labeling” or “tagging.”. There are different labeling strategies for different clinical applications, and here we focus only on the one we will use for our studies, pseudo continuous arterial spin labeling (PCASL) 121 . In PCASL, a series of small flip angle, slice selective RF pulses are applied to adiabatically invert the inflow spins. This labeling pulse train typically consists 1,000~2,000 Hanning-shaped RF pulses, separated in time by 1~2 ms 122 . The pulses will Figure 1.13: PCASL labeling pulse train and gradients. In the labeling sequence (red), the magnetization is inverted during a series of slice selective flips; in the control sequence (blue), the RF profile is reversed alternatively to preserve the equilibrium magnetization. 27 slowly rotate the M vector of stationary spins from the positive z to the negative z orientation. The spatial labeling plane is placed by applying a slice selective gradient and refocusing pulse (usually in z-direction for brain ASL) 123 . Due to its intrinsic labeling process, PCASL may not achieve perfect labeling efficiency, which may vary depending on RF strength, imperfections of B 0 and B 1 field, blood flow velocity, blood transverse relaxation and other physiologic parameters 124 . In mathematically simulations, the labeling efficiency approaches 85%, and image experiments in healthy adult subjects have meet this optimal labeling efficiency 125 ; however, labeling efficiency in subjects with high blood velocities, such as in anemic patients or children, requires a more specific estimation tailored to these individuals 124,126 . In the control pulse train, the RF pulse is reversed alternatively so that equilibrium magnetization is maintained, as illustrated in Figure 1.13 in blue. After labeling, there is a time delay between the end of the pulse train and image acquisition, and this delay is named post- labeling delay (PLD). Taken altogether, a PCASL sequence consists of pairs of labeled and control images, and the difference of each label-control pair is a perfusion-weighted map. Kinetic Model In order to quantitatively estimate CBF from a perfusion map, a kinetic tracer model is used. The simplest model is the single-compartment model proposed by Richard Buxton in the early 2000’s. In this model, arterial blood water is assumed to travel intravascularly from the labeling plane to the imaging voxel. The labeled spins diffuse from the vessel and homogenously fill the tissue. This model was advocated by the ASL White Paper, a consensus document produced by the ISMRM Perfusion Study Group, where CBF is expressed as 122 : (8). T 1b is longitudinal relaxation of blood (recommend as 1650ms), λ is the blood brain partition coefficient (0.9), α is the labeling efficiency (0.85), w is the PLD (2000ms), PD is a supplemental proton density image and t is the labeling duration (1600ms) 122 . However, this model is only optimal in CBF quantification when: a) the longitudinal relaxation of arterial blood is identical in vasculature and tissues; b) the transit time difference for distinct brain tissues is negligible 122,127,128 ; and c) grey matter is the tissue of interest. These 28 assumptions hold true in healthy adult individuals, but in pediatric and anemic cohorts, whose CBF are higher, both assumptions a) and b) are violated. This is especially true for transit time, a major factor in quantifying the proportion of the ASL signal in vascular and tissue/capillary compartment 77,129 . Therefore, a two-compartment model arose in response to these issues. A most commonly used two-compartment model is derived from Wang et. al.’s work 121 . In general, by carefully selecting the time delay, w, between the labeling pulse and the acquisition, sufficient time should allow the entire ASL signal to travel to the imaging voxel. However, depending on the location of the labeled arterial blood water in the vasculature at the time of acquisition, the perfusion signal can be modeled differently. This model depends on the time duration relationship between the transit time, post labeling delay and labeling duration. Therefore, the perfusion signal ∆M, acquired in two different time-dependent conditions, can be written as 𝛥𝑀 =2∗𝛼∗𝐶𝐵𝐹 exp qr s tuv 𝑑𝑡 xes y + z{| e } ~ z{| } ~ 𝑑𝑡 xes ,𝑓𝑜𝑟 (τ+𝑤)<δ (9) an d 𝛥𝑀 =2∗𝛼∗𝐶𝐵𝐹 z{| (ex/tuv) z{| ((qr s ex)/tuq) 𝑑𝑡,𝑓𝑜𝑟 𝑤 <δ y (10), where α is labeling efficiency, w is the post labeling delay, t is the labeling duration and δ is the arterial transit time 121 . These two equations can be solved to quantify CBF with a two-compartment model 124 : 𝐶𝐵𝐹 = yyy∗∗ qebv ∗ } ~ B∗∗t ∗∗( } ~ e } ~ ) (11). In this thesis, CBF quantification will be performed using equation (11), to incorporate a population based T 1t and a transit time in patients with SCD and chronic anemic syndrome 77 . Oxygen Delivery in the brain When CBF is estimated from ASL, cerebral oxygen delivery is simply the product of CBF and oxygen content 65 , where the latter is stated in equation (1). Oxygen Delivery = CBF × Oxygen Content (12). From equation (1) and (12), it is obvious that cerebral oxygen delivery is linearly dependent on hemoglobin and CBF. Previously, hematocrit-dependent cerebral O 2 delivery has been described in subjects with cardiovascular disease, with optimal delivery for a hematocrit of 40-45% 130 . 29 Section III: Diffusion-weighted Imaging As mentioned in previous sections, SCD is a white matter disease. In recent years, WM microstructural injury has gathered more attention from the research community. Conventional MRI is not capable of precisely delineating microscopic changes and the WM tracts of the brain. Diffusion weighted imaging (DWI), a newer imaging technique that utilize sequences in multiple diffusion-sensitizing gradient directions, allows us to measure the Brownian motion of water molecules in brain tissues 131 . DWI has become an important technique in analyzing WM, because of the latter’s intrinsic composition and anisotropic property. White matter is mainly made up of myelinated neuronal axons, bundled together to form WM tracts, also known as WM fibers. Trillions of WM tracts connect various grey matter areas of the brain. These long and myelinated axons restrict water molecules from diffusing in random directions. Instead, the parallel orientation on average of nerve fiber tracts makes WM highly anisotropic. The degree of anisotropy can be characterized by a diffusion ellipsoid with different hemi-axis lengths, which are measured by diffusion-weighted images. Diffusion Imaging Sequences Diffusion gradients can be readily incorporated in a conventional spin echo MR sequence 132 . DWI sequences used in the past decades are rooted in the pulsed gradient spin echo (PGSE) technique invented by Edward Stejskal and John TannerE, as illustrated in Figure 1.14. Firstly, images are acquired without the diffusion gradient, resulting in a set of “b0” images that will serve as the baseline for further calculations. These b0 images are typically T2-weighted and give the signal intensity S 0 . Then, diffusion-weighted images are obtained with the diffusion gradients turned on, in a combination with different gradient strengths and directions. The measured signal S k is sensitive to the small water displacements, as shown in red in Figure 1.14. The number of diffusion gradients have to be at least six to allow solving for symmetric 3x3 diffusion tensors. The reduction in the measured signal resulting from the application of a pulse gradient, is an exponential decay proportional to the amount of diffusion, according to the Stejskal-Tanner equation: 𝑆 =𝑆 y 𝑒 ev ¡ (13), 30 where S k is the image signal measured at k th gradient pulse, b is the b-value (typically ranges from 500-3000 s/mm 2 ), and D gk is the tensor given at k th gradient pulse that we will need to solve. g k is the gradients and can be expressed as: 𝑔 =[𝐺 ¤ 𝐺 ¥ 𝐺 ¦ ] t . (14), The diffusion tensor D, modeled as a 3x3 diffusion ellipsoid with different axis lengths and directions corresponding to the tensor eigenvalues λ 1 , λ 2 and λ 3 (λ 1 > λ 12 > λ 3 ) and eigenvectors, respectively (illustrated in Figure 1.15). D can be expressed as: 𝐷 = 𝐷 ¤¤ 𝐷 ¤¥ 𝐷 ¤© 𝐷 ¥¤ 𝐷 ¥¥ 𝐷 ¥© 𝐷 ©¤ 𝐷 ©¥ 𝐷 ©© (15), The metrics used in DTI are typically functions of the 3 eigenvalues from the diffusion tensor matrix. Diffusivities are simply linear combinations of the eigenvalues, whereas anisotropic measures are more complex transformations that capture the disparities of the diffusivities. Diffusion Tensor Modeling As mentioned above, eigenvalues and functions of them form the basis for most standard diffusion measures. The four most popular metrics used for diffusion tensor imaging (DTI) are: axial diffusivity (AD), radial diffusivity (RD), mean diffusivity (MD) and fractional anisotropy (FA). AD characterizes the value of λ 1 , the longest eigenvalue, RD is the average of λ 2 and λ 3 and represents the two shortest eigenvalues. MD provides an average of all three eigenvalues. The FA of each voxel of the brain is calculated as: 𝐹𝐴 = u B e d d r e « d r d e « d d r d d r « d (16), FA quantifies the normalized degree of anisotropic diffusion using a range of 0 to 1, where 0 is completely isotropic and 1 is completely anisotropic. Figure 1.14: Tensor shapes and their corresponding types of diffusion. Figure courtesy from: www.diffusion-imaging.com 31 All four metrics of DTI show differences in biological microstructure properties in the brain. For example, AD reflects WM changes and pathology; it decreases in axonal injury neurological disease, but it increases with brain maturation. RD shows dark (small values in a RD map) in highly organized and dense structures such as WM tracts, and shows bright (bigger voxel intensities in a RD map) in the regions filled with CSF. Changes in the axonal diameters and density or demyelation/dysmyelination can increase RD, whereas MD is more sensitive to cellularity, edema and necrosis. FA is a global measure of microstructural integrity, but it is less specific to the type of change. Motivation for microscopic white matter injury measurement in SCD As introduced in Part I, SCD is known to be associated with a high incidence of cerebrovascular complications, including stroke, SCI and brain atrophy. However, the exact mechanisms of cerebral damage, especially in WM are not fully understood in SCD. Prior studies showed that children with SCD who had normal clinical MRI findings still had declined neurocognitive performance, suggesting that there is a diffuse brain injury, or subtle changes in WM integrity in these patients. Even in patients whose structural images were identified with SCIs, the presence of SCIs was strongly associated with DTI metrics abnormalities 133 . Therefore, employing DTI in patients with SCD could be a great help in generating markers of WM health; this method can be used for tracking disease progression, and for studying neurological development prior to stroke. Current studies using DTI in SCD Numerous DWI-based measures have been proposed, but FA and MD have become the most widely used DTI-based metrics in brain MRI research. The earliest studies using DTI reported AD abnormalities visually, and was used as an additional imaging sequence to provide information about the timing of stroke 134–136 . With the quantification of FA, MD and RD maps, more studies emerged to quantitatively compare these metrics in the whole brain with healthy controls. Initial DTI based studies of SCD were focused on average behavior over specific regions of interests. Balci et. al. performed quantitative FA and AD analysis by outlining regions of interest (ROIs) 61 and found reduced FA and increased AD, as well as significant reduction in fiber counts in the most anterior areas of corpus callosum in patients with SCD. In addition, in the subgroup analysis, patients with SCIs had significantly lower FA values in the right centrum semiovale and 32 right superior frontal WM, and significantly higher AD values in the left occipital and right temporal WM. These findings were in line with the focal or global brain atrophy as predicted by diffuse brain injury in SCD patients without silent infarcts 14 . The decreased FA was believed to be the result of the axonal damage in the brain tissue undergoing chronic ischemia, while the increased AD could be the consequence of the increased extra-celluar water contents secondary to micro- and macroscopic cystic changes in the WM 137 . However, ROI-based methods are less sensitive to the specification of the certain bundles. The development of statistical group comparisons was a turning point in the field, and provided new means of performing detailed local analyses of the diffusion signal. These methods include for instance using permutation tests based on the alignment-invariant tract representation (mean FA skeleton), or continuous medial representation (cm-rep), which are the bases for the popular Tract-Based Spatial Statistics TBSS and Tract-Specific Analysis TSA, respectively 63,138 . Sun et. al. used TBSS to systematically explored microscopic regional WM changes in SCD. The results showed significantly lower FA in centrum semiovale in patients with SCD in whom conventional anatomic MRI scans showed no abnormalities. Using the same method, Kawadler and colleagues investigated differences of WM integrity in children with SCD without SCIs 62 . RD was found to be significantly higher in patients in their parietal and frontal lobes, and bilaterally in the corticospinal tract. Elevated RD was believed there to be due to axonal loss and acute demyelination 62 , led by ischemic infarction. The diffusion tensor model has important shortcomings, including most importantly failure to resolve multiple tract directions in regions with crossing fibers, thereby potentially underestimating FA values in regions where kissing, curving and branching fiber are prevalent. More complex models exist such as fiber orientation distribution (FOD) estimation, or constrained spherical deconvolution (CSD) estimation, both of which require high gradient multi-directional high angular resolution diffusion imaging (HARDI) data to better handles crossing fibers and dispersion. To resolve tissue mixture problems, multi-shell-multi-tissue (MSMT) models were also proposed 139 . One drawback of these important tools are the large scanning times are required for full acquisitions, which may not always be feasible in the context of clinical imaging research. Hence, given these constraints, for the work presented here we will focus exclusively on diffusion tensor-based methods. 33 Once we obtain 3D CBF and cerebral oxygen delivery map, as introduced in Part I, and quantitative DTI-derived metrics map, it is imperative to co-localize them with an SCI probability map, or a “heat map”, in order to ultimately identify precursor or imaging biomarkers of SCI (and further downstream, biomarkers of stroke). This will be made possible here through segmenting more than 700 SCIs in more than 200 T2-fluid attenuated inversion recovery (FLAIR) MR images (SIT trial data). In the next a few sections, learning-based segmentation techniques and related super-resolution topics will be introduced. Part III: Techniques: Statistics and deep learning SCIs exhibited higher image intensity values on brain T2-FLAIR images, therefore they are referred to as WM hyperintensities (WMHs). WMH on MRI are of different types, from large periventricular WMH to deep WM ischemia, thus WMH patterns are heterogeneous across different pathophysiologies 140,141 . Visual rating scales are easy to perform and commonly used in clinical and research settings, but they are limited due to high intra-subject and inter-subject variability 142 . In addition, manual delineation and quantification of WMH is cumbersome and time-consuming for the neuroradiologists. Previously, popular histogram or simple threshold based methods reported in the literature performed well on FLAIR images, but they required a final cut-point decision based on human experts 143–145 . Additionally, the threshold that worked well on one sequence may perform poorly when it comes to different sequence acquired from different scanners or medical sites. More recently, machine learning- and deep learning-based automated segmentation algorithms have been developed using different types of MRI sequences 146–148 . Section I: Automatic SCI segmentation Supervised and unsupervised-learning Machine learning-based segmentation algorithms could be broadly categorized as supervised and unsupervised methods. The former requires prior knowledge of what the output class (lesion or non-lesion) should be designated as “ground truth”, or “labels”. In brain lesion segmentation, this means that manual delineation and quantification of WMH is performed before the segmentation tasks. Common algorithms in supervised learning include logistic regression, Bayesian approach, support vector machine (SVM) 149 , k-nearest neighbors (kNN), random forests 150 , and neural networks. Traditional supervised learning methods performed well, but they relied on a set of handcrafted features for the classification, and could be limited by their 34 computationally complexity. In this context, unsupervised methods lifted the burden of manual segmentation, and were more easily portable to different scans from different manufactures or medical sites 151 . Deep neural networks, on the other hand, could be either supervised or unsupervised. Given that deep neural networks are a large topic and will be utilized in this work, they will be introduced in the last sub-section. Compared to supervised methods, unsupervised learning methods are leaner and easier to implement. Without using explicitly-provided labels, the method will learn the inherent structure of the data, and eventually create groups of data points (voxels of the brain image) such that points in different clusters are dissimilar (normal brain tissue voxels) while pints within a cluster are similar (WMH voxels). The most commonly used methods are: k-mean clustering, hierarchical clustering and principal component analysis (PCA). From the previous work, the unsupervised method has poor precision/accuracy in voxel-wise segmentation tasks. In this context, voxel-wise accuracy is required for SCI burden estimation and multi-modal analysis; therefore, unsupervised learning-based methods were not employed in this thesis, due to the difficulties in voxel-wise estimation. Segmentation Evaluation In order to evaluate the accuracy of the segmenting results, WMH volume and WMH voxels have been tested. Due to the lack of the information on the overlap between automatic segmented results and the ground truth, the similarity index, or dice similarity coefficient (DSC) has been more frequently used. It is scaled from 0 to 1. According to the literature, a rating above to 0.7 is considered to be good segmentation. The other widely used evaluation method is intra- class correlation coefficient (ICC), which is also adapted to set the criteria of the ground truth segmentation, such as to evaluate both intra-rater reliability and inter-rater reliability. Other overlap measurements include the Jaccard index, sensitivity, specificity and accuracy, DSC, ICC and other correlation measurements. Table 1 shows the formula to calculate the accuracy estimation metrics. If the segmented images have structural information, and perceptual quality of the resultant images are important, another evaluation metric, structural similarity index (SSIM) will be used. We will use SSIM along with peak signal-to-noise (PSNR) ratio to estimate the whole image in super-resolution problem, which will be introduced in the next Section. 35 Metric Formula Dice similarity index 2TP/(FP+FN_2TP) Sensitivity TP/(TP+FN) Specificity TN/(FP+TN) Accuracy (TP+TN)/(TP+FP+TN+FN) Jaccard index TP/(TP+FP+FN) Table 1. common measurements used to evaluate WM lesion segmentation. Abbreviations: TP=true positives, TN=true negatives, FP=false positives, FN=false negatives. Previous work using supervised-learning in SCI segmentation Lao et. al. used SVM to build a classification model, by defining the feature vector using intensity and spatial information of a small neighborhood of each voxel 96 . In this way, the method is more robust to the cases where image co-registration is not perfect. Lao also used a false positive correction step after classification model based on Hilbert distance 152 . Unlike Lao, combining the targeted voxel’s intensity and spatial information together, Anbeek and colleagues used kNN in T1-weighted, inversion recovery, proton density and T2-FLAIR scans separately, to build a WMH probability map where the value of each voxel was defined as the fraction of WMHs out of its 100 neighboring voxels 153 . However, the intensity and 3D spatial features were employed, and the performance was heavily dependent on the cut-off threshold. Steenwijk et. al improved kNN classification of WMHs by optimizing intensity normalization with the use of spatial tissue type priors (TPPs) 147 . We took our own research T2-FLAIR images in patients with SCD and reproduced this work. The preprocessing pipeline is demonstrated in Figure 1.15. The WM probability mask was extracted using FAST 154 , a FMRIB-based tissue-segmentation method. 36 After preprocessing, the WM-masked and pre-processed T2 FLAIR images were the input of the training process. The images were read in as feature vectors in a hyper-plane. Due to limited computational resources and abundant data points in feature space, we utilized Fisher’s discriminant analysis (FDA), a powerful method for supervised dimensionality reduction. After FDA, feature space dimension was decreased from 30 (size of 3 3 for intensities of the targeted voxel and spatial location x, y and z) to 5, such that the input to the kNN model will be a dimensional space of manageable size. The dataset was divided into two halves for training and testing to yield a learning model that could be tested on the separate patients. The scheme of the algorithm design is presented in Figure 1.16. Figure 1.15: Workflow of imaging preprocessing for a tissue-type prior supervised segmentation method. 37 Deep neural networks (DNN) Despite of the careful feature design, abundant preprocessing steps, and complex model, the accuracy of these methods relies on the training dataset 141,146 , which may need to be prohibitively large, or may not be universally applicable due to the heterogeneous nature of WMH. Another limitation of traditional learning-based method is that it could be very computationally expensive when large dataset is available for the experiment. Given our large dataset (205 children with more than 750 lesions) and relatively homogeneous WMH type (SCI) with small lesion load 21 , we designed an improved learning algorithms based on convolutional neural networks (CNNs) to Figure 1.17: A: One neuron unit and its input connections, ⍬ i is the i th neuro unit of the layer; B: Neuroal network with multiple hidden layers. Figure 1.16: The scheme of the algorithm design using k-nearest neighbor. FDA=Fisher’s discriminant analysis. 38 better manage unbalanced data, which is more capable of handling high dimensional space, and by implementing three CNNs in parallel with different patch size, we further can decrease false positive rates 155,156 DNNs, or Deep Nets, are widely used in image recognition and have been applied successfully on a variety of biomedical segmentation problems, including WM lesion segmentation 157–159 . Unlike classical machine learning methods, DNNs do not require a set of handcrafted features for the classification tasks. Instead, the neural network takes input data points directly and enters them into the “neuron”, as illustrated in Figure 1.17A. Generally, DNN is a collection of “neurons” with “synapses” connecting them with inputs, weights, neighboring layers, and outputs. Figure 1.17A simply shows one neural unit with its input and output connections. However, in practice, many hidden layers are designed with the term “deep learning,” implying multiple hidden layers as shown in Figure 1.17B. To get the final output, the activation function is applied to the hidden layer sums to yield an output based on the threshold ⍬ i . Figure 1.18 shows the most commonly used six activation functions in neural network design raining the network means repeatedly adjusting the weights by two steps, forward propagation and back propagation. In forward propagation, a set of weights are applied to the input data, and this process yields an Figure 1.18: Examples of activation functions. A: Linear, f(a i )=f(w T x-⍬i); B: Hard threshold, f(a i ) = -1 if x<0, f(a i ) = 1 otherwise; C: f(a i )=tanh(a i ); D: Rectified linear unit (ReLU), f(a i ) = 0 if x<0, f(a i ) = x otherwise; E: Leaky ReLU, f(a i ) = 0 if x<0, f(a i ) = 0.01x otherwise; F: Sigmoid, f(a i )=S(a i ). 39 output. In back propagation, the margin of error of the outputs are estimated, and they in turn adjust the weights in each layer accordingly to minimize the error. The network repeats forward and back propagation until convergence, meaning that the weights are “learned” to accurately predict an output. In image processing tasks, the filter (or kernel) is slid across the image, which can make traditional DNN methods very computationally expensive. However, convolutional neural network (CNN) utilizes its filters to convolve or slide across the image to produce an activation at every slide position, producing a feature map. Since the filter can screen the whole image (either 2D or 3D), it allows the network to extract features based on the data points incorporating in their neighboring information. This efficient feature extract manners of CNN boosted many applications in MRI with good performance, such as tissue segmentation, brain tumor segmentation and SCI segmentation. 160–162 . A CNN consists of an input and an output layer, connected by multiple hidden layers. Typically, the hidden layers include a series of convolutional layers, pooling layers and deconvolutional (or transposed convolutional in some publications) layers, connected by dot production. CNNs learn sets of convolution kernels that are specifically trained for classification by optimizing sets of kernels based on the provided training data. In this way, the model can automatically extract objects that are relevant to the task. Particularly, the spatial (the lesions are located in the WM) and intensity information (the lesions appear brighter) are important to distinguish between classes (lesion/non-lesion), and optimized sets of kernels can be learned from the provided information and mimic how a human observer would distinguish WMH out of normal-appearing WM. We implemented a CNN for a WMH classification task, taking T2 FLAIR images in patients with SCD as input data. To reduce the computational complexity, each input image was subdivided into equal-size patches and sampled according to the intensity histogram (shown on Figure 1.19). The patches were screened to preserve the ones containing WMHs, as these patches of interests showed higher mean intensity levels or maximum intensity values than normal appearing ones. Then selected patches were entered a multiple-layer CNN in parallel. Our designed CNN mainly consisted of convolution layers and pooling layers. For each convolutional layer, the image patches were scanned by a convolution kernel in a specific stride and done convolution 40 operations. The local features such as intensity and texture were then extracted and mapped to another feature space, which had lower dimension and improved generalization (i.e. for specific patches in different patient images, the contrast between lesion/normal will be stronger and the variance of feature values for internal classes will be smaller). For each pooling layer, inputs were scanned in a filter-size stride so that there was no overlap. During the pooling process, inputs were represented by local maximum (max pooling) or mean values (mean pooling) to reduce the computational complexity. In the last layer of CNN, features were fully connected as feature map used for training and prediction. The training step is the optimization of convolution kernels and pooling filters based on comparing training results and ground truth. After training, the optimized CNN system predicts the lesion voxels based on a feature map and generate WMH detection results. The architecture of CNN classification model is presented in Figure 1.19. To create a SCI density map in a standard atlas, an accurate lesion segmentation is merely the first step. Given that SIT trail dataset were collected following clinical imaging protocol, more than 60% of the scans were in non-isotropic low resolution. Registering a low, non-isotropic resolution image to a high, isotropic resolution template while keeping the fidelity of lesions is quite challenging using conventional neuroimaging post processing tools. Fortunately, with the development of adversarial network in image generation, it is feasible to produce a high-resolution like image of the specific scan, to make the registration much easier. In the next section, adversarial generative network 163 will be introduced and tailored to our application in super-resolution reconstruction. Figure 1.19: The architecture of CNN classification model using T2-FLAIR MR images. 41 Section II: Super-resolution recovery Figure 1.20 presents an example of a clinical scan. In most of the multi-slice two- dimensional (2D) acquisitions, the through-plane (axial in this case) image is in high resolution (0.86 by 0.86 in this example) but with wide slice thickness (6mm in Figure 1.20). Therefore, it exhibits a salient anatomic difference between neighboring slices (cut position shown in yellow and blue in Figure 1.20), when comparing Figure 1.20C and D. Another problem lies in the non- isotropic resolution, because the in-plane views are either 0.86 by 6 or 6 by 0.86 mm 2 , as displayed in Figure 1.20 A and B. This situation makes it difficult for atlas registration, multi-modality analysis, and group comparisons. Therefore, through-plane slices are to be retrieved to produce isotropic resolutions in the orthogonal views, which is termed as single image super-resolution (SR) in the literature. Since the known variables in low resolution (LR) images are generally outnumbered by the unknowns in the desired high-resolution (HR) images, this problem is highly ill-posed, and can be mathematically represented as 𝒀=𝑓 𝑿 , where Y is the observed data, or LR image, X is the unknown HR images, and 𝑓 is the image degradation. The aim of SR recovery is to find 𝑓 eu ∙ and minimize the error between the estimated result 𝑿 and HR images X. Over the past decades, a Figure 1.20: A clinical T2-weighted image from SIT Trail dataset. A: coronal view; B: sagittal view; C and D: axial view of the neighboring slices shown on A and B in yellow and blue, respectively. Resolution: 0.86 x 0.86 x 6mm 3 , dimension: 256*256*22. 42 large number of single image SR methods have been proposed in computer vision tasks, many have been implemented in MRI SR recovery applications 164 . Previous work on single image super resolution using deep learning Conventional interpolation methods are simple and fast; these methods include nearest neighbor, bi-linear, and bi-cubic. But they generally cause artefactual noise, due to heavy reliance on neighboring voxel intensities. Recently, drawing on the success achieved by CNN implementation in other computer vision tasks, there have been increasingly more applications in MRI SR recovery using CNNs. Pham et. al. proposed a fully-connected a 3D CNN 165 , using 3D- based SR-CNN approach to produce HR 3D images by fusing 2D HR images based on the parameters published in the Conference on Computer Vision and Pattern Recognition (CVPR) 166 . The network achieved better performance in PSNR and SSIM than spline interpolation, non-local means upsampling, and low-rank total variation 167 . Yuhua et. al presented 3D Densely Connected Super-Resolution Networks (DCSRN) 168 , and the results outperformed their 2D counterparts and a previous method for faster SR-CNN 169 . CNN-based SR method can generate satisfying HR results. However, the widely used optimization methods of CNNs are voxel-wise error based between estimated and the ground truth images, such as mean squared error (MSE), or signal-to-noise (SNR), which lead to low visual quality. More recently, adversarial generative network (GAN), proposed by Goodfellow, et al 163 , provided a powerful framework for generating plausible-looking images in SR problems, referred to as SRGAN. SRGAN was firstly proposed by Ledig and colleagues in CVPR, with a perceptual loss function. SRGAN can generate photo-realistic natural images, and was adapted to MRI SR problems by Sanchez and Vilaplana 170 . Compared to SR-CNN, SRGAN can restore images with more realistic characteristics, including more local textures and details. It is approximately six times faster when implementing same publicly available brain structural MRI dataset 171 . The basic framework of GAN will be introduced below. Adversarial Generative Network (GAN) GANs are generally composed of two components: generator(s) and discriminator(s). A generator is a generative model since it tries to model the data distribution, or the features given a label, and then generate new data that look like the given data. A discriminator, on the other hand, tries to distinguish between the distribution of a real data and the “faked” data. The error is back propagated to update the discriminator learning weights. Then the discriminator is fixed while the 43 generator re-samples the data and is updated by back-propagated errors. In this way, the generator is “learned” to be a better at faking more similar data to fool the current discriminator, and the discriminator is trained to be a better classifier. The reason that a GAN is named “adversarial” is because these two components compete and try to outwit each other. This process iterates until the discriminator and generator reach equilibrium. To extend this framework to generate images, the generator, as a neural network, takes in a random variable vector (typically following Gaussian distribution) and generates an image I generator with a distribution p g (G). The discriminator (also a neural network) takes in the I generator and the real image, I real with the distribution p x and outputs a probability P generator to decide how likely the image is real. More concretely, the discriminator maximizes log(1- P generator ) when converging to 𝐷 𝑥 = ¯ ° ¯¤r ¯ ¡ . After updating the generator, by resampling the data with the mapping 𝑥 =𝐺 𝑧 , the updated distribution p g flows to the region that is more likely to be classified as real data. After a few rounds of training, they will meet at a point where both networks cannot improve because p g =p x. The discriminator can no longer differentiate between the two distributions. Mathematically, the discriminator is updated by ascending its stochastic gradient: ∇ ³´ u µ 𝑙𝑜𝑔𝐷 𝑥 (¶) +log 1−𝐷 𝑧 (¶) µ ¶·u (17), where ∇ is denoted as the gradient, m is the number of data points, and z is the sample mini-batch of m noise samples 𝑧 u ,𝑧 µ ,…,𝑧 µ . By maximizing the first item in the equation above, the discriminator can better recognize the real image while by maximizing the second item, it can better recognize the generated images. On the other hand, the generator is updated by stochastic gradient descend (SGD): ∇ ³´ u µ log 1−𝐷 𝑧 (¶) (18). By minimizing equation (18), the generator is better at fooling the discriminator. Taken together, GAN follows a two-player mini-max game with a value function V(G,D): min ¹ max 𝑉 𝐷,𝐺 = 𝔼 ¤~¯ ½¾¾ ¤ 𝑙𝑜𝑔𝐷 𝒙 +𝔼 ©~¯ ½¾¾ © log 1−𝐷 𝒛 (19), 44 where the first term is the entropy that the data from p x passes through the discriminator, and the discriminator tries to maximize it to 1. The second term is the entropy that the data from the random samples p g (z) passes through the generator, and the discriminator tries to minimize to 0. In most of the applications using GANs, the value function V is designed with regularization parameters, and the individual network, discriminator and generator are designed with various mathematical and statistical elements. For example, conditional-GAN uses a specific condition or characteristics rather than a generic sample from Gaussian noise distribution to feed to the generator. W-GAN utilizes Wasserstein distance instead of entropy in the value function V(G,D). Another novelty of the GAN framework is cycle-GAN, which contains two generators and two discriminators and the loss-functions of adversarial loss and cycle consistency loss. Cycle- GAN performed very well in image translation in natural images. The detailed work using GAN- based super-resolution is arranged in Chapter 3. 45 Chapter 2: Impaired white matter oxygen delivery in anemic patients Abstract Although modern medical management has lowered overt stroke occurrence in patients with sickle cell disease (SCD), progressive white matter (WM) damage remains common. It is known that cerebral blood flow (CBF) increases to compensate for anemia, but sufficiency of cerebral oxygen delivery, especially in the WM, has not been systematically investigated. Cerebral perfusion was measured by arterial spin labeling in 32 SCD patients (age range: 10-42 years old, 14 males, 7 with HbSC, 25 HbSS) and 25 age and race-matched healthy controls (age range: 15- 45 years old, 10 males, 12 with HbAS, 13 HbAA); 8/24 SCD patients were receiving regular blood transfusions and 14/24 non-transfused SCD patients were taking hydroxyurea. Imaging data from control subjects was used to calculate maps for CBF and oxygen delivery in SCD patients and their T-score maps. Whole brain CBF was increased in SCD patients with a mean T-score of 0.5 and correlated with lactate dehydrogenase (r 2 = 0.58, p<0.0001). When corrected for oxygen content and arterial saturation, whole brain and grey matter (GM) oxygen delivery were normal in SCD, but WM oxygen delivery was 35% lower than in controls. Age and hematocrit were the strongest predictors for WM CBF and oxygen delivery in patients with SCD. There was spatial co- localization between regions of low oxygen delivery and white matter hyperintensities on T2 FLAIR imaging. To conclude, oxygen delivery is preserved in the GM of SCD patients, but is decreased throughout the WM, particularly in areas prone to WM silent strokes Introduction Stroke is the most devastating complication in patients with sickle cell disease (SCD). It is not only a leading cause of death, but also impairs mobility, cognitive function and quality of life in these patients. The cooperative study of sickle cell disease demonstrated that 11% of patients had an overt stroke by age 20 and a cumulative lifetime stroke risk of 40% 16,172 . Routine transcranial Doppler (TCD) screening and blood transfusion therapy have lowered this risk ten- fold 22,23,26 . Despite this, risk of cerebral silent infarction (SCI) in SCD patients increases linearly with age, with a prevalence of 50% by the age of 30 37,38 . Non-invasive neuroimaging has been a powerful tool to improve the selection of patients with high stroke risk for treatment 173 . Nonspecific white matter (WM) atrophy, including volumetric decrease and microscopic alterations have been demonstrated in SCD using magnetic 46 resonance imaging (MRI) 48,49,61 . Although recent studies have reported WM cerebral blood flow (CBF) is lowest in the regions of highest SCI density 113 , and that SCI occur in regions of increased oxygen extraction fraction 174 , no one has actually determined whether WM oxygen delivery is impaired to WM regions at risk. We reported that SCD and other anemic patients preserve resting whole brain O 2 delivery, despite their impaired O 2 carrying capacity, by increasing their cerebral blood flow (CBF) 106,175 . However, it is unknown whether O 2 delivery is preserved regionally, across the brain. Characterization of O 2 delivery and its spatial association with WM lesions might help to identify brain regions that are most vulnerable to overt stroke. Arterial spin labeling (ASL) is a non-invasive and contrast-free MRI imaging method to measure CBF, from which cerebral O 2 delivery can be derived. In this work, we compare CBF and O 2 delivery between 32 patients with SCD and 25 healthy age and race-matched control subjects. Imaging data from control subjects was used to calculate maps of CBF and O 2 delivery in SCD patients. We used regression analysis to identify clinical and laboratory predictors of impaired O 2 delivery in white and grey matter (GM), and tested whether silent cerebral infarctions occurred in regions of poor O 2 delivery. Methods Population The study was approved by the Institutional Review Board (CCI-11-00083) at Children’s Hospital Los Angeles (CHLA). All participants were recruited from June 2016 to August 2017, with consent or assent. Two cohorts were studied: the first consisted of SCD subjects with hemoglobin SS (N=25) and SC (N=7) genotypes. Exclusion criteria included previous overt stroke, pregnancy and hospitalization within the month prior to the study visit. Children younger than nine years of age were excluded because of inability to cooperate with study procedures but there was no upper limit for age. All the SCD patients at our institution are regularly screened by TCD ultrasonography to identify patients at risk of stroke. The patients who have a blood flow velocity > 200 cm/sec are placed on monthly blood transfusions to suppress hemoglobin S% to less than 30%, according to standard clinical practice. Patients with normal Doppler velocities are placed on hydroxyurea and titrated to maximum tolerated dose 176 . The second cohort was composed of age- matched healthy controls, with or without sickle cell trait. Most were first or second-degree relatives of the patients studied or recruited from friends of patients’ families, so that two cohorts 47 are ethnically and socioeconomically-matched. Controls were excluded if they had developmental delay, a prior history of neurologic insult, or serious, uncontrolled, chronic illness. Data acquisition On the same day as the imaging study, complete blood count, hemoglobin electrophoresis, and cell free hemoglobin levels were analyzed in the clinical laboratory. Brain MRI was performed on the same day using a Philips Achieva 3 Tesla scanner with an 8-channel head coil. Arterial O 2 saturation was measured from a fingertip pulse oximeter during the MRI scan. In the patients undergoing blood transfusion therapy, MRI was performed immediately prior to the blood transfusion. For each subject, a 3D T1-weighted image (echo time (TE) = 3.8 ms, repetition time (TR) = 8.3 ms, resolution 1 mm 3 with a SENSE factor 2) and T2-weighted fluid attenuated inversion recovery (FLAIR) image (TE = 2.5 ms, TR = 4.8 ms, resolution = 1.3 × 1.0 × 1.0 mm 3 ) were acquired for co-registration and screening for WM lesions. High resolution time of flight MR angiography (resolution = 0.5 × 0.5 × 1.5 mm) was also performed at the Circle of Willis to screen for arterial stenosis. The field of view also spanned the supraclinoid, cavernous, petrous and 1-2 cm of the cervical portions of the internal carotid artery. Extracerebral vessels were also profiled by 3D angiography collected at a resolution of 1.5 x 1.7 x 3 mm. Pseudo continuous arterial spin labeling (PCASL) scans were performed using an unbalanced, Hanning shaped RF pules labeling train (mean gradient of 1 G/cm, interpulse interval of 1 ms and pulse duration of 0.5 ms), and a 3D GRASE two-shot readout (TE = 9.8 ms, TR = 3800 ms, resolution = 3.7 × 3.7 × 10 mm 3 , labeling duration of 2000 s, post labeling delay of 1600 s with 10 dynamics, EPI factor of 5). Two timed inversion pulses for background suppression following labeling were used, with an inversion efficiency of each background suppression pulse of 95%. Pre-processing and WMH segmentation We co-registered T1- and T2-weighted images to the 2 mm 3 Montreal Neurological Institute (MNI) atlas using FLIRT and FNIRT 177 . The probability maps of WM and GM were obtained using FMRIB’s Automated Segmentation Tool from registered T1 images 154 . The T2 FLAIR images were read for white matter hyperintensities (WMH) by a licensed neuroradiologist who was blinded to disease status. A WMH was considered significant if it was greater or equal to three millimeters in two orthogonal planes 35 . Although some WMH were observed in the control population, we limited WMH atlas generation to SCD patients. We used our in-house MATLAB 48 toolbox published previously to semi-manually segment WMHs 178 from several subjects. These lesions, and transformations of these lesions, were used to train a deep learning based method 179 capable of detecting all of the WMH in the remaining patients. While the sensitivity of the deep learning method was excellent, all detected lesions were manually edited using ITK-SNAP 180 to remove false alarms; final lesion morphometry was confirmed by a neuroradiologist (BT). WMH maps for each SCD subject were transformed to the MNI atlas, and fused to create a binary mask that localized WMHs across all the SCD subjects. Regional CBF was calculated using the PCASL images, rigidly co-registered to the T1 images and then to MNI space using FLIRT 177 . Voxel-wise gray and WM probability maps were derived at the same resolution as the PCASL images in order to quantify tissue CBF in native space as well as to perform partial volume effect correction 181 . Finally, we applied the whole brain CBF to the cerebral arterial territory masks to measure regional CBF in three major cerebral arterial territories: anterior cerebral artery (ACA), middle cerebral artery (MCA) and posterior cerebral artery (PCA). The mask of major vascular territories was based on a published template of vascular territories in both hemispheres 90 . The territories were drawn based on an extensive overview of anatomic studies of cerebral vascularization, and evaluated on the biocommissural plane. CBF quantifications CBF quantification was performed using a modified two-compartment kinetic quantification model proposed by Wang et al 121 . The parameters in the equation for quantification used in this work were previously described in detail by Bush et al 124 for a smaller cohort, using identical imaging protocols and quantification parameters 121 . Briefly, we calculated labeling efficiency on a patient-by-patient basis, correcting for uneven radio-frequency excitation and differences in blood flow velocity 124,126 . Flow quantitation also employed disease specific estimation of blood T1, and patient and tissue specific arterial transit time estimates 182,183 . Complex difference was performed between tag and control images, so the norm of the signal difference was always positive. The median of 10 CBF dynamics was used for quantification and comparison. To alleviate the blurring due to the point spread function and mixing of tissue-specific signals, we performed partial volume effect correction, using a 3 × 3 × 3 regression model 181 . During registration to common atlas space, rare peripheral voxels would be designated Not-a- Number (NaN) by FLIRT. We converted these flow values to zero. 49 Oxygen delivery maps CBF maps were converted into cerebral O 2 delivery maps to correct for patient-specific differences in hemoglobin and O 2 saturation. The following equations show the relationship of hemoglobin, O 2 content and O 2 delivery 184 : Oxygen Delivery = CBF × Oxygen Content (20) Oxygen Content = 1 .34 × Hemoglobin×SpO 2 + 0.003 × pO 2 , (21) where SpO 2 is the arterial O 2 saturation, and pO 2 is the partial pressure of O 2 , which is assumed to be 100 torr for room air. Statistical Analysis We used two-tailed, two sample t-tests to compare CBF and O 2 delivery in SCD and control subjects across the cerebral vascular territories. Predictors of whole brain, total GM and total WM CBF/ O 2 delivery were identified using univariate and stepwise multivariate regression. Test statistics with two-sided p<0.05 were considered to be statistically significant. All statistical analyses were performed using JMP Pro 11 (SAS, Cary, NC). Systematic spatial differences in CBF and O 2 delivery were identified using T-maps. The voxelwise mean and standard deviation, µ and σ, were first calculated from the control subjects. Voxel-wise 3D T-score maps were calculated for both CBF and O 2 delivery by calculating t = (X– µ) / σ on a voxelwise basis where t is the T-score value of each voxel, and X is the group average of CBF (or O 2 delivery) value in patients with SCD. To determine the spatial concordance of CBF and O 2 delivery between areas with white WMHs and normal appearing WM (NAWM), we performed a permutation analysis of WM lesions on the CBF maps and O 2 delivery maps. We created a WM lesion risk map by summating all the WMH segmented from T2-FLAIR images in 17 patients with SCD. The test distribution consisted of CBF values associated with each lesion, generating a distribution of CBF values in regions known to be at high risk from strokes. To create the corresponding null statistic, we randomly positioned each WM lesion 100 times within the NAWM regions and recalculated the CBF. The shape of the two distributions was compared by Kolmogorov-Smirnov test, while the mean and variance of the distributions were compared by t-test and ratio test, respectively. The permutation test on O 2 delivery map followed the same procedure. This was done using our in-house MATLAB code. 50 CTL (N=25) SCD (N=32) p-value Transfusion Status Not Applicable Transfused N=8 Non-transfused N=24 Age 23.0±8.6 19.8±7.2 23.2±9.1 0.4 Sex F15, M10 F6, M2 F10, M14 1.0 Ethnicity 22 African American, 2 Hispanic, 1 Middle East 28 African American, 4 Hispanic Hemoglobin (g/dl) 13.2±1.2 8.4±1.2 10.1±2.0 <0.05 Hematocrit (%) 40.0±3.6 24.8±3.0 28.1±5.2 <0.0001 %A Hemoglobin 81.7±17.8 2.8±0.2 4.4±8.9 <0.0001 %S Hemoglobin 18.5 ± 18.2 25.2±13.3 68.3±22.1 <0.0001 %F Hemoglobin 0.7±2.4 3.5±3.0 8.5±9.0 <0.0001 Plasma Hemoglobin 7.5±6.5 29.6±24.8 18.1±12.1 <0.05 Oxygen Saturation % 98.2±1.4 98.3±1.7 96.0±1.9 <0.05 Hydroxyurea Treatment 0 0 14 (12HbSS) NSAID Treatment 0 5 (4HbSS) 1 (HbSS) Table 2.1 Patient Demographis. Continuous variables presented in mean ± standard deviation. Comparison of continuous variables was made using pooled, two-tailed t-tests. Comparison of categorical variables (sex) was made using a Fisher’s exact test. All the tests and p-value was the comparison between patients with SCD (N=32) and healthy controls (N=25). CTL = controls, SCD = sickle cell disease, WMH = white matter hyperintensity, Hb=hemoglobin. Results Demographics Patient demographic information is summarized in Table 2.1. There is no significant difference in age and sex between two cohorts. The healthy control (CTL) group included 12 participants with sickle cell trait (hemoglobin AS) and 13 with hemoglobin AA. The sickle trait subjects did not differ in any detectable way from the control subjects (hemoglobin AA) other than their hemoglobin electrophoresis. The SCD group consisted of 7 patients with hemoglobin SC, and 25 hemoglobin SS. Eight of the SS patients were receiving chronic transfusion therapy for a history of abnormal transcranial Doppler (N=6) or recurrent acute chest syndrome (N=2). Chronically transfused patients were maintained with a pre-transfusion hemoglobin S level below 30% and had relatively low hemoglobin F%, reflecting good suppression of endogenous erythropoiesis. Chronically transfused patients were studied at their hemoglobin nadir to better match their hemoglobin levels with non-transfused subjects. Hydroxyurea was being taken by 12/17 nontransfused SS patients and 2/7 SC patients. In addition to having decreased hemoglobin 51 and hematocrit, SCD patients were mildly desaturated compared to healthy controls; three SCD patients has O 2 saturations below 93% while the rest had O 2 saturations from 97% to 95%. MR angiography images were read by a licensed neuroradiologist for all participants. All subjects had normal MRA except one 37-year-old patient with bilateral anterior cerebral artery stenosis; this patient was excluded from the analysis. Small, WMH were found in 12/24 non-transfused SCD patients and 5/8 transfused SCD patients. Five of 17 WMH+ patients were HbSC phenotype. CBF and Oxygen Delivery The top panel of Figure 2.1 compares the spatial distribution of CBF in CTL and SCD patients. In both groups, CBF is greater in grey matter than white matter, but appears to be particularly increased in SCD patients. To facilitate a comparison across groups, the top-right panel of Figure 2.1 demonstrates a T-score map of regional CBF between SCD and control subjects. CBF is equally elevated in grey matter structures (T-score > 2) but is normal or decreased in white matter. The average T-score value for whole brain CBF is 0.5. Table 2.2 shows the quantitative summary of all whole brain and regional CBF measurements for two groups. CBF in both whole Figure 2.21: Whole brain CBF map, oxygen delivery map and T-score maps in axial view. Top row: average CBF map of healthy controls (N=25), sickle cell patients (N=32) and T-score. The color-bar represents absolute perfusion in ml/100g/min. Bottom row: average oxygen delivery map of healthy controls (N=25), sickle cell patients (N=32) and T-score. The color-bar depicts oxygen delivery in milliliter of oxygen /100g brain tissue per minute. 52 brain and three territories are significantly higher for SCD patients than healthy controls (p<0.01). In grey matter, CBF is elevated across the brain, while in WM, there was no significant difference. Regions CTL (N=25) SCD (N=32) P-value (adjusted) CBF (ml/100g/min) Whole brain Total 54.7 ± 12.9 72.2 ± 21.0 0.0006 GM 63.0 ± 14.7 84.0 ± 26.5 0.0006 WM 41.2 ± 11.6 37.62 ± 9.6 0.2 ACA Total 46.5 ± 9.7 74.7 ± 27.3 <0.0001 GM 53.4 ± 11.5 87.0 ± 33.8 <0.0001 WM 34.9 ± 8.4 38.8 ± 12.4 0.2 MCA Total 58.4 ± 14.0 79.1 ± 23.5 0.0003 GM 67.0 ± 16.7 92.1 ± 29.6 0.0004 WM 44.0 ± 12.1 41.3 ± 11.0 0.4 PCA Total 56.4 ± 12.5 74.3 ± 22.9 0.0009 GM 64.7 ± 12.9 86.5 ± 28.9 0.001 WM 42.5 ± 14.7 38.6 ± 9.7 0.2 Oxygen Delivery (ml/100g/min) Whole brain Total 11.3 ± 3.0 10.4 ± 2.2 0.2 GM 12.9 ± 3.5 12.0 ± 2.8 0.3 WM 8.5 ± 2.7 5.5 ± 1.3 <0.0001 ACA Total 9.6 ± 2.2 10.6 ± 2.9 0.1 GM 11.0 ± 2.6 12.3 ± 3.6 0.1 WM 7.2 ± 2.0 5.6 ± 1.5 0.001 MCA Total 12.1 ± 3.4 11.2 ± 2.4 0.3 GM 13.8 ± 4.0 13.1 ± 3.1 0.4 WM 9.1 ± 2.9 5.9 ± 1.4 <0.0001 PCA Total 11.6 ± 2.9 10.6 ± 2.3 0.1 GM 13.4 ± 3.5 12.3 ± 3.0 0.2 WM 8.8 ± 2.7 5.5 ± 1.2 <0.0001 Table 2.2: Global/regional CBF and oxygen delivery. Measurements presented as global (GM/WM) with mean ± standard deviation, values in bold indicates significant difference with p<0.01. CTL = controls, SCD = sickle cell disease, CBF = cerebral blood flow, ACA = anterior cerebral artery, MCA = middle cerebral artery, PCA = posterior cerebral artery. To control for the significant differences in O 2 content between the two groups, we converted CBF maps into O 2 delivery maps following Equation (1). The bottom panel of Supplemental Figure 1 compares regional O 2 delivery maps between the two groups; the rightmost panel depicts the corresponding T-score map. The WM areas exhibit cooler color on the T-score map, indicating decreased O 2 delivery in SCD group; while in GM, the T-score is around zero, meaning the O 2 delivery in GM is quite similar between SCD and CTL groups. The average value 53 of the T-score for white matter O 2 delivery was -0.4 (0.4 standard deviations below the mean). The quantitative whole brain and regional O 2 delivery results are summarized in the bottom panel of Table 2.2. Whole brain and grey matter O 2 delivery in SCD was completely normal (p=0.2 and 0.3 respectively, using one-way Anova test), but was 35% lower in the white matter (p<0.001, using one-way Anova test), consistent with what was observed on the T-score map. Brain Perfusion Sickle Cell Disease Patients Predictor Total Perfusion GM Perfusion WM Perfusion Plasma Hemoglobin r 2 =0.38, p=0.0002 r 2 =0.38, p=0.0002 - Lactate Dehydrogenase r 2 =0.58, p<0.0001 r 2 =0.54, p=<0.0001 r 2 =0.23, p=0.006 Reticulocyte Count r 2 =0.27, p=0.002 r 2 =0.27, p=0.002 - Mean Platelet Volume r 2 =0.23, p=0.005 r 2 =0.22, p=0.007 - Heart Rate r 2 =0.29, p=0.002 r 2 =0.24, p=0.004 r 2 =0.19, p=0.01 Weight - - r 2 =0.16, p=0.02 Age - - r 2 =0.20, p=0.01 MCHC r 2 =0.12, p=0.05 r 2 =0.12, p=0.05 - Oxygen Content r 2 =0.37 p=0.0002 r 2 =0.34 p=0.0004 r 2 =0.14, p=0.03 Hemoglobin r 2 =0.36, p=0.0003 r 2 =0.33, p=0.0006 r 2 =0.17, p=0.03 Hematocrit r 2 =0.37, p=0.0002 r 2 =0.33, p=0.0006 r 2 =0.22, p=0.007 RDW r 2 =0.09, p=0.09 r 2 =0.10, p=0.07 r 2 =0.14, p=0.03 White blood cell count r 2 =0.12, p=0.05 r 2 =0.11, p=0.06 - Oxygen Delivery in Sickle Cell Disease Patients Predictor Total O 2 Delivery GM O 2 Delivery WM O 2 Delivery Plasma Hemoglobin r 2 =0.30, p=0.001 r 2 =0.33, p=0.0007 - Lactate Dehydrogenase r 2 =0.10, p=0.07 r 2 =0.11, p=0.07 - Mean Platelet Volume r 2 =0.17, p=0.02 r 2 =0.17, p=0.02 - Hemoglobin S% r 2 =0.10, p=0.07 r 2 =0.11, p=0.06 - Age - - r 2 =0.24, p=0.005 MCHC - - r 2 =0.22, p=0.007 Oxygen Content - - r 2 =0.18, p=0.02 Hemoglobin - - r 2 =0.15, p=0.03 Hematocrit - - r 2 =0.09, p=0.09 RDW - - r 2 =0.14, p=0.03 Reticulocyte Count - - r 2 =0.14, p=0.04 Table 2.3: Predictors of SCD brain perfusion and oxygen delivery. Bold lettering indicates retention on multivariate analysis. Abbreviations: Mean corpuscular hemoglobin concentration (MCHC), Red cell distribution width (RDW), Grey matter (GM), White matter (WM), Oxygen (O 2 ) 54 To further investigate the relative contributions to CBF and O 2 delivery, we performed univariate and multivariate regressions using demographic factors and blood test results (Table 2.1). In the SCD group, whole brain CBF and GM CBF were inversely correlated with O 2 content, hemoglobin, and hematocrit, but positively correlated with lactate dehydrogenase (LDH), cell-free hemoglobin, reticulocytes, mean platelet volume (MPV) and heart rate on univariate analysis. On multivariate analysis, only LDH remained in the model for both whole brain (r 2 = 0.58, p<0.001) and GM CBF (r 2 = 0.54, p<0.001). The WM CBF had similar univariate correlates except that cell- free hemoglobin, reticulocytes, and MPV did not reach statistical significance, and WM CBF decreased significantly with age (r 2 =0.20, p=0.01). On multivariate analysis, LDH, age, and hematocrit predicted WM CBF with a combined r 2 of 0.45 (p<0.001). When O 2 delivery was examined instead of CBF, many factors related to anemia were no longer significant for whole brain and grey matter O 2 delivery. Cell-free hemoglobin and mean platelet volume were the only significant predictors, with only cell-free hemoglobin surviving multivariate analysis (Table 2.2). WM O 2 delivery was positively correlated with hemoglobin, hematocrit, O 2 content, and mean corpuscular hemoglobin concentration (MCHC) and inversely Figure 2.22: White matter lesion mask overlaid on T-score map. The scaled WM lesion mask in its axial view is overlaid on top of the T-score for CBF (top row) and O2 delivery (bottom row). The number z indicates the coordinates in the standard atlas 55 correlated with age, red cell distribution width (RDW) and reticulocytes. Age, O 2 content, and hematocrit remained in multivariate analysis with a combined r 2 of 0.46. Figure 2.2 compares the spatial relationship between CBF, O 2 delivery and WMH. The color-bar represents the T-score and Z indicates the position of the axial slice in the atlas. The lesion mask was gray-scaled and shown in white. WMH were all distributed in areas having negative T-values on the O 2 delivery map. Figure 2.3A and B (dashed lines) summarize the CBF and O 2 delivery in the WMH prone regions, compared with values in normal appearing white matter (NAWM, solid lines). The average value for CBF in WMH prone regions was 30.0 compared with 38.9 ml/100g/min in NAWM (p<0.001). The average value for O 2 delivery in WMH prone regions was 4.3 ml/100g/min compared with 5.9 ml/100g/min in NAWM regions 02468 10 12 14 16 18 0.05 0.1 0.15 Oxygen Delivery in WMH areas Oxygen Delivery in NAWM areas (B) Oxygen Delivery (ml/100g/min) -2 -1.5 -0.5 0.5 1.5 2.5 0 0.05 0.1 0.15 T-score in WMH areas T-score in NAWM areas -2 -1.5 -0.5 0.5 1.5 2.5 0 0.05 0.1 0.15 T-score in WMH areas T-score in NAWM areas T-score of Oxygen Delivery T-score of CBF (D) 020 40 60 80 100 120 140 0.05 0.1 0.15 CBF in WMH areas CBF in NAWM areas (A) CBF (ml/100g/min) (C) Figure 2.23: Boxplot comparisons of PC CBF estimates (left) and relative error between PCASL and PC CBF (right) among subjects with and without venous outflow. Patients with venous outflow displayed larger absolute CBF and relative error than subjects without venous outflow (p<.05). 56 (p<0.001). Figure 2.3C and D depict the same relationships but converts them into T-scores. The mean CBF T-scores were 0.3 and 0.4 (p<0.001) in WMH prone and NAWM respectively, while the corresponding mean T-scores for O 2 delivery were -0.5 and -0.3 (p<0.001), respectively. Discussion There have been several studies focusing on quantifying CBF using ASL in SCD cohorts 104,107,124 , however, we are the first to quantify O 2 delivery in both healthy controls and SCD populations. We found that whole brain and GM CBF was increased in SCD patients, as a result of cerebral vasodilation to meet GM metabolic demands. Surprisingly, we did not observe a similar compensatory hyperemia in the WM. As a result, whole brain and GM O 2 delivery were preserved in SCD patients, but WM O 2 delivery was only 2/3 as high as for control subjects. Regions commonly associated with WMH had lower O 2 delivery than normal appearing WM. After performing univariate and multivariate regression, age and O 2 content were the strongest predictors for WM O 2 delivery in SCD patients. Increased CBF in patients with SCD has also been described in other studies using MRI 175 , PET 185 , and Xenon Computed Tomography 103 . The elevated CBF was found to be in part due to cerebral vasodilation in response of low hematocrit levels and high hemoglobin S concentration 186 . The mechanism by which anemia mediates increased CBF is not entirely known. Although blood viscosity decreases linearly with hematocrit, CBF increase cannot be explained by passive alterations in arterial blood viscosity 187,188 . Anemia, whether caused by SCD or acute hemodilution, produces cerebral tissue hypoxia 189,190 . In turn, cerebral hypoxia triggers arterial vasodilation acutely, and promotes capillary proliferation in GM through hypoxia inducible factor mediated signaling to shorten O 2 diffusion distances and lower cerebrovascular resistance 191,192 . On a whole organ basis, the increased CBF completely compensates for the inadequate O 2 carrying capacity 175,193 . However, compensatory vasodilation was not observed in WM, causing WM O 2 delivery to be directly proportional to hemoglobin concentration. Hematocrit-dependent cerebral O 2 delivery has been described in subjects with cardiovascular disease, with optimal delivery for a hematocrit of 40-45% 130 . WM O 2 delivery also declined dramatically with age. The etiology of this decline is unclear, but likely reflects accelerated geometric and morphologic alterations, or even destruction of the microvascular network 194 . Our observed decline in WM O 2 delivery with age parallels the prevalence of WMH in SCD patients observed by Kassim 38 and DeBaun 46 . 57 In contrast, GM O 2 delivery is normal in patients with SCD, reflecting the powerful inverse relationship between CBF and O 2 carrying capacity 101,195 . The preserved GM O 2 delivery we observed is consistent with the lack of GM atrophy we observed in our previous reports 11 . GM O 2 delivery was stable with age, however the median age of our SCD cohort was only 21.4 years of age. Thus, we may have been underpowered to observe age-related reductions in GM O 2 delivery. Surprisingly, cell-free hemoglobin was the strongest predictor of GM O 2 delivery. Cell free hemoglobin is a toxic byproduct of hemolysis and is associated with impaired nitric oxide metabolism 6 and endothelial dysfunction 196 . However, its impact on cerebral nitric oxide generation and cerebral endothelial function has not been reported. A healthy microvasculature is necessary for matching O 2 supply and consumption. Bush et. al. have previously documented evidence of arterio-venous shunting and impaired O 2 unloading in SCD patients 197,198 , both suggestive O 2 supply-demand mismatch. Alternatively, cell-free hemoglobin could have a more direct vasodilatory effect. Cell free hemoglobin is metabolized by heme oxygenase, producing free iron and carbon monoxide, a potent cerebral vasodilator 6,199 . In a parallel vascular circuit, excessive vasodilation of one bed could produce ischemia in another. Whether the decreased WM O 2 delivery represents vascular stealing by the grey matter, or simply inadequate neovascularization in response to chronic hypoxia could not be determined. The WMHs and O 2 delivery map overlay demonstrates two important points. Firstly, CBF and O 2 delivery were lower in regions prone to WMH, than in those with NAWM, independent of disease state (Figure 2.3C and D). It implies that WMHs are occurring in borderzone regions where perfusion is inherently decreased in anyone. Studies from Numaguchi, Baldeweg, Guilliams and Ford 64,103,113,200 in SCD patients also support that conclusion. Our O 2 delivery T-score maps agrees with the observations of Fields, et. al., who demonstrated increased O 2 extraction fraction in watershed areas 95 . However, the T-score for O 2 delivery was lower in regions at risk for WM stroke (T=-0.5) than in NAWM (T=-0.3) as shown in Figure 2.3D. The use of a T-score between groups and permutation corrections for the expected reduction in O 2 delivery based upon a watershed effect alone. Therefore, O 2 delivery is disproportionately impaired with WMH prone regions, as exhibited on Figure 2.2. Our study has several limitations. Firstly, measurement of WM CBF and O 2 delivery could be complicated by differences in blood transit time and longitudinal relaxation of blood, grey matter, and white matter 105,181,201 . While we used velocity-based corrections for transit time 58 derived from previous studies in SCD patients 183 , patient-wise measurement of transit times might improve the accuracy of the CBF estimation. Blood magnetic properties (T1 relaxation) are different in SS patients. While we used values appropriate for SCD patients, patient specific T 1 blood measurements could the CBF accuracy in SCD patients 202 . Blood T1 in SC patients has never been reported. Secondly, the intrinsic low resolution of ASL images and the use of a 3D ASL acquisition, produce admixture of CBF values from grey and white matter, known as partial volume effects. Our partial volume correction may not be sufficient performed to eliminate GM and WM cross- contamination. This will cause underestimation of GM CBF and overestimation of WM CBF, leading to a low GM/WM flow ratio; the thin, convoluted cerebral cortex (GM) is particularly vulnerable to CBF underestimation from this mixing effect. However, partial volume effects cannot explain low WM O 2 delivery observed in SCD patients. Nor can it explain the dependence of WM O 2 delivery on age and O 2 content because those effects were not observed in GM. The use of T-scores for group comparison controls for most of the partial volume biases because both SCD patients and healthy controls suffer from comparable GM/WM admixture. Thirdly, hematocrit and O 2 delivery at the microvascular level are different than in conduit vessels. It is possible that actual O 2 delivery could be slightly better, or worse, than our calculations based on venous hematocrit values. However, use a global, venous hematocrit value nonetheless yields intuitive results that provide important insights into WM disease in SCD patients. Lastly, our sample size was too small to explore important covariates such as the role of blood transfusion therapy or the compliance with hydroxyurea, nor could we explore links between O 2 delivery, brain volumes, and neurocognitive outcomes. Lastly, our SCD population was heterogeneous with respect to genotype and treatment. The goal of the parent study (CCI-11-00083) was to examine the impact of hemoglobin level and hemoglobin S% on cerebral blood flow and oxygenation, requiring a cohort having a broad range in those variables. We accomplished this objective, but at the cost of cohort homogeneity. While we were careful to control for possible group effects during statistical analysis, our subgroups were too small to make any inferences about the impact of transfusions, hydroxyurea, or genotype. Our study would have been strengthened if we could have studied chronically transfused patients prior to and following their transfusion visit but it was not feasible with our study budget. 59 In conclusion, we demonstrated that WM O 2 delivery is impaired in SCD patients. Cerebral compensatory mechanisms appear optimized to protect GM O 2 delivery, even while WM remains hypoxic. WMH occur in watershed areas, where CBF and O 2 delivery are intrinsically low, but O 2 delivery in these regions is even lower than uniquely predicted by watershed effects. Taken together, compensatory hyperemia preserves O 2 delivery to the GM in SCD patients but is insufficient to maintain O 2 delivery to the WM, explaining the distribution and progressive evolution of SCIs in SCD patients. 60 Chapter 3: White matter microscopic injury is associated with neurological performance Abstract SCD is a hereditary blood disorder in which the oxygen-carrying hemoglobin molecule in red blood cells is abnormal. SCD patients are at increased risks for strokes, cortical atrophy, and neurocognitive deficits. Although neurovascular screening and treatments have lowered the rate of overt strokes tenfold, progressive white matter (WM) damage and cognitive deficits remain common in SCD patients. Tract-specific analysis is a statistical method to evaluate microstructural WM damage in neurodegenerative disorders, using diffusion tensor imaging. We utilized TSA and compared 11 major brain WM tracts between SCD patients with no history of overt stroke, anemic controls and age-comparable healthy controls. Disruption of WM microstructure orientation dependent metrics for the SCD patients were found in the genu of the corpus callosum (CC), cortico-spinal tract, inferior fronto-occipital fasciculus, right inferior longitudinal fasciculus, superior longitudinal fasciculus, and left uncinate fasciculus. Neurocognitive test performance indicated slower processing speed and lower response inhibition skills in SCD patients compared to controls, implicating a relationship between microscopic damage to WM structural damage and neurocognitive function. TSA abnormalities in the CC were significantly associated with measures of processing speed, working memory, and executive functions, highlighting the important role the CC plays in core neurocognitive processes. Introduction Sickle cell disease (SCD) is an inheritable genetic disorder of red blood cells, in which a single base pair DNA mutation causes hemoglobin to polymerize upon deoxygenation, producing sickle-shaped red blood cells. It is a major public health concern, with over 300,000 children born with SCD each year worldwide, with incidence rates projected to increase to 400,000 by 2050 203 . Individuals living with the disease experience lifelong complications, including anemia, infections, stroke, tissue damage, organ failure, pain crises, and premature death 5,204 . While SCD affects many vital organs, damage to brain tissue is amongst the most concerning due to the profound personal, professional and social cost to patients 28,176 . Until recently, ten percent of children with SCD had a symptomatic, overt stroke, with incidence rates rising to 24% by 45 years of age 16 , though in the last 16 years, routine transcranial Doppler (TCD) 61 screening of the Circle of Willis and chronic transfusion therapy when indicated have lowered the risk tenfold 26,205 . However, while the risk for silent cerebral infarcts (SCI) is also reduced by transfusion, SCI remain a serious concern due to associated neurocognitive deficits in comparison to lesion-free patients 21,36 . Approximately 27% of children experience SCI by six years of age and 37% by 14 years of age, with ongoing risk of progressive injury, with the size and number of lesions increasing with age 37 , reaching up to 53% by 30 years of age 38 In addition to SCIs, which primarily damage the white matter (WM) in the brain, SCD patients also show reduced WM volume relative to controls 49,58,206 . These abnormalities may lead to affective and cognitive consequences over time, including depression and isolation due to physical and mental limitations, as well as poor neurocognitive outcomes 39,207,208 . In adolescent SCD patients, a range of academic and neurocognitive difficulties compared to healthy peers has been reported, including lower verbal IQ scores, poorer math performance, and impairments in visual-motor functions 44 . Additionally, there is a preponderance of evidence that individuals with SCD experience difficulties with working memory, executive functions and processing speed 45,46,209 , neurocognitive domains sensitive to WM compromise 210,211 . Brain magnetic resonance imaging (MRI) is commonly used for the detection of cerebrovascular damage. While clinical imaging protocols of MRI typically rely on qualitative assessments of T1 and T2-weighted images, diffusion tensor imaging (DTI) has proven to be a more sensitive technique to probe WM microstructure, characterize abnormalities of WM pathways in various neurological disorders 212 , and to detect preclinical neurologic ischemia 132 . Previous studies have used methods such as region of interest (ROI) analysis 61 , voxel-based analyses as in voxel-based morphometry (VBM) 64 , and WM tract-based methods such as Tract- Based Spatial Statistics (TBSS) 62,138 and Tract-Specific Analysis 48,63 , of which the latter two are WM tract-based methods. Balci et. al 61 , the first to use DTI in SCD cohorts, reported significantly reduced fractional anisotropy (FA) values and increased apparent diffusion coefficient values in the corpus callosum (CC) and cortico-spinal tracts (CST) using an ROI-based analytic approach. Likewise, Sun et. al 50 used TBSS and found similar outcomes in both the CC and CST for asymptomatic SCD patients. Nevertheless, few studies have systematically explored microscopic regional WM changes in relation to neurocognitive performance in patients with SCD 213 , but exploring this relationship may inform SCI prevention and early intervention efforts. 62 Although TBSS is a widely used tool, TSA has several advantages for examining WM differences in specific tracts across populations. TBSS lacks anatomical specificity because it constructs the skeleton for the entire WM, instead of separately for each individual WM tract, like TSA. TSA also makes it possible to distinguish between adjacent WM tracts, such as the CC and CST, two of the most prominent WM tracts. In TSA, WM fibers are segmented on a population specific template, and statistical comparisons are performed on DTI derived measurements projected on a medial sheet generated from these tracts. By focusing on specific tracts, it provides better localization of areas affected by neurological disorders. Additionally, recent studies comparing congenitally blind versus sighted control groups 214 , investigating 22q11.2 deletion syndrome 215 and preterm infants 216 , suggest that TSA can outperform TBSS in terms of statistical detection power. In this study, we utilize TSA to explore differences in major brain WM tracts: corpus callosum, right and left cortico-spinal tracts, inferior fronto-occipital fasciculus (IFO), inferior longitudinal fasciculus (ILF), superior longitudinal fasciculus (SLF), and uncinate fasciculus (UNC). The studied groups are: patients with SCD, anemic controls with normal hemoglobin, and healthy controls. We additionally examine performance on measures of processing speed, working memory and executive functions, as well the relationship between the most commonly used DTI metric of WM tracts and neurocognitive performance in the SCD patients and healthy controls. Methods Participants Following Institutional Review Board at Children’s Hosptial Los Angeles (CHLA), all participants provided consent or assent. Patients and controls were recruited from the CHLA patient population, their families and the community. Study participants completed a 75-minute MRI examination without sedating medications, limiting our study to subjects older than 10 years of age. Exclusion criteria for SCD patients included previous overt stroke, significant cerebral vasculopathy on previous imaging studies, acute chest syndrome, pregnancy, and pain crisis hospitalization within one month of the study. SCD patients at CHLA are regularly screened with TCD ultrasonography to identify patients at higher risk of stroke. Patients with a blood flow velocity faster than 200cm/sec read from TCD are placed on monthly blood transfusions to suppress the percentage of hemoglobin S to less than 30%, in accordance with standard clinical practice. 63 The healthy control group (CTL) consisted of subjects with no known chronic medical conditions, prior history of neurologic insults, or developmental delay. CTL subjects were recruited from friends and family of the SCD patients to better match for ethnicity, socioeconomic status, and other environmental factors. To control for possible changes related to anemia, independently of sickle hemoglobin, we also recruited patients with non-sickle chronic anemia syndromes (ACTL), including beta-thalassemia major, beta-thalassemia intermedia, and congenital dyserythropoietic anemia. Some of the patients in the SCD and ACTL cohort were on chronic transfusion. For these patients, MRI examination was performed prior to a regularly scheduled transfusion to better match the hematocrit to the non-transfused anemic patients. Neuroimaging Each subject underwent a MRI study using a 3T Philips Achieva with an 8-element array coil. Whole brain T2-FLAIR images were acquired to screen for WM hyperintensities. Acquisition parameters were: TR=4800 ms, TE=363 ms, FOV=256 × 256 mm, and voxel size of 1.3 × 1 × 1 mm. T2-FLAIR images were read for WM abnormalities by a board certified neuroradiologist, who was blinded to disease status. White matter hyperintensities (WMHs), indicating SCIs, were classified as 3-5 mm lesions on T2-FLAIR, observed in two orthogonal planes. These lesions had no known neurological sequelae. Diffusion weighted images (DWI) were obtained in 30 encoding directions at a b-value of 1000m/s 2 , and a reverse gradient b=0 using a single shot echo-planar imaging sequence (TR=6700 ms, TE=86 ms, FOV=240 × 240 mm, voxel size of 2.5 × 2.5 × 2.5 mm). Registration and 3D representations DWI data was visually inspected for major artifacts and signal drop off. The 4-D images were corrected for eddy-current introduced geometric distortions using the Tensor Toolkit (TTK), as part of the ITK software package (www.itk.org). In addition, we applied susceptibility-induced off-resonance field corrections by calculating field maps from reverse gradient b=0 images using TOPUP 217 . DWI images were skull-stripped using DSI studio (http://dsi-studio.labsolver.org), followed by tensor estimation using TTK. WM microstructural differences were compared in atlas space. The atlas in the current context is a study-specific DTI template constructed from healthy adults, i.e., a volume that captures the average shape and diffusion-based features of the entire cohort 218 . We conducted rigid, affine and non-linear registration using Diffusion Tensor Imaging Toolkit (DTI-TK) 219 to align 64 tensor images to the atlas space. We did not use scalar-based images such as FA because they discard orientation information, making it difficult to distinguish neighboring tracts with similar FA values but different orientations. Additionally, registering tensor images onto a template yields better alignment of the dominant diffusion orientation 215 . Rigid alignment was employed first to find an initial linear estimation of the original image in the template space. Affine alignment and a non-linear registration were performed using a deformation field map to improve alignment quality 220 . Tracts of interest were generated as binary segmentation using deterministic tractography 221 . Then a continuous medial representation (cm-rep) model 222 , a deformable modeling and shape analysis technique was formulated and a medial surface was approximated for each vertex, so that the skeleton and boundary were defined simultaneously for each tract 223 . Diffusion data from every subject was then projected onto the skeleton, by searching along the unit normal from the vertex to the tract boundary, which defines the stopping criteria 215 . This sampling strategy limits potential voxel mis-assignments from neighboring tracts. Maximum or mean value of the DTI metrics can be projected and sampled 224 . Statistical analysis of the projected diffusion metrics at each vertex on the skeleton are given in the following sections. In this work, maximum FA values were computed locally and projected onto the cm-rep model for the tracts of interest in 3D, as shown in Figure 3.1: the corpus callosum (CC), cortico- spinal tracts (CST), inferior fronto-occipital fasciculi (IFO), inferior longitudinal fasciculi (ILF), superior longitudinal fasciculi (SLF), and uncinate fasciculi (UNC). To quantitatively compare and correlate the FA values with neurocognitive scores, we calculated mean FA values for each subject in the atlas space in each tract. Figure 3.1: Medial representation of the 11 white matter tracts in coronal, axial and sagittal view. Red: corpus callosum (CC), green: cortico-spinal tract (CST), orange: inferior fronto-occipital tracts (IFO), magenta: inferior longitudinal tracts (ILF), cyan: superior longitudinal tracts (SLF) and purple uncinates (UNC). 65 Neurocognitive assessment All participants completed a three- to four-hour battery of standardized psychometric measures. From this battery, we chose to examine measures directly related to white matter integrity: processing speed (i.e. the speed of information processing), working memory (i.e. the ability to briefly hold and manipulate information in one’s mind) and executive functions (i.e. higher level cognitive skills that involve mental control and self-regulation). Executive functions assessed included inhibition (i.e. the ability to resist prepotent or impulsive responses), cognitive flexibility (i.e. the ability to shift one’s focus within or between tasks), and phonemic and semantic verbal fluency (i.e. rapidly generating words beginning with particular letters or in specific categories). For participants on chronic blood transfusions, testing was performed within one week of transfusion to minimize fatigue effects. Additionally, the NIH PROMIS fatigue scale was given to all participants to screen for fatigue. Testing was performed by the study neuropsychologist or by doctoral trainees under the supervision. Working memory was assessed with Digit Span from the Wechsler Intelligence Scale for Children, Fourth Edition (WISC-IV) 225 or the Wechsler Adult Intelligence Scale, Fourth Edition (WAIS-IV) 226 , processing speed with Coding and Symbol Search from the WISC-IV or WAIS-IV, and executive functioning with the Color-Word Interference, Trail Making and Verbal Fluency subtests of the Delis-Kaplan Executive Function System (D-KEFS) 227 . Statistical Analysis To compare the demographic characteristics and psychometric performance between the groups, two-sample t-tests and Pearson’s X 2 tests were used for continuous and categorical variables respectively. Bonferroni correction was used to correct for multiple comparisons. All neurocognitive tests were scored using the non-medical, age-adjusted normative data provided by the test publishers. To correct for the effect of outliers, we performed a nonparametric, one-way Wilcoxon Rank test on the results using JMP (JMP®12.1.0). Statistical differences in FA between the patient groups and healthy controls were assessed using a supra-threshold statistical model 219 . A two-sample t-test at each point on the skeleton surface of a tract was computed. An arbitrarily chosen t 0 was set to extract clusters on the surface for which t values are less than t 0 . The mass of each cluster (the area of the cluster in this context) was compared to a histogram of maximal cluster masses computed from a large number of identical experiments, so that the family-wise error rate (FWER) correction takes into account the 66 number of WM tracts included 228 . The threshold p-value was set to 0.01 and the number of permutations to 10,000. We included age and sex in the general linear model in the TSA pipeline to control for relevant confounding factors. To explore the quantitative FA value differences between groups, one-way ANOVA with repeated measures was performed on the FA values of the 11 tracts between the tree groups: SCD, ACTL and controls. To investigate how WMH would affect FA within SCD group, we compared FA values of the 11 WM tracts of a subgroup of SCD patients with WMHs and without WMHs to healthy controls, respectively. Finally, for the SCD group FA values were correlated for the 11 tracts with the nine study measures of neurocognitive performance, using the Benjamini-Hochberg procedure 229 , to decrease the false discovery rate. Results Basic clinical characteristics of participants Participants’ demographic information, WMHs, and neurocognitive test performance are summarized in Table 1. SCD participants (N=26) were African American or white Hispanic, ages 24.2 ± 9.7 years old with a balanced number of males and females (F=13, M=13). Four of the 26 SCD patients were on chronic transfusion therapy for this indication. The control group did not demonstrate significant differences in age and sex with participants with SCD. Twelve control participants were first or second-degree relatives of the SCD patients. There was a non-significant gender imbalance (p>0.05) compared to SCD group. To control for possible changes related to anemia, independently of sickle hemoglobin, we also recruited 19 participants with non-sickle chronic anemia syndromes (ACTL), including beta-thalassemia major (N=11), beta-thalassemia intermedia (N=7) and congenital dyserythropoietic anemia (N=1). ACTL age (26.1 ± 11.7) and sex (F=10, M=9) distributions were comparable to those of the SCD patients. Fourteen of the 19 ACTL patients were receiving chronic transfusion therapy. Due to relatively more various types of anemia, it was not possible to ethnically match the ACTL and SCD populations. Tract specific analysis We observed lower FA in most of the 11 tracts tested in SCD patients compared with healthy controls (Figure 3.2.A). In each group comparison, red indicates lower FA in SCD in comparison to controls and blue indicates lower FA values in controls; none of the blue regions approached significance. 67 Group 1 Group 2 Demographics CTL (N=21) SCD (N=26) p- Value CTL (N=21) ACTL (N=19) p-Value Age (mean ± SD) 22.6±8.9 24.2±9.7 0.5 22.6±8.9 26.1±11.7 0.4 Female 13 (62%) 13 (50%) 0.7 13 (62%) 8 (42%) No. transfused (%) 0 8 (31%) - 0 13 (68%) - No. participants with WMHs 4 14 - 4 7 - Distribution of WMHs in most affected regions (L/R) Frontal (47/39) Parietal (8/10) Temporal (2/2) Percentage (L/R) 29.5% / 35.6% 6.0% / 7.5% 1.5% / 1.5% Neuropsychologic al tests† Group 1 Group 2‡ Digit Span 9.4 (4-18) 9.4 (3-15) 0.9 9.4 (4-18) 9.6 (7-12) - Coding* 9.1 (1-14) 7.2 (1-13) 0.03 9.1 (1-14) 9.7 (5-13) - Symbol Search* 10.0 (6-15) 8.6 (4-12) 0.03 10.0 (6- 15) 9.8 (7-13) - Inhibition 10.4 (8-13) 8.4(1-12) 0.05 10.4 (8- 13) 10.0 (1- 13) - Inhibition/Switchin g 10.1 (6-13) 8.9 (5-13) 0.08 10.1 (6- 13) 10.1 (4- 16) - Letter Fluency 11.0 (6-17) 8.9 (1-15) 0.06 11.0 (6- 17) 9.8 (5-16) - Category Fluency 9.6 (6-13) 8.2 (1-16) 0.2 9.6 (6-13) 9.9 (6-14) - Switching 9.9 (5-16) 8.7 (2-13) 0.3 9.9 (5-16) 10.3 (4- 14) - Trail Making Test 9.4 (4-18) 9.1 (3-15) 0.8 9.4 (4-18) 9.5 (5-12) - Table 3.1. Demographical characteristics and neuropsychological performance results. Abbreviations: CTL: controls, SCD: sickle cell disease, ACTL: anemic controls, WMH: white matter hyperintensity, L/R= left hemisphere/right hemisphere. FSIQ: full-scaled IQ.†Psychometric scores are presented as scaled scores, which have a mean of 10 and a standard deviation of 3, followed by the range. ‡In group 2, neuropsychological tests are only available for 9 patients in ACTL group, therefore, they are not analyzed. *Significant difference between two groups using t-test with Bonferroni correction. 68 The most pronounced differences in FA between SCD patients and healthy controls are found along the CC. The largest cluster along the CC is found on the genu on the left hemisphere. Noticeable areas of difference observed bilaterally along the CST are mostly located inferiorly. The IFO shows larger and more regions of significant difference on the left hemisphere in comparison to the right. Significant FA differences were observed on the lateral portion of the right ILF, SLF and left UNC while no significant FA differences were discernable on either the left ILF or the left SLF between SCD patients and controls. All the significant FA different areas are shown in red clusters in Figure 3.2A. To determine the impact of anemia alone on WM integrity, Figure 3.2.B shows FA comparison of 11 WM tracts between the ACTL and CTL groups. The clusters on the right CST and right ILF in Figure 3.2.A disappear in Figure 3.2.B. The significant areas on left genu of the CC are also seen in ACTL with CTL, as well as that on the left IFO. Figure 3.3: The significant clusters of reduced FA in SCD patients with white matter hyperintensities (panel A) and patients without hyperintensities (panel B) compared to healthy controls (in red), overlaid on the corresponding t-statistics maps on the skeleton surfaces of 11 tracts. Figure 3.3: The significant clusters of reduced FA in patients with SCD (panel A) and non-sickle anemic syndromes (panel B) compared to healthy controls (in red), overlaid on the corresponding t-statistics maps on the skeleton surfaces of 11 tracts. 69 From the T2-FLAIR images, white matter hyperintensities (WMHs) were observed in 14 out of 26 subjects in the SCD group, with accumulated WMHs of 132 distributed mostly in frontal and parietal lobes (Table 3.1). Figure 3.3 demonstrates FA differences in the subgroup of SCD patients with WMHs (Figure 3.3.A) and without WMHs (Figure 3.3.B) compared to CTL. The patterns are similar in both comparisons for ILF, SLF, UNC and IFO. However, when we directly compared the patients with WMHs against the ones without WMHs, there were no significant clusters on the inferior portion of CST or the left genu of CC. The only significant clusters were observed on the anterior potion of the left SLF, of which the location and size are consistent with the ones observed in Figure 3.3.A, therefore, the figure was not included. In addition, we compared mean FA values of each tract in each subject between the SCD, ACTL and CTL groups, after masking the FA maps with the tracts in the atlas space. Mean FA values of the CC and right CST are significantly lower in SCD than the ones in controls. In comparison, only mean FA in the CC (p=0.007) and the right CST (0.042) showed a significant difference between in the repeated one-way ANOVA test. Neurocognitive data analysis and correlation with neuroimaging Given the prevalence of FA abnormalities observed in Figure 3.2, we compared neurocognitive performance between the SCD and CTL subjects. Neurocognitive testing was only completed in 6 of the 19 ACTL subjects, who additionally were not comparable in terms of race or parents’ education and income level to either the CTL or SCD groups. Therefore, the ACTL group was not included in neurocognitive analysis. While the SCD group had lower scores in all but one metric, the only statistically significant group differences were found on Symbol Search (p=0.027) and Coding (p=0.030), both measures of processing speed and Inhibition (p=0.053), a measure of response inhibition. A non-significant trend was suggested for Letter Fluency, a lexical verbal fluency task (p=0.061). No significant group differences were found for Category Fluency (semantic verbal fluency), Digit Span (working memory), Color-Word Interference (response inhibition), and Trail-Making (cognitive flexibility). To examine the potential relationship between neurocognitive performance and WM microstructural alterations in the SCD group, we performed correlations for each neurocognitive test by mean FA values in each of the 11 tracts. Only the FA changes in CC were associated with the neurocognitive data, and significant correlations were observed across all nine neurocognitive measures (despite correction for multiple comparisons): Processing Speed: Symbol Search 70 (p=0.012, r=0.30) and Coding (p=0.006, r=0.34); Working Memory: Digit Span (p=0.011, r=0.27); Response Inhibition: Inhibition (p=0.007, r=0.51); Cognitive Flexibility: Inhibition/Switch (p=0.007, r=0.57), Trail Making (p=0.011, r=0.29), and Verbal Fluency: Category Switch (p=0.011, r=0.25), and lexical (Letter, p=0.037, r=0.23) and semantic (Category, p=0.015, r=0.24) verbal fluency. Discussion We examined the differences in 11 major WM tracts between patients with SCD, non- sickle anemic controls, and healthy controls, using the most widely used DTI metric: FA. FA has been classically seen as a measure of white matter integrity and cohesiveness, and it increases over time with axonal growth and myelination 230–232 . The results were calculated using a medial representation of tracts, and a deformable shape analysis technique, TSA, which directly projects areas of significant differences between WM tracts onto surfaces 63 . Additionally, we analyzed the neurocognitive performance of the SCD patients and healthy controls and further examined the relationship between FA values and neurocognitive performance in SCD. To investigate the impact of anemia alone on WM microscopic changes, we also compared FA differences between ACTL and healthy controls. SCD patients demonstrated widespread lower WM FA on each of 11 tracts except the SLF and UNC. Many of the changes observed in the SCD group were also observed in the ACTL group, but in smaller size and with a sparser distribution, especially in the CC, and the right and left CST. The corpus callosum (CC), the largest and most prominent WM structure in the brain, connects the left and right cerebral hemispheres and consists of more than 200 million axonal projections to the various parts of the cerebral cortex. Our observed decrease in FA parallels the structural alterations of the CC, including shape deformation and reduced FA specific to genu 50,61 . It has been demonstrated that WM has impaired oxygen transport in patients with either sickle or non-sickle anemic syndromes 108 . This chronic hypoxic condition may irreversibly reduce the extent of myelination in the CC, as suggested in a study in mice 233 . However, the effect of chronic hypoxia on myelination of the CC, or more generally, longitudinal WM alterations due to anemia or hypoxia, have not yet been studied in humans. Nevertheless, the reduced FA in the genu, possibly reflecting synaptic pruning and dysmyelination at the microscopic neuronal level, could be one of the consequences of chronic hypoxia on WM. 71 Changes in the inferior right CST appear unique to SCD patients in our overall sample and may be related to the pathophysiology mechanism of sickled hemoglobin. The CST interconnects the motor cortex with the brainstem and spinal cord, thus playing a vital role in controlling muscular movements in the body. Our results for the CST are consistent with the results in 61 , who observed fewer fiber counts on both left and right CST using ROI-based analysis. The clusters on the inferior portion imply degenerated corticospinal pathways, which would suggest an impact on motor functioning. While motor functioning was not assessed in our study, obvious motor deficits were not observed in our participants, consistent with previous reports in patients with SCD with no overt neurological symptoms 52,61 . On the other hand, changes in the IFO were found in both SCD and ACTL patients. The IFO is a prominent WM tract that connects the frontal lobe with the occipital cortex and temporo- basal areas 234 . Due to its connectivity, the functional influences are: semantic elaboration of language for the superficial layer and the posterior region the deep layer; integration of multimodal sensory inputs and motor planning functions for the middle region; emotional and behavioral impacts for the anterior region 235 . The left middle portion of IFO in Figure 3.2.A and B, both show significant difference, which indicates that these areas are affected by anemia in general, regardless of phenotype. Consistent with Van der Land et. al. 94 , we found the presence or absence of WMHs was a marker of disease severity in the SCD group, but this did not fundamentally alter the distribution of significant areas of FA abnormality: reduced FA present on Figure 3.2.A were also present in patients with WMHs (Figure 3.3B). The significant areas of CC and SLF in Figure 3.3.B are more widely distributed than Figure 3.3.A, however, we did not observe a significant difference when directly comparing patients with WMH against patients without WMH within SCD group. On the contrary, the patterns in the CST, left ILF and IFO are reversed: significant differences are less discernable in Figure 3.3.B than those of Figure 3.2.A. Given that the WMHs were distributed mostly in frontal (65%) and parietal lobe (14%), the FA changes were found in varying locations across different patients, which suggests that they are biomarkers of more diffuse WM disease, independent on the location of the WMH. That is, WMH represent an iceberg phenomenon, having microscopic structural damage far exceeding the areas exhibiting WMH on T2-FLAIR images. The significant lower FA regions on the genu of the CC and the anterior region of IFO, ILF and left SLF in both Figure 3.2.A and 3.3 may indicate the vulnerability of the WM in the frontal and 72 parietal lobes, which are the regions with the lowest cerebral blood flow 108,113 . Our findings are consistent with Ford et. al, who reported a large SCI density in the pediatric patients with SCIs, in the frontal lobes (90%), followed by the parietal lobes (53%) in a multi-center pediatric SCD study 113 . Regarding neurocognitive test findings in SCD patients, we found performance was significantly lower on measures of processing speed in comparison to healthy controls, in line with previous reports in adults with SCD who have no known history of overt stroke 52,209,236–238 , Processing speed is a basic neurocognitive process sub-serving other cognitive functions 239 . In its broadest sense, it is defined as the speed at which one can perform mental operations, though many measures also involve a fine motor component. Processing speed is a critical component in the acquisition of new learning, as well as in the efficient retrieval and integration of previous learning 240 . In addition to its impact on cognitive domains such as learning and retrieval, processing speed tests themselves are often a component of Full Scale Intelligence Quotients (FSIQ). In addition to verbal and nonverbal reasoning, FSIQ may include working memory and processing speed, depending on the measure. This would partially account for the lower FSIQ scores reported in studies of SCD that used intelligence measures that included processing speed 39,207,241–243 . Deficits in processing speed can stem from focal or diffuse injuries to the white matter, as well as from subcortical lesions, for example, in the caudate 244 . Processing speed deficits are commonly seen with axonal injury in traumatic brain injury 245 , following cranial irradiation for cancer treatment 207,246 and in demyelinating disorders such as multiple sclerosis 247–249 . Our finding of lower processing speed in SCD patients with no known history of overt stroke is consistent with microscopic structural damage to WM tracts. Additionally, our SCD group’s neurocognitive performance was also significantly lower, in comparison to healthy controls, on a measure of inhibition. Inhibition is an executive function involving resisting a prepotent, overlearned, or impulsive response. Inhibition and other psychometric measures of executive functions are timed and therefore performance is mediated by processing speed. Deficits in inhibition have been associated with WM integrity 250,251 and SCD 252,253 . More broadly, deficits in executive functions, including working memory have been reported in SCD 39,54,236 . In our examination, we found that lower FA values along the CC was significantly associated to lower neurocognitive performance in our SCD group. Significant correlations were not indicated for the other tracts, likely due to less spatial sensitivity to capture potential 73 relationships 254 . Given the CC is by far the largest commissural tract responsible for communication between the left and right hemispheres, the integrity of its axons would play a critical role in the speed of information transfer. Regarding our significant findings between FA values and measures of executive functions, the genu of the CC has fibers which connect the prefrontal cortices of the two hemispheres while the orbital-frontal cortices are connected by fibers from the rostrum of the CC 255 . The relationship between the prefrontal cortex and executive functions has been examined since 1848, when Dr. Harlow began treating and studying Phineas Gage after an industrial accident sent an iron rod through Gage’s left frontal lobe, leading to significant change in his personality and behavior 256 . Of note is the orbital-frontal aspect of Gage’s injury and his subsequent inability to inhibit behavior 256 . More recent studies 257–259 have elucidated the role of the prefrontal cortex as the primary coordinating region for executive functions, with connections to diverse brain regions involved in specific executive functions. The role of the CC in transferring information between the hemispheres, as well as between the prefrontal cortices, is consistent with our significant findings supporting the role of CC integrity and functional outcomes in processing speed and executive functions in SCD. Our study has several limitations. Neuronal connections of WM extend beyond the continuous medial representation of the sheet-like 11 tracts, if observed using tractography 260 . TSA is not able to reveal the full WM connections, especially in the distal areas of deep WM structures, which explains the very small clusters shown on the edge of right SLF, right CC in Figure 3.2A.and left UNC in Figure 3.2B. In addition, the diffusion tensor model itself is less capable of evaluating FA values in regions with crossing fibers, thereby potentially underestimating FA values in the regions where fiber kissing, curving and branching are pronounced. Hence, the next stage of our work will be to employ alternative modeling methods with the current dataset. It would also be beneficial to rescan our study participants with multi-shell acquisition methods, in order to obtain a finer understanding of the different microstructural differences. Lastly, our SCD patients are heterogeneous with respect to genotype and treatment (e.g, blood transfusion status, hydroxyurea), as are our ACTL patients. While we were able to obtain a broad range of hemoglobin values (by study design), we did not have the statistical power to characterize the impact of transfusions, hydroxyurea, or genotype, nor the power to compare the neurocognitive performance differences between SCD patients with WMH and without 209 . Despite this, microstructural differences 74 between our anemic and non-anemic participants were observed and in line with previous work, highlighting the sensitivity of the TSA methods. In conclusion, our study is the first to investigate the impact of anemia and sickled hemoglobin separately on the microstructural changes in the 11 major WM tracts. Decreased FA in the genu of CC, left inferior CST and IFO was observed in chronically anemic patients, regardless of anemia subtype, while FA reductions in the CST and SLF were unique to SCD patients. Patients with WMHs had more significant FA abnormalities, which were found in the same areas as patients without WMHs. Slower processing speed and response inhibition skills were observed in SCD patients, consistent with WM involvement. Decreased FA values in the CC significantly correlated with all nine neurocognitive measures, which included processing speed, working memory, and executive functions, suggesting a critical importance for CC fiber integrity in core neurocognitive processes. Future work on this study will include examining the thickness and the integrity of WM interconnections, to further understand WM damage and neurocognitive functioning in SCD patients. 75 Chapter 4: Silent cerebral infarcts stroke map in pediatric cohort: its implications and research significance Introduction Three-dimensional (3D) MRI acquisitions offer outstanding signal-to-noise and resolution, however they are time-consuming and subject to motion artifacts 261 . Hence many clinical institutions (particular pediatric hospitals) use multi-slice two-dimensional (2D) acquisitions, instead, which are poorly suited for multiplayer reformatting and atlas registration because of voxel asymmetry. The problem, as shown in Figure 4.1 (a), is that given a stack of 2D slices with wide slice thickness, through-plane slices are to be retrieved to produce isotropic resolutions in the orthogonal views, and super-resolution (SR) techniques have been developed to tackle this issue. Conventional interpolation methods are simple and fast, but generally cause artefactual noise, due to relying on neighboring voxel intensities. Recently, deep convolution neural networks (CNN) have shown explosive popularity and powerful capability to improve high resolution (HR) results. Figure 4.1: In 2D slice stacks, through-plane (axial) slices have large thickness, making the resolution much lower than in-plane resolution. Non-isotropic SR is needed to retrieve the missing slices. (b); The slices to be restored are modeled as 1-valued masks, when viewing in other two planes (coronal and sagittal), Image in-painting could reproduce fine details, especially for the complex gyrus and fissures. The edge generator hallucinates edges in the missing areas, supposing that edge recovery is easier than image completion. Then the image in-paint network combines the edges in the missing regions with texture information of the rest of the image to fill the missing areas. 76 SR-CNNs can generate satisfying HR results 165,166 . However, the widely used optimization methods of CNNs are voxel-wise error between estimated and the ground truth images, such as mean squared error (MSE), or signal-to-noise (SNR), which lead to overall blurring and low visual quality. More recently, generative adversarial networks (GAN) 163 provided a framework for generating plausible-looking images in SR problems, referred to as SRGAN 262 . While SRGAN can restore images with more realistic characteristics, it does not guarantee voxel- wise performance, because of its emphasis on learning patterns 170,263 . In this work, we address the challenge of restoring high-resolution, isotropic 3D images from 2D MRI slices. Instead of learning the inverse mapping directly from low resolution (LR) images to HR images using neural networks 169 . we achieve the goal by edge recovering and contrast completion. The detailed design and experiments are articulated in the following sections. Methodology The SR problem can be mathematically represented as 𝒀=𝑓 𝑿 , where Y is the observed data (LR images), X is the unknown HR images, and 𝑓 is the image degrading mapping. The aim of SR task is to find 𝑓 eu ∙ and minimize the difference between estimated result 𝑿 and HR images X. In this section, we proposed a tailored data representation for non-isotropic clinical MR image super resolution, and employed an edge-guided GAN (EG-GAN) which embedded dual generators and discriminators to progressively solve SR problem in two steps: edge connection and contrast completion. Data Representation of LR images and Proposed Framework In this work, we analogize the missing through-plane slices as 1-valued masks, which present as masked rows in the other two planes, as illustrated in Figure 4.1(b); retrieving missing slices is achieved by estimating the masked rows in the other two orthogonal planes. The first step is to connect the broken edges in the masked rows; then the contrast completion network combines the bridged edges and LR images to estimate the voxel intensities in the missing rows. This approach is achieved by edge-guided GAN (EG-GAN) to enhance perceptually consistent results. As shown in Figure 4.2, EG-GAN firstly connects the missing edges of the LR images by edge generator, taking LR images, and masks generated from missing slices in through-plane as input, supervised by the edges generated from HR images. As input and ground truth, edges of LR images and HR images are extracted by Canny edge detector. Then contrast generator fills the intensities based on the original contrast from LR images, guided by edges generated from the first 77 step and supervised by HR images. Both steps follow an adversarial model, each consists of a generator and a discriminator pair. Our architecture follows 264 , which has achieved impressive results in image restoration for natural images. In our networks, edge connection and contrast completion are optimized by two objective functions, respectively. The objective function of edge connection is designed as: min ¹ Ã max Ã 𝑙𝑜𝑠𝑠𝐺 Å = min ¹ Ã λ ¹ÇÈu max Ã 𝑙𝑜𝑠𝑠 ¹ÇÈu + λ É 𝑙𝑜𝑠𝑠 É (22) where G E , D E are the edge generator and paired discriminator, and 𝑙𝑜𝑠𝑠 ¹ÇÈu is the adversarial loss. To stabilize the network, feature-matching loss 𝑙𝑜𝑠𝑠 É is added, and 𝑙𝑜𝑠𝑠 ¹ÇÈu and 𝑙𝑜𝑠𝑠 É are regularization parameters. The edge mapping performed by generator can be represented as: 𝑪 ¯´ = 𝐺 Å 𝑰 Ì ,𝑪 Ì ,𝑴 (23) where 𝑪 ¯´ is the predicted image contour, 𝑰 Ì and 𝑪 Ì are the LR input and corresponding image contour. 𝑴 is the mask indicating the missing slice. Therefore, the adversarial loss can be written as: 𝑙𝑜𝑠𝑠 ¹ÇÈu =𝔼 𝑪 ¡ ,𝑰 ¡ log𝐷 Å 𝑪 𝒈𝒕 ,𝑰 𝒈𝒕 +𝔼 𝑰 𝒈𝒕 log 1−𝐷 Å 𝑪 𝒑𝒓𝒆𝒅 ,𝑰 𝒈𝒕 , (24) Figure 4.2: An overview of the proposed method. Edge generator (in green) firstly connects the missing edges from the LR images. Then contrast generator fills the intensities based on the original contrast from LR images, guided by edges generated from the first step and supervised by HR images. Both steps follow an adversarial model. 78 where 𝑰 q and 𝑪 q are the HR ground truth images and their contour. The feature-matching loss loss 𝑙𝑜𝑠𝑠 É calculates the difference of the activation maps generated by the hidden layers in the discriminator, is defined as: 𝑙𝑜𝑠𝑠 É =𝔼 u È ¶·u 𝐷 Å ¶ 𝑪 𝒈𝒕 −𝐷 Å ¶ 𝑪 𝒑𝒓𝒆𝒅 u , (25) where 𝐿 is the number of hidden layers in the discriminator, and 𝑁 ¶ is the number of elements in the map of the 𝑖 q× layer. 𝐷 ¶ represents the 𝑖 q× activation. Similarly, the contrast mapping 𝐺 Ø can be represented as: 𝑰 ¯´ = 𝐺 Å 𝑰 Ì ,𝑪 Ì (26) where 𝑰 ¯´ is the prediction of HR image. Therefore, the adversarial loss is: 𝑙𝑜𝑠𝑠 ¹ÇÈB =𝔼 𝑰 ¡ ,𝑪 ÙÚÛ½ log𝐷 𝑰 𝒈𝒕 ,𝑪 𝒑𝒓𝒆𝒅 +𝔼 𝑪 𝒑𝒓𝒆𝒅 log 1−𝐷 𝑪 𝒑𝒓𝒆𝒅 ,𝑰 𝒑𝒓𝒆𝒅 , (27) where 𝐷 is the corresponding discriminator. To ensure the reconstructed image has both high voxel-wised accuracy and good perceptual quality, we utilized another three loss: 𝑙 u loss, perceptual loss, 𝑙𝑜𝑠𝑠 ¯ , and style loss 𝑙𝑜𝑠𝑠 Ü . Therefore, the overall loss is: 𝑙𝑜𝑠𝑠 ¹ Ý = λ 𝑙 1 𝑙𝑜𝑠𝑠 𝑙 1 + λ 𝐺𝐴𝑁 2 𝑙𝑜𝑠𝑠 ¹ÇÈB + λ 𝑝 𝑙𝑜𝑠𝑠 ¯ + λ 𝑠 𝑙𝑜𝑠𝑠 Ü . (28) Proposed Edge-guided Adversarial Network To further stabilize networks, spectral normalization is applied in both generator and discriminator. Note that edge generator has spectral normalization (SN) and instance normalization across all the layers, while contrast generator only uses instance normalization, because learning high frequency information such as edges requires more restrictions to maintain the stability of network. However, for low frequency contrast information, SN is not necessary and might slow the training procedure. Therefore, the SN is removed from contrast generator. The discriminator for edge connection is the same with the one used in contrast completion. Experiments and Evaluation Data Preparation To demonstrate the generalization of EG-GAN, we used a large publicly available T1- weighted brain images in the Human Connectome Project (HCP) S1200 datasets. We downloaded images after preprocessing pipelines, including distortion correction and brain extraction. We randomly chose 600 subjects in our whole experiments. The images come in 0.7 mm isotropic high resolution, and we removed boundary all-zero slices and fit them to 256 x 256 x 256. The whole 79 dataset is split into 50% of training, 25% of validation and 25% of testing, without subjects overlapping. The predicted results from testing set only were used in our final performance evaluation and comparison. The original images were used as ground truth HR images, and artificially generated thick-slice images, or LR images, were generated based on their paired-HR images. While only one slice in every three slices in the LR images were from HR images, the two other slices will be filled with 1-valued masks (illustrated in Figure 4.1b), therefore, the LR images have the same size as reference images (256 x 256 x 256). Training Procedure The models were implemented in PyTorch on High-Performance Computing clusters with Nvida Tesla P100 GPUs. To keep and continuously improve the generalization of the proposed method, we used parameters kindly provided by Nazeri, K et. al, pre-trained on three publicly available natural datasets, to initialize our model with batch size of 8. Learning rate was set to 10- 4 for generators and 10-4 for discriminators. The model was optimized using ADAM optimizer. We set the maximum training iteration as 20,000. Results To evaluate the similarity between reference HR images and our results, we compared the voxel-wise intensity value accuracy, peak signal-to-noise ration (PSNR). Furthermore, to measure perceptual quality of resultant images, we also evaluated the results with structural similarity index (SSIM). Figure 4.3 compares our results against other five methods: low-rank total variation (LRTV), nearest neighbor (NN), bicubic, SR-CNN and SR-GAN. As shown in the second row, the magnified region exhibits the degree of blurring and fine details of the brain cortex. Figure 4.3: An axial slice of: the ground truth, reconstructed by nearest neighbor (NN) interpolation, bicubic interpolation, low-rank total variation (LRTV), super-resolution CNN (SR-CNN), super-resolution GAN (SR- GAN) and our proposed method (EGGAN) resulting image. 80 Visually, the most comparable method to the proposed method is SR-GAN, also showing only slightly lower SSIM than EG-GAN. Conclusion and Discussion Our work modeled non-isotropic MR SR problem as edge guidance and contrast completion. Both steps were implemented using GAN, and both networks incorporate losses based on deep features to enhance perceptually consistent results. Our results showed great improvements on both PSNR and SSIM over interpolation-based and metric completion-based methods. When we tested on the CNN-based method, PSNR is more similar to the proposed method, however, SSIM is 14% lower than EG-GAN. It is as expected because of "checkerboards" artifacts due to deconvolution. Compared to the GAN-based method, SSIM are identical yet PSNR is much lower than that in EG-GAN. Partially, it is because GAN-based methods are generally more capable of generating images with high SSIM, by sacrificing PSNR. However, in EG-GAN, we added more constrains to enhance the consistency of the network to achieve higher intensity accuracy. Therefore, our proposed method has superior PSNR, conspicuity, and qualitative “look- and-feel” compared with the SR-GAN method, which is important for human interpretation of the data. Further work is necessary to determine whether EG-GAN is truly diagnostically superior to SR-GAN with respect to lesion detection, volumetric, and other metrics reflecting clinical utility. 81 Chapter 5: Conclusion and ongoing work In this thesis, we demonstrated that WM O 2 delivery is impaired in SCD patients using arterial spin labeling MRI. Cerebral compensatory mechanisms appear optimized to protect GM O 2 delivery, even while WM remains hypoxic. WMH occur in watershed areas, where CBF and O 2 delivery are intrinsically low, but O 2 delivery in these regions is even lower than uniquely predicted by watershed effects. Taken together, compensatory hyperemia preserves O 2 delivery to the GM in SCD patients but is insufficient to maintain O 2 delivery to the WM, explaining the distribution and progressive evolution of SCIs in SCD patients. However, this work did not take account regional arterial transit time heterogeneity when quantifying CBF or O 2 delivery. Arterial transit time could even differ across the brain in healthy subjects, being longer in distal branches, especially in the watershed area 105 . In this work, PCASL sequence was designed with one single inversion time, and therefore transit time is not possible to be retrieved. Time-encoding ASL, or TE-ASL, utilizes scans with time encoding blocks to record transit time information without the penalty of decreasing SNR 265 . Assessment of regional arterial transit time enables us to improve the accuracy of CBF quantification and confirm the hypoperfusion that we observed in this thesis, by eliminating the inaccuracies induced by arterial transit time 266 . More importantly, regional arterial transit time tailored to each flow territory could be highly informative to identify vascular occlusions and hemodynamic disturbances, especially in patients with chronic anemia 129 . Other than ASL, Gadolinium-based dynamic susceptibility contrast (gDSC) is another commonly used MR technique to measure brain perfusion, but its invasiveness and long-term contrast agent accumulation have significantly limited its repeatability 267 . Alternatively, hypoxia- induced deoxyhemoglobin contrast generated by transient inhalation of a hypoxic gas can be used to estimate brain perfusion map, named as deoxyhemoglobin DSC (dDSC), following the same principle of gDSC. The changes in the MRI signal can be further used to quantify cerebral blood volume and capillary transit time. In addition, the hypoxic contrast of dDSC is safe and tolerable, therefore it allows repeated experiments in pediatric cohorts and renal-impaired subjects 268 . While the future work mentioned above focus on estimating brain perfusion metrics, exploring the adequacy of oxygenated blood transported to the brain tissue, the other part of the future work goes to the measurement of venous oxygenation, a vital information of oxygen supply- demand balance, which is abnormal in most cerebral vascular disease 102 . Asymmetric spin Echo 82 (ASE) provides the contrast depending on the amount of deoxyhemoglobin in the veins and its volume fraction 269 . Based on a two-compartment model (extravascular brain tissue and intravascular venous blood vessel) proposed by Yablonskiy and Haacke 270 , venous cerebral blood volume (vCBV) and oxygen extraction fraction can be estimated. Studies using TRUST and ASE has demonstrated increased whole brain OEF in patients with SCD, however, many confounders were not considered and the physiological model was not perfect 271 . Future work is required to explore accurate estimation of OEF to further detect chronic hypoxia in the brain. One of the consequences of impaired hemodynamic balance in SCD is microstructural damage in the WM. In this thesis, we investigated the impact of anemia and sickled hemoglobin separately on the microstructural changes in the 11 major WM tracts. Decreased FA in the genu of CC, left inferior CST and IFO was observed in chronically anemic patients, regardless of anemia subtype, while FA reductions in the CST and SLF were unique to SCD patients. Patients with WMHs had more significant FA abnormalities, which were found in the same areas as patients without WMHs. Slower processing speed and response inhibition skills were observed in SCD patients, consistent with WM involvement. Decreased FA values in the CC significantly correlated with all nine neurocognitive measures, which included processing speed, working memory, and executive functions, suggesting a critical importance for CC fiber integrity in core neurocognitive processes. Unlike DTI metrics, high-gradient multi-shell DWI images could resolve tissue mixture problems using multi-shell-multi-tissue in fiber orientation distribution (FOD) estimation, or constrained spherical deconvolution (CSD) estimation 272 . Multi-shell DWI scans are available for an ongoing project, and FOD or CSD metrics will better handle crossing fibers and dispersion. As number of patients recruited increases over time, this work would provide more sensitive and accurate WM microstructural damage assessments in patients with SCD 273 . The exact mechanism of microstructural WM injury in SCD is not completely understood. The possible explanations are either chronic hypoxia condition, or a leaky blood-brain barrier (BBB), or a combination of both. It is of the vital importance of estimating the blood that pass through the microvasculature to enter the venous vessels. This could be measured using water extraction with phase contrast arterial spin labeling 274 in the future. As no studies have probed the relation of microvascular injury and BBB in patients with SCD, future work on assessing BBB 83 could shed light on better understanding WM microstructural injury and potentially the progression of silent infarction 275 . In order to create a silent infarction “heat map” in the common space, we utilized convolutional neural network (CNN) and achieved improved lesion segmentation performance by combining three CNNs with different patch size in parallel. To register low and anisotropic resolution images to high and isotropic template, we generated high resolution T2 MR scans using GAN. However, the proposed method is not yet able to produce high-resolution in other two orthogonal planes with anatomical meaningful information. There are ongoing efforts to exploit orthogonal low-resolution images to improve the super-resolution, such as maximum a-posteriori (MAP) and Gaussian Mixture Models 276,277 . To make full use of information provided by two orthogonal scans, iterative optimization between the two reconstructed images using the method in Chapter 4 is one option. The optimization could be performed either in image space or in k- space. In this thesis, our super-resolution reconstruction method can be extended from more accurate registration and lesion segmentation to more reliable brain extraction and brain tissue parcellation 278 , especially for clinical anisotropic low resolution scans. On the other hand, the promising performance of the edge-guided image reconstruction method relies on the degree of edge recovering, a more cohort-specific edge detector that is tailored to the dataset, such as traumatic brain injuries or tumors, will greatly improve the overall generalization by allowing the model to learn meaningful covariation. Therefore, a wide range of clinical datasets contain the anatomical consistency is required for further experiments and validations. As for the ultimate goal, generating a more reliable lesion map, prior knowledge from T1 images could be used to enhance the performance of the lesion classification networks. On the one hand, T1-images could provide a WM binary mask to spatially restrain the false positives outside of the WM. On the other hand, a silent infarct shows hyperintensities on a T2 images could turn out to be darker on T1 scans, which could be used to differentiate the candidate lesion voxels from the ones neutral on T1 images. In conclusion, measurement of cerebral oxygen delivery and microstructural injury is necessary and imperative to identify an imaging biomarker of stroke, to unveil the onset, progression, and development of brain infarction, and to better understand how cerebral vascular disease would physiologically and hematologically impact the brain. The next breakthrough could 84 lay on the evolving of MR sequence, optimizing the quantification models, as well as advancing the artificial neural networks. 85 Reference: 1. Studies S, Polymerization DS, Hemoglobins O, Mechanism D. Sickle cell hemoglobin polymerization. 2. Lervolino LG, Eduardo P, Baldin A, et al. Prevalence of sickle cell disease and sickle cell trait in national neonatal screening studies. 2011;33(1). doi:10.5581/1516-8484.20110015. 3. Hassell KL. Population estimates of sickle cell disease in the U.S. Am J Prev Med. 2010;38(4 Suppl):S512-21. doi:10.1016/j.amepre.2009.12.022. 4. Creary S, Badawy SM, Cronin RM, et al. Patient-Centered eHealth Interventions for Children , Adolescents , and Adults With Sickle Cell Disease : Systematic Review. 2018;20(7). doi:10.2196/10940. 5. Rees DC, Williams TN, Gladwin MT. Sickle-cell disease. Lancet. 2010;376(9757):2018- 2031. doi:10.1016/S0140-6736(10)61029-X. 6. Reiter CD, Wang X, Tanus-Santos JE, et al. Cell-free hemoglobin limits nitric oxide bioavailability in sickle-cell disease. Nat Med. 2002;8(12):1383-1389. doi:10.1038/nm1202-799. 7. Hospital G, Progam B, Cross AR. Irreversibly Sickled Cells and Red Cell Survival in Sickle Cell Anemia A Study with Both DF3 * P and 51CR. 1978;64(February):253-258. 8. Milnm PF, Charaghe S. Life Span of Carbamylated Red Cells in Sickle Cell Anemia. 1973;52(December):3161-3171. 9. Rees DC. Sickle Cell Disease. 2017:1561-1573. doi:10.1056/NEJMra1510865. 10. Hulbert ML, Panepinto JA, Scott JP, et al. Red blood cell transfusions during sickle cell anemia vaso-occlusive crises: a report from the magnesium in crisis (MAGiC) study. Transfusion. 2017;00(April):1-7. doi:10.1111/trf.14155. 11. Adams RJ, Ohene-Frempong K, Wang W. Sickle cell and the brain. Hematology Am Soc Hematol Educ Program. 2001:31-46. doi:10.1182/asheducation-2001.1.31. 12. Strouse JJ, Jordan LC, Lanzkron S, Casella JF. The excess burden of stroke in hospitalized adults with sickle cell disease. Am J Hematol. 2009;84(9):548-552. doi:10.1002/ajh.21476. 13. Rees DC. How I manage red cell transfusions in patients with sickle cell disease. 2018;(January):607-617. doi:10.1111/bjh.15115. 14. Moser FG, Miller ST, Bello JA, et al. The Spectrum of Brain MR Abnormalities in Sickle- 86 Cell Disease : A Report from the Cooperative Study of Sickle Cell Disease. 15. Strouse JJ, Lanzkron S, Urrutia V, Lanzkron S. Expert Review of Hematology stroke in adults with sickle cell disease The epidemiology , evaluation and treatment of stroke in adults with sickle cell disease. 2014;4086. doi:10.1586/ehm.11.61. 16. Ohene-Frempong K, Weiner SJ, Sleeper L a, et al. Cerebrovascular accidents in sickle cell disease: rates and risk factors. Blood. 1998;91(1):288-294. 17. Scothorn DJ, Price C, Schwartz D, et al. Risk of recurrent stroke in children with sickle cell disease receiving blood transfusion therapy for at least five years after initial stroke. J Pediatr. 2002;140(3):348-354. doi:10.1067/mpd.2002.122498. 18. Platt, Orah S, Donald J. Brambilla WFR. Mortality in Sickle Cell Disease: Life Expectancy and Risk Factors for Early Death. N Engl J Med. 1994;330(23):1639-1644. 19. Powars D, Wilson B, Imbus C, Pegelow C, Allen J. The natural history of stroke in sickle cell disease. Am J Med. 1978;65(3):461-471. doi:10.1016/0002-9343(78)90772-6. 20. Article O. Silent Cerebral Infarction is Associated with Incident Stroke and TIA Independent of Carotid Intima-Media Thickness. 2009:817-822. doi:10.2169/internalmedicine.49.3211. 21. Debaun MR, Armstrong FD, Mckinstry RC, Ware RE, Vichinsky E, Kirkham FJ. Silent cerebral infarcts : a review on a prevalent and progressive cause of neurological injury in sickle cell anemia Silent cerebral infarcts : a review on a prevalent and progressive cause of neurological injury in sickle cell anemia Department of Pedia. Blood. 2012;119(20):4587-4597. doi:10.1182/blood-2011-02-272682. 22. DeBaun MR, Gordon M, McKinstry RC, et al. Controlled Trial of Transfusions for Silent Cerebral Infarcts in Sickle Cell Anemia. N Engl J Med. 2014;371(8):699-710. doi:10.1056/NEJMoa1401731. 23. Thompson A a, Sarnaik S a, Woods GM, et al. Silent cerebral infarcts occur despite regular blood transfusion therapy after first stroke in children with sickle cell disease. Blood. 2011;117(3):772-780. doi:10.1182/blood-2010-01-261123. 24. Caplan LR. Silent brain infarcts. Cerebrovasc Dis. 1994;4(Suppl. 1):32-39. 25. Switzer JA, Hess DC, Nichols FT, Adams RJ. Pathophysiology and treatment of stroke in sickle-cell disease : present and future. 2006;5(June):501-512. 26. Adams RJ, McKie VC, Hsu L, et al. Prevention of a First Stroke by Transfusions in 87 Children with Sickle Cell Anemia and Abnormal Results on Transcranial Doppler Ultrasonography. N Engl J Med. 1998;339(1):5-11. doi:10.1056/NEJM199807023390102. 27. Estcourt LJ, Fortin PM, Hopewell S, Trivella M, Doree C, Abboud MR. Interventions for preventing silent cerebral infarcts in people with sickle cell disease. Cochrane Database Syst Rev. 2017;2017(5). doi:10.1002/14651858.CD012389.pub2. 28. Debaun MR, Kirkham FJ. Central nervous system complications and management in sickle cell disease. Blood. 2016;127(7):829-838. doi:10.1182/blood-2015-09-618579. 29. Debaun MR. Secondary Prevention of Overt Strokes in Sickle Cell Disease : Therapeutic Strategies and Efficacy. 2011:427-433. 30. Verduzco LA, Nathan DG. Review article Sickle cell disease and stroke. 2019;114(25):5117-5126. doi:10.1182/blood-2009-05-220921. 31. Scott JP, Neahring B, Medical KP, Insti- B, Orleans N. Neurologic events after partial exchange transfusion for priapism in sickle cell disease. 1992. 32. Otrock ZK, Thibodeaux SR, Jr RJ. Vascular access for red blood cell exchange. 2018;58(February):569-579. doi:10.1111/trf.14495. 33. Hema- P, City K, Oli- N, Driscoll C, Wang W, Hurlett A. Prevention of a first stroke by transfusions i nchildren with sickle cell anemia and abnormal results on transcranial doppler. 1998:5-11. 34. Kumar R, Rakesh A, Patel K. Hydroxyurea in Sickle Cell Disease : Drug Review. 2014;30(June):91-96. doi:10.1007/s12288-013-0261-4. 35. Casella JF, King AA, Barton B, et al. Design of the Silent Cerebral Infarct Transfusion (Sit) Trial. Pediatr Hematol Oncol. 2010;27(2):69-89. doi:10.3109/08880010903360367. 36. DeBaun MR, Gordon M, McKinstry RC, et al. Controlled Trial of Transfusions for Silent Cerebral Infarcts in Sickle Cell Anemia. N Engl J Med. 2014;371(8):699-710. doi:10.1056/NEJMoa1401731. 37. Bernaudin F, Verlhac S, Arnaud C, et al. Chronic, acute anemia and eICA stenosis are independent risk factors for silent cerebral infarcts in sickle cell anemia. Blood. 2014;125(10):1653-1662. doi:10.1182/blood-2014-09-599852. 38. Kassim AA, Pruthi S, Day M, et al. Silent cerebral infarcts and cerebral aneurysms are prevalent in adults with sickle cell anemia - Letter to the Editor. Blood. 2016;127(16):2038-2041. doi:10.1136/neurintsurg-2014-011572.8. 88 39. Daniel Armstrong F, Thompson Jr RJ, Winfred Wang MD, et al. Cognitive functioning and brain magnetic resonance imaging in children with sickle cell disease. Pediatrics. 1996;97(6 I):864-870. http://www.scopus.com/inward/record.url?eid=2-s2.0- 15844366370&partnerID=40&md5=79fabc394a8ad669abf88e6c3e155278. 40. Sarri G, Bhor M, Abogunrin S, et al. Systematic literature review and assessment of patient-reported outcome instruments in sickle cell disease. 2018:1-13. 41. Scantlebury N, Mabbott D, Janzen L, et al. White Matter Integrity and Core Cognitive Function in Children Diagnosed With Sickle Cell Disease. J Pediatr Hematol Oncol. 2011;33(3):163-171. doi:10.1097/MPH.0b013e3182036f33. 42. Neurology C. Multicenter prospective study of children with sickle cell disease : ... 2000;(May). 43. Wang W, Enos L, Gallagher D, et al. N europsychologic performance in school-aged children with sickle cell disease : A report from the Cooperative Study of Sickle Cell Disease. 2001:391-397. doi:10.1067/mpd.2001.116935. 44. Prussien K V, Jordan LC, Debaun MR, D BECPH. Cognitive Function in Sickle Cell Disease Across Domains , Cerebral Infarct Status , and the Lifespan : A Meta-Analysis. 2019:1-11. doi:10.1093/jpepsy/jsz031. 45. van der Land V, Hijmans CT, de Ruiter M, et al. Volume of white matter hyperintensities is an independent predictor of intelligence quotient and processing speed in children with sickle cell disease. Br J Haematol. 2015;168(4):553-556. doi:10.1111/bjh.13179. 46. Kral MC, Brown RT, Connelly M, et al. Radiographic Predictors of Neurocognitive Functioning in Pediatric Sickle Cell Disease. J Child Neurol. 2006;21(1):37-44. doi:10.2310/7010.2006.00008. 47. Kirk GR, Haynes MR, Palasis S, et al. Regionally specific cortical thinning in children with sickle cell disease. Cereb Cortex. 2009;19(7):1549-1556. doi:10.1093/cercor/bhn193. 48. Chai Y, Coloigner J, Qu X, et al. Tract specific analysis in patients with sickle cell disease. In: In 11th International Symposium on Medical Information Processing and Analysis. Vol 9681. ; 2015:968108. doi:10.1117/12.2213617. 49. Choi S, Bush AM, Borzage MT, et al. Hemoglobin and mean platelet volume predicts diffuse T1-MRI white matter volume decrease in sickle cell disease patients. NeuroImage Clin. 2017;15(April):239-246. doi:10.1016/j.nicl.2017.04.023. 89 50. Sun B, Brown RC, Hayes L, et al. White matter damage in asymptomatic patients with sickle cell anemia: screening with diffusion tensor imaging. AJNR Am J Neuroradiol. 2012;33(11):2043-2049. doi:10.3174/ajnr.A3135. 51. Pegelow CH, Macklin EA, Moser FG, et al. Longitudinal changes in brain magnetic resonance imaging findings in children with sickle cell disease. 2019;99(8):3014-3019. 52. Vichinsky E, Neumayr LD, Gold JI, et al. Neuropsychological Dysfunction and Neuroimaging Abnormalities in Neurologically Intact Adults With Sickle Cell Anemia. J Am Med Assoc. 2010;303(18):1823-1831. 53. Kirk GR, Haynes MR, Palasis S, et al. Regionally Specific Cortical Thinning in Children with Sickle Cell Disease. 2009;(July):1549-1556. doi:10.1093/cercor/bhn193. 54. Mackin RS, Insel P, Truran D, et al. Neuroimaging abnormalities in adults with sickle cell anemia Associations with cognition. 2014. 55. Chen R, Pawlak MA, Flynn TB, Krejza J, Herskovits EH, Melhem ER. Brain Morphometry and Intelligence Quotient Measurements in Children With Sickle Cell Disease. J Dev Behav Pediatr. 2009;3(6):3. doi:10.1097/DBP.0b013e3181c35ddf. 56. Kawadler JM, Jonathan D, Kirkham FJ, Cox C, Saunders DE, Clark CA. Subcortical and cerebellar volumetric deficits in paediatric sickle cell anaemia. 2013;(July). doi:10.1111/bjh.12496. 57. Sowell ER, Thompson PM, Leonard CM, Welcome SE, Kan E, Toga AW. Longitudinal Mapping of Cortical Thickness and Brain Growth in Normal Children. 2004;24(38):8223- 8231. doi:10.1523/JNEUROSCI.1798-04.2004. 58. Steen RG, Emudianughe T, Hunte M, et al. Brain Volume in Pediatric Patients with Sickle Cell Disease: Evidence of Volumetric Growth Delay? AJNR Am J Neuroradiol. 2005;26(3):455-462. http://www.ajnr.org/cgi/content/abstract/26/3/455. 59. Steen RG, Hankins GM, Xiong X, et al. Prospective Brain Imaging Evaluation of Children with Sickle Cell Trait: Initial Observations. Radiology. 2003;228(1):208-215. doi:10.1148/radiol.2281020600. 60. Chen R, Krejza J, Arkuszewski M, Zimmerman R, Herskovits E, Melhem E. Brain morphometric analysis predicts decline of intelligence quotient in children with sickle cell disease: a preliminary study. Adv Med Sci. 2016;62(1):151-157. doi:10.1016/j.advms.2016.09.002. 90 61. Balci A, Karazincir S, Beyoglu Y, et al. Quantitative brain diffusion-tensor MRI findings in patients with sickle cell disease. Am J Roentgenol. 2012;198(5):1167-1174. doi:10.2214/AJR.11.7404. 62. Kawadler JM, Kirkham FJ, Clayden JD, et al. White Matter Damage Relates to Oxygen Saturation in Children with Sickle Cell Anemia Without Silent Cerebral Infarcts. Stroke. 2015;46(7):1793-1799. doi:10.1161/STROKEAHA.115.008721. 63. Zhang H, Awate SP, Das SR, et al. A tract-specific framework for white matter morphometry combining macroscopic and microscopic tract features. Med Image Anal. 2010;14(5):666-673. doi:10.1016/j.media.2010.05.002. 64. Baldeweg T, Hogan AM, Saunders DE, et al. Detecting white matter injury in sickle cell disease using voxel-based morphometry. Ann Neurol. 2006;59(4):662-672. doi:10.1002/ana.20790. 65. Pittman RN. Regulation of Tissue Oxygenation. Morgan & Claypool Life Sciences; 2011. 66. Guyton A HJ. Textbook of Medical Physiology. 6th ed.; 2006. 67. Pearls SM. A Patient With Sickle Cell Disease and a Low Baseline Sleeping Oxygen Saturation. 2007;3(3):313-315. 68. Akinsheye I, Alsultan A, Solovieff N, et al. Review article Fetal hemoglobin in sickle cell anemia. 2019;118(1):19-28. doi:10.1182/blood-2011-03-325258. 69. Fathallah H, Atweh GF. Induction of Fetal Hemoglobin in the Treatment of Sickle Cell Disease. :58-62. 70. Bhatt R, Labus J, Coloigner J, Coates T, Zeltzer L, Wood J. Systems ( Physiology , Anatomy , Animal Models ). J Pain. 19(3):S2-S3. doi:10.1016/j.jpain.2017.12.017. 71. Supply THE, Oxygen OF, The TO. of Zoophysiology, University of Copenhagen.). (8). 72. Hayashi T, Watabe H, Kudomi N, et al. A Theoretical Model of Oxygen Delivery and Metabolism for Physiologic Interpretation of Quantitative Cerebral Blood Flow and Metabolic Rate of Oxygen. J Cereb Blood Flow Metab. 2003;(July 2014):1314-1323. doi:10.1097/01.WCB.0000090506.76664.00. 73. Zheng Y, Johnston D, Berwick J, Chen D, Billings S, Mayhew J. A three-compartment model of the hemodynamic response and oxygen delivery to brain. Neuroimage. 2005;28(4):925-939. doi:10.1016/j.neuroimage.2005.06.042. 74. Oxygen supply to tissues : The Krogh model and its assumptions. 1982;38. 91 75. Kuschinsky, W. and OBP. capillary circulation in the brain. Cerebrovasc Brain Metab Rev. 1992;4(3):261-286. 76. Roy, Charles Smart and CSS. On the regulation of the blood-supply of the brain. J Physiol. 1899;11(1-2):85-158. 77. Jespersen SN, Østergaard L. The roles of cerebral blood flow, capillary transit time heterogeneity, and oxygen tension in brain oxygenation and metabolism. J Cereb Blood Flow Metab. 2012;32(2):264-277. doi:10.1038/jcbfm.2011.153. 78. Rasmussen, Peter M., Sune N. Jespersen and LØ. The effects of transit time heterogeneity on brain oxygenation during rest and functional activation. J Cereb Blood Flow Metab. 2015;3(35):432-442. 79. Angleys, H., & Ostergaard L. The combined effects of capillary transit time heterogeneity and hematocrit on brain oxygenation. J Cereb Blood Flow Metab. 2019;In Press. 80. Villringer, Arno, Andreas Them, Ute Lindauer, K. Einhäupl and UD. Capillary perfusion of the rat brain cortex. An in vivo confocal microscopy study. Circ Res. 1994;75(1):55-62. 81. Buxton RB, Frank LR. A model for the coupling between cerebral blood flow and oxygen metabolism during neural stimulation. J Cereb Blood Flow Metab. 1997;17:64-72. doi:10.1097/00004647-199701000-00009. 82. Buxton RB, Frank LR. A Model for the Coupling Between Cerebral Blood Flow and Oxygen Metabolism During Neural Stimulation. 1997:64-72. 83. Renkin EM. BW Zweifach Award lecture: regulation of the microcirculation. Microvasc Res. 1985;3(30):1985. 84. H G. Anatomy of the Human Body.; 1918. 85. Kretschmann, H.J. and Weinrich W. No Title.; 1985. 86. Damasio H. A computed tomography guide to the identification of cerebral vascular territories. Arch Neurol. 1983;3(40):138-142. 87. Van Der Zwan, Albert ; Hillen B. Review of the Variability of the Territories of the Major Cerebral Arteries. Stroke. 1991;8(22):1078-1084. 88. Van Der Zwan, A ; Hillen, B ; Tulleken, C A ; Dujovny, M ; Dragovic L. Variability of the territories of the major cerebral arteries. J Neurosurg. 1992;77(6):927-940. 89. Laar PJ Van, Hendrikse J, Golay X, Lu H, Osch MJP Van, Grond J Van Der. In vivo flow territory mapping of major brain feeding arteries. 2006;29:136-144. 92 doi:10.1016/j.neuroimage.2005.07.011. 90. Duvernoy HM. Arterial territories of the human brain(Neurol).pdf. 1999;30:99-110. doi:10.1159/000333602. 91. Bladin, C. F.; Chambers, B. R.; Donnan GA. Confusing stroke terminology: Watershed or borderzone infarction? Stroke. 1993;3(24):477-478. 92. Evrard S, Woimant F, Coz P Le, et al. Watershed cerebral infarcts : Retrospective study of 24 cases Watersh ed cerebra l infarcts : Retrospective study of 24 cases. 2016;6412:6-9. doi:10.1080/01616412.1992.11740022. 93. Derdeyn CP, Khosla A, Videen TO, et al. Impairment and Border Zone – Region Infarction 1. 2001:195-201. 94. van der Land V, Mutsaerts HJMM, Engelen M, et al. Risk factor analysis of cerebral white matter hyperintensities in children with sickle cell disease. Br J Haematol. 2016;172(2):274-284. doi:10.1111/bjh.13819. 95. Fields ME, Guilliams KP, Ragan DK, et al. Regional oxygen extraction predicts border zone vulnerability to stroke in sickle cell disease. Neurology. 2018;90(13):10.1212/WNL.0000000000005194. doi:10.1212/WNL.0000000000005194. 96. Lao Z, Shen D, Liu D, et al. Computer-Assisted Segmentation of White Matter Lesions in 3D MR Images Using Support Vector Machine. Acad Radiol. 2008;15(3):300-313. doi:10.1016/j.acra.2007.10.012. 97. MJ C. The Cerebral Circulation: Morgan & Claypool Life Sciences.; 2009. 98. Takahashi T, Shirane R, Sato S, Yoshimoto T. Developmental Changes of Cerebral Blood Flow and Oxygen Metabolism in Children. 1999;(May):917-922. 99. Wu C, Honarmand AR, Schnell S, et al. Age-Related Changes of Normal Cerebral and Cardiac Blood Flow in. :1-13. doi:10.1161/JAHA.115.002657. 100. Chen JJ, Rosas HD, Salat DH. Age-associated reductions in cerebral blood flow are independent from regional atrophy. Neuroimage. 2011;55(2):468-478. doi:10.1016/j.neuroimage.2010.12.032. 101. Bush AM, Borzage MT, Choi S, et al. Determinants of resting cerebral blood flow in sickle cell disease. Am J Hematol. 2016;91(9):912-917. doi:10.1002/ajh.24441. 102. Jordan LC, Gindville MC, Scott AO, et al. Non-invasive imaging of oxygen extraction fraction in adults with sickle cell anaemia. Brain. 2016;139(3):738-750. 93 doi:10.1093/brain/awv397. 103. Numaguchi Y, Haller JS, Humbert JR, et al. Cerebral blood flow mapping using stable Xenon-enhanced CT in sickle cell cerebrovascular disease. Neuroradiology. 1990;32(4):289-295. doi:10.1007/BF00593047. 104. Oguz KK, Golay X, Pizzini FB, et al. Sickle cell disease: continuous arterial spin-labeling perfusion MR imaging in children. Radiology. 2003;227(2):567-574. doi:10.1148/radiol.2272020903\r2272020903 [pii]. 105. Petersen ET, Zimine I, Ho YCL, Golay X. Non-invasive measurement of perfusion: A critical review of arterial spin labelling techniques. Br J Radiol. 2006;79(944):688-701. doi:10.1259/bjr/67705974. 106. Chai Y, Coloigner J, Bush AM, Wood JC, Lepore N. Regional Cerebral Blood Flow Measurement in Patients with Sickle Cell Disease Using Pseudo Continuous Arterial Spin Labeling. 2016:2016. 107. Van Den Tweel XW, Nederveen AJ, Majoie CBLM, et al. Cerebral blood flow measurement in children with sickle cell disease using continuous arterial spin labeling at 3.0-tesla MRI. Stroke. 2009;40(3):795-800. doi:10.1161/STROKEAHA.108.523308. 108. Chai Y, Bush AM, Coloigner J, et al. White matter has impaired resting oxygen delivery in sickle cell patients. Am J Hematol. 2019;4(94):467-474. doi:10.1002/ajh.25423. 109. Mutsaerts HJMM, Richard E, Heijtel DFR, Van Osch MJP, Majoie CBLM, Nederveen AJ. Gray matter contamination in arterial spin labeling white matter perfusion measurements in patients with dementia. NeuroImage Clin. 2014;4:139-144. doi:10.1016/j.nicl.2013.11.003 Review Article. 110. Perfusion asymmetries and flow in children with Sickle Cell Disease assessed by pseudo- continuous Arterial Spin Labeling and Phase Contrast MRI. Perfusion. 2011;19:2011. 111. Mutsaerts HJMM, van Dalen JW, Heijtel DFR, et al. Cerebral Perfusion Measurements in Elderly with Hypertension Using Arterial Spin Labeling. PLoS One. 2015;10(8):e0133717. doi:10.1371/journal.pone.0133717. 112. Marstrand JR. Cerebral Perfusion and Cerebrovascular Reactivity Are Reduced in White Matter Hyperintensities. Stroke. 2002;33(4):972-976. doi:10.1161/01.STR.0000012808.81667.4B. 113. Ford AL, Ragan DK, Fellah S, et al. Silent infarcts in sickle cell anemia occur in the 94 borderzone region and are associated with low cerebral blood flow. Blood. 2018;(314):blood-2018-04-841247. doi:10.1182/blood-2018-04-841247. 114. Zhipei Liang PCL. Principles of Magnetic Resonance Imaging: A Signal Processing Perspective.; 1999. 115. Nishimura DG. Principles of Magnetic Resonance Imaging.; 2010. 116. Detre JA, Zhang W R DA. Tissue specific perfusion imaging using arterial spin labeling. NMR Biomed. 1994;9(5):45-82. 117. Fan AP, Jahanian H, Holdsworth SJ ZG. Comparison of cerebral blood flow measurement with [15O]-water positron emission tomography and arterial spin labeling magnetic resonance imaging: A systematic review. J Cereb Blood Flow Metab. 2016;36(5):842- 861. 118. Yonas H, Pindzola RP JD. Xenon/computed tomography cerebral blood flow and its use in clinical management. Neurosurg Clin N Am. 1996;7(4):605-616. 119. Wolf RL, Alsop DC, McGarvey ML, Maldjian J a, Wang J, Detre J a. Susceptibility contrast and arterial spin labeled perfusion MRI in cerebrovascular disease. J neuroimaging. 2003;13(1):17-27. doi:10.1177/1051228402239713. 120. Golay X, Hendrikse J, Lim TCC. Perfusion imaging using arterial spin labeling. Top Magn Reson Imaging. 2004;15(1):10-27. doi:10.1097/00002142-200402000-00003. 121. Wang J, Alsop DC, Li L, et al. Comparison of quantitative perfusion imaging using arterial spin labeling at 1.5 and 4.0 Tesla. Magn Reson Med. 2002;48(2):242-254. doi:10.1002/mrm.10211. 122. et. al. DC. Recommended implementation of arterial spin-labeled perfusion MRI for clinical applications: A consensus of the ISMRM perfusion study group and the European consortium for ASL in dementia. Magn Reson Med. 2015;73(1):102-116. 123. Jung Y, Wong EC, Liu TT. Multiphase pseudocontinuous arterial spin labeling (MP- PCASL) for robust quantification of cerebral blood flow. Magn Reson Med. 2010;64(3):799-810. doi:10.1002/mrm.22465. 124. Bush A, Chai Y, Choi SY, et al. Pseudo continuous arterial spin labeling quantification in anemic subjects with hyperemic cerebral blood flow. Magn Reson Imaging. 2018;47(August 2017):137-146. doi:10.1016/j.mri.2017.12.011. 125. Wu WC, Fernández-Seara M, Detre JA, Wehrli FW, Wang J. A theoretical and 95 experimental investigation of the tagging efficiency of pseudocontinuous arterial spin labeling. Magn Reson Med. 2007;58(5):1020-1027. doi:10.1002/mrm.21403. 126. Chai Y, Bush A, Coloigner J, et al. An experimental investigation of labeling efficiency for pseudo-continuous arterial spin labeling. In: Proceedings - International Symposium on Biomedical Imaging. Vol 2016-June. ; 2016:1101-1104. doi:10.1109/ISBI.2016.7493458. 127. et. al. SA. Estimation of labeling efficiency in pseudocontinuous arterial spin labeling. Magn Reson Med. 2010;63(3):765-771. 128. Aslan S. et. al. Estimation of labeling efficiency in pseudocontinuous arterial spin labeling. Magn Reson Med. 2010;63(3):765-771. 129. Alsop DC, Detre JA. Reduced Transit-Time Sensitivity in Noninvasive Magnetic Resonance Imaging of Human Cerebral Blood Flow. J Cereb Blood Flow Metab. 1996:1236-1249. doi:10.1097/00004647-199611000-00019. 130. Blinman T, Maggard M. Rational manipulation of oxygen delivery. J Surg Res. 2000;92(1):120-141. doi:10.1006/jsre.2000.5830. 131. Beaulieu, Christian and PSA. Determinants of anisotropic water diffusion in nerves. Magn Reson Med. 1994;31(4):394-400. 132. Basser, Peter J., James Mattiello and DL. Estimation of the effective self-diffusion tensor from the NMR spin echo. J Magn Reson. 1994;B(103):247-254. 133. Review P. Neuroimaging findings in sickle cell disease. 2014;(April). doi:10.1259/bjr.20130699. 134. Debaun MR, Derdeyn CP, III RCM. Etiology of Strokes in Children with Sickle Cell Anemia. Ment Retard Dev Disabil Res Rev. 2007;13(August):313 – 320. doi:10.1002/mrdd. 135. Calamante F, Kirkham FJ, Bynevelt M, et al. Diffusion and Perfusion Magnetic Resonance Imaging in Childhood Stroke. 1999. 136. Moritani T, Numaguchi Y, Lemer NB, et al. Sickle cell cerebrovascular disease Usual and unusual findings on MR imaging and MR angiography. 2004;28:173-186. doi:10.1016/S0899-7071(03)00121-9. 137. Sotak CH. The role of diffusion tensor imaging in the evaluation of ischemic brain injury ± a review. 2002:561-569. doi:10.1002/nbm.786. 96 138. Smith SM, Jenkinson M, Johansen-berg H, et al. Tract-based spatial statistics : Voxelwise analysis of multi-subject diffusion data. 2006;31:1487-1505. doi:10.1016/j.neuroimage.2006.02.024. 139. Tuch DS. Q-ball imaging. Magn Reson Med. 2004;52(6):1358-1372. doi:10.1002/mrm.20279. 140. Maillard P, Fletcher E, Harvey D, et al. White matter hyperintensity penumbra. Stroke. 2011;42(7):1917-1922. doi:10.1161/STROKEAHA.110.609768. 141. Klöppel S, Abdulkadir A, Hadjidemetriou S, et al. A comparison of different automated methods for the detection of white matter lesions in MRI data. Neuroimage. 2011;57(2):416-422. doi:10.1016/j.neuroimage.2011.04.053. 142. Scheltens P, Erkinjunti T, Leys D, et al. White matter changes on CT and MRI: an overview of visual rating scales. European Task Force on Age-Related White Matter Changes. Eur Neurol. 1998;39(2):80-89. doi:10.1159/000007921. 143. Gibson E, Gao F, Black SE, Lobaugh NJ. Automatic segmentation of white matter hyperintensities in the elderly using FLAIR images at 3T. J Magn Reson Imaging. 2010;31(6):1311-1322. doi:10.1002/jmri.22004. 144. Admiraal-Behloul F, Van Den Heuvel DMJ, Olofsen H, et al. Fully automatic segmentation of white matter hyperintensities in MR images of the elderly. Neuroimage. 2005;28(3):607-617. doi:10.1016/j.neuroimage.2005.06.061. 145. Smart SD, Firbank MJ, O’Brien JT. Validation of automated white matter hyperintensity segmentation. J Aging Res. 2011;2011:391783. doi:10.4061/2011/391783. 146. Lao Z, Shen D, Liu D, et al. Computer-Assisted Segmentation of White Matter Lesions in 3D MR Images Using Support Vector Machine. Acad Radiol. 2008;15(3):300-313. doi:10.1016/j.acra.2007.10.012. 147. Steenwijk MD, Pouwels PJW, Daams M, et al. Accurate white matter lesion segmentation by k nearest neighbor classification with tissue type priors (kNN-TTP). NeuroImage Clin. 2013;3:462-469. doi:10.1016/j.nicl.2013.10.003. 148. Caligiuri ME, Perrotta P, Augimeri A, Rocca F, Quattrone A, Cherubini A. Automatic Detection of White Matter Hyperintensities in Healthy Aging and Pathology Using Magnetic Resonance Imaging: A Review. Neuroinformatics. 2015;13(3):261-276. doi:10.1007/s12021-015-9260-y. 97 149. M.A. Hearst ; S.T. Dumais ; E. Osuna ; J. Platt ; B. Scholkopf. Support vector machines. IEEE Intell Syst their Appl. 1994;3(4):18-28. 150. Duda, R. O., Hart, P. E., & Stork DG. Pattern Classification. New York: Wiley; 1973. 151. Caligiuri ME, Perrotta P, Augimeri A, Rocca F, Quattrone A, Cherubini A. Automatic Detection of White Matter Hyperintensities in Healthy Aging and Pathology Using Magnetic Resonance Imaging: A Review. Neuroinformatics. 2015;13(3):261-276. doi:10.1007/s12021-015-9260-y. 152. Yi L. Support vector machines and the Bayes rule in classification. Data Min Knowl Discov. 2002;6(5):259–275. 153. Anbeek P, Vincken KL, Van Osch MJP, Bisschops RHC, Van Der Grond J. Probabilistic segmentation of white matter lesions in MR imaging. Neuroimage. 2004;21(3):1037-1044. doi:10.1016/j.neuroimage.2003.10.012. 154. Zhang Y, Brady M, Smith S. Segmentation of brain MR images through a hidden Markov random field model and the expectation-maximization algorithm. IEEE Trans Med Imaging. 2001;20(1):45-57. doi:10.1109/42.906424. 155. Chyzhyk D, Price CC. Brain White Matter Lesion Segmentation with 2D/3D CNN. 2017;(Md):394-403. doi:10.1007/978-3-319-59740-9. 156. Dolz J, Desrosiers C, Ayed I Ben. NeuroImage 3D fully convolutional networks for subcortical segmentation in MRI : A large-scale study. Neuroimage. 2017;(April):1-15. doi:10.1016/j.neuroimage.2017.04.039. 157. Brosch T, Yoo Y, Tang LYW, Li DKB, Traboulsee A, Tam R. Deep Convolutional Encoder Networks for Multiple Sclerosis Lesion Segmentation. 2015:3-11. doi:10.1007/978-3-319-24574-4. 158. Prasoon A, Petersen K, Igel C. Deep Feature Learning for Knee Cartilage Segmentation Using a Triplanar Convolutional Neural Network. 2011;(09). 159. Kushner BH, LaQuaglia MP, Wollner N, et al. Desmoplastic small round-cell tumor: Prolonged progression-free survival with aggressive multimodality therapy. J Clin Oncol. 1996;14(5):1526-1531. doi:10.1002/pbc. 160. Moeskops P, Viergever MA, Mendrik AM, De Vries LS, Benders MJNL, Isgum I. Automatic Segmentation of MR Brain Images with a Convolutional Neural Network. IEEE Trans Med Imaging. 2016;35(5):1252-1261. doi:10.1109/TMI.2016.2548501. 98 161. Kamnitsas K, Ledig C, Newcombe VFJ, et al. Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation. Med Image Anal. 2017;36:61-78. doi:10.1016/j.media.2016.10.004. 162. Chen H, Dou Q, Yu L, Qin J, Heng PA. VoxResNet: Deep voxelwise residual networks for brain segmentation from 3D MR images. Neuroimage. 2017;(April):1-10. doi:10.1016/j.neuroimage.2017.04.041. 163. Goodfellow IJ, Pouget-abadie J, Mirza M, Xu B, Warde-farley D. Generative Adversarial Nets. :1-9. 164. Reeth EVAN, Tham IWK, Tan CH, Poh CLOO. Super-Resolution in Magnetic Resonance Imaging : A Review. Concepts Magn Reson Part A. 2012;40(6):306-325. doi:10.1002/cmra. 165. Pham C. BRAIN MRI SUPER-RESOLUTION USING DEEP 3D CONVOLUTIONAL NETWORKS Aur ´ Institut Mines T ´ el ´ Institut Mines T ´. 2017:197-200. doi:10.1109/ISBI.2017.7950500. 166. Kim J, Lee JK, Lee KM. Accurate Image Super-Resolution Using Very Deep Convolutional Networks. 2016 IEEE Conf Comput Vis Pattern Recognit. 2016:1646-1654. doi:10.1109/CVPR.2016.182. 167. Shi F, Cheng J, Wang L, et al. LRTV : MR Image Super-Resolution With Low-Rank and Total Variation Regularizations. 2015;34(12):2459-2466. 168. Chen Y, Xie Y, Zhou Z, Shi F, Christodoulou AG, Li D. MRI super-resolution using deep 3D convolutional networks. In: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). IEEE; 2018:739-742. doi:10.1109/ISBI.2018.8363679. 169. Dong, C., Loy, C.C. and Tang X. Accelerating the super-resolution convolutional neural network. In: In European Conference on Computer Vision. ; 2016:391-407. 170. Sánchez I, Vilaplana V. Brain MRI super-resolution using 3D generative adversarial networks. 2018;(Midl):1-8. 171. Chen, Y., Shi, F., Christodoulou, A.G., Xie, Y., Zhou, Z. and Li D. Efficient and Accurate MRI Super-Resolution Using a Generative Adversarial Network and 3D Multi-level Densely Connected Network. In: In International Conference on Medical Image Computing and Computer-Assisted Intervention. ; 2018:91-99. 172. Miller ST, Macklin EA, Pegelow CH, et al. Silent infarction as a risk factor for overt 99 stroke in children with sickle cell anemia: A report from the Cooperative Study of Sickle Cell Disease. J Pediatr. 2001;139(3):385-390. doi:10.1067/mpd.2001.117580. 173. Hoppe C. Defining stroke risk in children with sickle cell anaemia. Br J Haematol. 2005;128(6):751-766. doi:10.1111/j.1365-2141.2004.05310.x. 174. Guilliams KP, Fields ME, Ragan DK, et al. Red cell exchange transfusions lower cerebral blood flow and oxygen extraction fraction in pediatric sickle cell anemia. Blood. 2018;131(9):1012-1021. doi:10.1182/blood-2017-06-789842. 175. Bush AM, Borzage MT, Choi S, et al. Determinants of resting cerebral blood flow in sickle cell disease. Am J Hematol. 2016;91(9):912-917. doi:10.1002/ajh.24441. 176. Adams RJ, Ataga KI, Ballard H, et al. The management of sickle cell disease. Natl Institutes Heal. 2002:1-206. 177. Smith S, Jenkinson M, Woolrich M. Advances in functional and structural MR image analysis and implementation as FSL. Neuroimage. 2004;23(November):S208-S219. doi:10.1016/j.neuroimage.2004.07.051. 178. Vu CQ, Wood JC. Title: White Matter Hyperintensities Segmentation and Stereological Bias Correction. In: International Society for Magnetic Resonance in Medicine (ISMRM). ; 2017:1. 179. Xu B, Chai Y, Galarza CM, et al. Orchestral Fully Convolutional Networks for Small Lesion Segmentation in Brain Mri. In: In Biomedical Imaging (ISBI 2018), 2018 IEEE 15th International Symposium on, Pp. 889-892. IEEE, 2018. ; 2018:889-892. 180. Yushkevich PA, Piven J, Hazlett HC, et al. User-guided 3D active contour segmentation of anatomical structures: Significantly improved efficiency and reliability. Neuroimage. 2006;31(3):1116-1128. doi:10.1016/j.neuroimage.2006.01.015. 181. Asllani I, Borogovac A, Brown TR. Regression algorithm correcting for partial volume effects in arterial spin labeling MRI. Magn Reson Med. 2008;60(6):1362-1371. doi:10.1002/mrm.21670. 182. Lu H, Clingman C, Golay X, Zijl P Van. What is the longitudinal relaxation time ( T1 ) of blood at 3 . 0 Tesla ? Blood. 2003;11(3):21205-21205. 183. MacIntosh BJ, Filippini N, Chappell MA, Woolrich MW, Mackay CE, Jezzard P. Assessment of arterial arrival times derived from multiple inversion time pulsed arterial spin labeling MRI. Magn Reson Med. 2010;63(3):641-647. doi:10.1002/mrm.22256. 100 184. Jensen JH, Helpern JA, Ramani A, Lu H, Kaczynski K. Diffusional kurtosis imaging: The quantification of non-Gaussian water diffusion by means of magnetic resonance imaging. Magn Reson Med. 2005;53(6):1432-1440. doi:10.1002/mrm.20508. 185. Herold S, Brozovic M, Gibbs J, et al. Measurement of regional cerebral blood flow, blood volume and oxygen metabolism in patients with sickle cell disease using positron emission tomography. Stroke. 1986;17(4):692-698. doi:10.1161/01.STR.17.4.692. 186. Prohovnik I, Hurlet-Jensen A, Adams R, De Vivo D, Pavlakis SG. Hemodynamic etiology of elevated flow velocity and stroke in sickle-cell disease. J Cereb Blood Flow Metab. 2009;29(4):803-810. doi:10.1038/jcbfm.2009.6. 187. Brown M, Marshall J. Regulation of Cerebral Blood Flow in Response to Changes in Blood Viscosity. Lancet. 1985:604-609. 188. Brown MM, Marshall J. Effect of plasma exchange on blood viscosity and cerebral blood flow. Br Med J. 1982;284(6331):1733-1736. 189. Cahill LS, Gazdzinski LM, Tsui AK, et al. Functional and anatomical evidence of cerebral tissue hypoxia in young sickle cell anemia mice. J Cereb Blood Flow Metab. 2017;37(3):994-1005. doi:10.1177/0271678X16649194. 190. Hare GMT, Tsui AKY, Ozawa S, Shander A. Anaemia: Can we define haemoglobin thresholds for impaired oxygen homeostasis and suggest new strategies for treatment? Best Pract Res Clin Anaesthesiol. 2013;27(1):85-98. doi:10.1016/j.bpa.2012.12.002. 191. Ni WW, Christen T, Rosenberg J, Zun Z, Moseley ME, Zaharchuk G. Imaging of cerebrovascular reserve and oxygenation in Moyamoya disease. J Cereb Blood Flow Metab. 2017;37(4):1213-1222. doi:10.1177/0271678X16651088. 192. Kedar A, Drane WE, Shaeffer D, Nicole M, Adams C. Measurement of cerebrovascular flow reserve in pediatric patients with sickle cell disease. Pediatr Blood Cancer. 2006;46(2):234-238. doi:10.1002/pbc.20475. 193. Borzage MT, Bush AM, Choi S, et al. Predictors of Cerebral Blood Flow in Patients With and Without Anemia. J Appl Physiol. 2016:jap.00994.2015. doi:10.1152/japplphysiol.00994.2015. 194. Moeini M, Lu X, Avti PK, et al. Compromised microvascular oxygen delivery increases brain tissue vulnerability with age. Sci Rep. 2018;8(1):1-17. doi:10.1038/s41598-018- 26543-w. 101 195. Detterich J, Alexy T, Rabai M, et al. Low-shear red blood cell oxygen transport effectiveness is adversely affected by transfusion and further worsened by deoxygenation in sickle cell disease patients on chronic transfusion therapy. Transfusion. 2013;53(2):297- 305. doi:10.1111/j.1537-2995.2012.03822.x. 196. Detterich JA, Kato RM, Rabai M, Meiselman HJ, Coates TD, Wood JC. Chronic Transfusion Therapy Improves but Does Not Normalize Systemic and Pulmonary Vasculopathy in Sickle Cell Disease. Vol 126.; 2015. doi:10.1182/blood-2014-12-614370. 197. Bush A, Borzage M, Detterich J, et al. Empirical model of human blood transverse relaxation at 3 T improves MRI T2 oximetry. Magn Reson Med. 2017;77(6):2364-2371. doi:10.1002/mrm.26311. 198. Bush AM, Coates TD, Wood JC. Diminished cerebral oxygen extraction and metabolic rate in sickle cell disease using T2 relaxation under spin tagging MRI. Magn Reson Med. 2018;80(1):294-303. doi:10.1002/mrm.27015. 199. Rebel A, Ulatowski JA, Kwansa H, Bucci E, Koehler RC. Cerebrovascular response to decreased hematocrit: effect of cell-free hemoglobin, plasma viscosity, and CO2. Am J Physiol Heart Circ Physiol. 2003;285(4):H1600-H1608. doi:10.1152/ajpheart.00077.2003. 200. Guilliams KP, Fields ME, Ragan DK, et al. Large-Vessel Vasculopathy in Children With Sickle Cell Disease: A Magnetic Resonance Imaging Study of Infarct Topography and Focal Atrophy. Pediatr Neurol. 2017;69:49-57. doi:10.1016/j.pediatrneurol.2016.11.005. 201. Wang J, Licht DJ, Jahng GH, et al. Pediatric perfusion imaging using pulsed arterial spin labeling. J Magn Reson Imaging. 2003;18(4):404-413. doi:10.1002/jmri.10372. 202. Vaclavu L, Van Der Land V, Heijtel DFR, et al. In vivo T1 of blood measurements in children with sickle cell disease improve cerebral blood flow quantification from arterial spin-labeling MRI. Am J Neuroradiol. 2016;37(9):1727-1732. doi:10.3174/ajnr.A4793. 203. Piel FB, Patil AP, Howes RE, et al. Global epidemiology of Sickle haemoglobin in neonates: A contemporary geostatistical model-based map and population estimates. Lancet. 2013;381(9861):142-151. doi:10.1016/S0140-6736(12)61229-X. 204. King AA, Rodeghier MJ, Panepinto JA, et al. Silent cerebral infarction, income, and grade retention among students with sickle cell anemia. Am J Hematol. 2014;89(10):E188-E192. doi:10.1002/ajh.23805. 102 205. Bernaudin F, Verlhac S, Arnaud C, et al. Impact of early transcranial Doppler screening and intensive therapy on cerebral vasculopathy outcome in a newborn sickle cell anemia cohort CME article Impact of early transcranial Doppler screening and intensive therapy on cerebral vasculopathy outcome . 2014;117(4):1130-1140. doi:10.1182/blood-2010-06- 293514. 206. Choi S, Neil SHO, Joshi AA, et al. Anemia predicts lower white matter volume and cognitive performance in sickle and non-sickle cell anemia syndrome. 2019;(June):1-11. doi:10.1002/ajh.25570. 207. Schatz J, Brown RT, Pascual JM, Hsu L, DeBaun MR. Poor school and cognitive functioning with silent cerebral infarcts and sickle cell disease. Neurology. 2001;56(8):1109-1111. doi:10.1212/WNL.56.8.1109. 208. Prussien K V, Debaun MR, Yarboi J, et al. Cognitive Function , Coping , and Depressive Symptoms in Children and Adolescents with Sickle Cell Disease. 2018;43(November 2017):543-551. doi:10.1093/jpepsy/jsx141. 209. Balfour P. White matter integrity and processing speed in sickle cell anemia. 2018;0. doi:10.1212/WNL.0000000000005644. 210. Jacobs, Heidi IL, Elizabeth C. Leritz, Victoria J. Williams, Martin PJ Van Boxtel, Wim van der Elst, Jelle Jolles, Frans RJ Verhey, Regina E. McGlinchey, William P. Milberg and DHS. Association between white matter microstructure, executive functions, and processing speed in older adults: the impact of vascular health. Hum Brain Mapp. 2013;34(1):77-95. 211. Turken, U., Susan Whitfield-Gabrieli, Roland Bammer, Juliana V. Baldo, Nina F. Dronkers and JDG. Cognitive processing speed and the structure of white matter pathways: convergent evidence from normal variation and lesion studies. Neuroimage. 2008;42(2):1032-1044. 212. Sundgren PC, Dong Q, Gómez-Hassan D, Mukherji SK, Maly P, Welsh R. Diffusion tensor imaging of the brain: Review of clinical applications. Neuroradiology. 2004;46(5):339-350. doi:10.1007/s00234-003-1114-x. 213. Daniel F, Pegelow H. Cognitive Functioning and Brain Magnetic Children Cell Resonance Disease Imaging in. 2016;97. 214. Lao, Yi, Yue Kang, Olivier Collignon, Caroline Brun, Shadi B. Kheibai, Flamine Alary, 103 James Gee, Marvin D. Nelson, Franco Lepore and NL. A study of brain white matter plasticity in early blinds using tract-based spatial statistics and tract statistical analysis. Neuroreport. 2015;26(18):1151-1154. 215. Yushkevich PA, Zhang H, Simon TJ, Gee JC. Structure-specific statistical mapping of white matter tracts. 2008;41:448-461. doi:10.1016/j.neuroimage.2008.01.013. 216. Pecheva D, Yushkevich P, Batalle D, et al. NeuroImage A tract-speci fi c approach to assessing white matter in preterm infants. Neuroimage. 2017;157(March):675-694. doi:10.1016/j.neuroimage.2017.04.057. 217. Andersson JLR, Skare S, Ashburner J. How to correct susceptibility distortions in spin- echo echo-planar images: application to diffusion tensor imaging. 2003;20:870-888. doi:10.1016/S1053-8119(03)00336-7. 218. Zhang H, Yushkevich PA, Rueckert D, Gee JC. Unbiased White Matter Atlas Construction Using Diffusion Tensor Images. In: Medical Image Computing and Computer-Assisted Intervention – MICCAI 2007. ; 2017:211-218. 219. Yang W, Zhang X, Tian Y, Wang W, Xue J. Deep Learning for Single Image Super- Resolution : A Brief Review. :1-15. 220. Zhang H, Avants BB, Yushkevich PA, et al. High-dimensional spatial normalization of diffusion tensor images improves the detection of white matter differences: An example study using amyotrophic lateral sclerosis. IEEE Trans Med Imaging. 2007;26(11):1585- 1597. doi:10.1109/TMI.2007.906784. 221. Lawes INC, Barrick TR, Murugam V, et al. Atlas-based segmentation of white matter tracts of the human brain using diffusion tensor tractography and comparison with classical dissection. 2008;39:62-79. doi:10.1016/j.neuroimage.2007.06.041. 222. Yushkevich, Paul A., Hui Zhang and JCG. Continuous medial representation for anatomical structures. IEEE Trans Med Imaging. 2006;25(12):1547-1564. 223. Yushkevich PA. Continuous medial representation of brain structures using the biharmonic PDE. Neuroimage. 2009;45(1):99-110. 224. Zhang H, Awate SP, Das SR, et al. A tract-specific framework for white matter morphometry combining macroscopic and microscopic tract features. Lect Notes Comput Sci (including Subser Lect Notes Artif Intell Lect Notes Bioinformatics). 2009;5762 LNCS(PART 2):141-149. doi:10.1007/978-3-642-04271-3_18. 104 225. Wechsler D. Wechsler Intelligence Scale for Children–Fourth Edition (WISC-IV). San Antonio, TX; 2003. 226. Wechsler D. Wechsler Adult Intelligence Scale–Fourth Edition (WAIS–IV). 4th ed. San Antonio, TX: NCS Pearson; 2008. 227. Delis DC, Kaplan E, Kramer J. Delis-Kaplan Executive Function System (D-KEFS).; 2001. 228. Nichols TE, Holmes AP. Nonparametric Permutation Tests For Functional Neuroimaging : A Primer with Examples. 2001;25(August 1999):1-25. 229. Benjamini, Yoav and YH. Controlling the false discovery rate: a practical and powerful approach to multiple hypothesis testing. J R Stat Soc Ser B. 2006;1(57):289–300. 230. Klingberg T, Vaidya CACJ. Myelination and organization of the frontal white matter in children : a diffusion tensor MRI study. 1999;10(13):2817-2821. 231. Ulu AM. Monitoring brain development with quantitative diffusion. 2002;3:286-292. 232. Schmithorst VJ, Wilke M, Dardzinski BJ, Holland SK. Correlation of White Matter Diffusivity and Anisotropy with Age during Childhood and Adolescence : A Cross- sectional Diffusion-Tensor MR Imaging Study 1. 2002;(11). 233. Kanaan A, Farahani R, Douglas RM, et al. Effect of chronic continuous or intermittent hypoxia and reoxygenation on cerebral capillary density and myelination. 2019;0831. doi:10.1152/ajpregu.00535.2005. 234. Martino, J., Brogna, C., Robles, S. G., Vergani, F., & Duffau H. Anatomic dissection of the inferior fronto-occipital fasciculus revisited in the lights of brain stimulation data. Cortex. 2010;5(46):691-699. 235. Sarubbo S, De BA, Maldonado IL, Basso G, Duffau H. Frontal terminations for the inferior fronto-occipital fascicle: anatomical dissection, DTI study and functional considerations on a multi-component bundle. Brain Struct Funct. 2013;4(312):235-236. 236. Crawford RD, Jonassaint CR. Adults with sickle cell disease may perform cognitive tests as well as controls when processing speed is taken into account : a preliminary case – control study. 2015. doi:10.1111/jan.12755. 237. Schatz J, Schlenz AM, Smith KE, Roberts CW. Predictive validity of developmental screening in young children with sickle cell disease : a longitudinal follow-up study. 2018. doi:10.1111/dmcn.13689. 105 238. Steen RG, Miles MA, Helton KJ, et al. Cognitive Impairment in Children with Hemoglobin SS Sickle Cell Disease: Relationship to MR Imaging Findings and Hematocrit. AJNR Am J Neuroradiol. 2003;24(3):382-389. http://www.ajnr.org/content/24/3/382.long. 239. Viana-Baptista, M., Bugalho, P., Jordão, C., Ferreira, N., Ferreira, Á., Secca MF. Cognitive function correlates with frontal white matter apparent diffusion coefficients in patients with leukoaraiosis. J Neurol. 2008;10(255):360-366. 240. Brébion, G., David, A.S., Bressan, R.A., Pilowsky L. Role of processing speed and depressed mood on encoding, storage, and retrieval memory functions in patients diagnosed with schizophrenia. J Int Neuropsychol Soc. 2007;(13):99-107. 241. Kral, Mary C. and RTB. Transcranial Doppler ultrasonography and executive dysfunction in children with sickle cell disease. J Pediatr Psychol. 2004;3(29):185-195. 242. Schatz J, White DA, Moinuddin A, Armstrong M, DeBaun MR. Lesion burden and cognitive morbidity in children with sickle cell disease. J Child Neurol. 2002;17(12):891- 895. doi:10.1177/08830738020170122401. 243. Brown, Ronald T., Iris Buchanan, Karla Doepke, James R. Eckman, Kevin Baldwin, Brian Goonan and SS. Cognitive and Academic Functioning in Children With Sickle-Cell Disease. J Clin Child Psychol. 2007;2(22):207-218. 244. Righart, R., Duering, M., Gonik, M., Jouvent, E., Reyes, S., Hervé, D., Chabriat, H., Dichgans M. Impact of regional cortical and subcortical changes on processing speed in cerebral small vessel disease. NeuroImage Clin. 2013;2(48):854-861. 245. Felmingham, Kim L., Ian J. Baguley and AMG. Effects of diffuse axonal injury on speed of information processing following severe traumatic brain injury. Neuropsychology. 2004;3(18):564. 246. McRobbie D, Moore E, Graves M PM. MRI: From Picture to Proton.; 2006. 247. Barker-Collo. Quality of life in multiple sclerosis: Does information-processing speed have an independent effect? Arch Clin Neuropsychol. 2006;8(61):167-174. 248. Demaree, H.A., DeLuca, J., Gaudino, E.A., Diamond B. Speed of information processing as a key deficit in multiple sclerosis: implications for rehabilitation. J Neurol Neurosurg Psychiatry. 1999;4(67):661-663. 249. Genova, H.M., Hillary, F.G., Wylie, G., Rypma, B., Deluca J. Examination of processing 106 speed deficits in multiple sclerosis using functional magnetic resonance imaging. J Int Neuropsychol Soc. 2009;2(15):383-393. 250. Hinton KE, Lahey BB, Villalta-gil V, et al. Right Fronto-Subcortical White Matter Microstructure Predicts Cognitive Control Ability on the Go / No-go Task in a Community Sample. 2018;12(April):1-11. doi:10.3389/fnhum.2018.00127. 251. Bessette, Katie L. and MCS. Neurocognitive pathways in attention-deficit/hyperactivity disorder and white matter microstructure. Biol Psychiatry Cogn Neurosci Neuroimaging. 2019;4(3):233-242. 252. Hijmans CT, Fijnvandraat K, Grootenhuis MA, Geloven N Van, Peters M, Oosterlaan J. Neurocognitive Deficits in Children With Sickle Cell Disease : A Comprehensive Profile. 2011;(December 2010):783-788. doi:10.1002/pbc. 253. Daly, Brian, Mary C. Kral, Ronald T. Brown, David Elkin, Avi Madan-Swain, Monica Mitchell, Lori Crosby, David DeMatteo, Angela LaRosa and SJ. Ameliorating attention problems in children with sickle cell disease: a pilot study of methylphenidate. J Dev Behav Pediatr. 2014;33(3):244. 254. Tournier J, Mori S, Leemans A. Diffusion Tensor Imaging and Beyond. 2011;1556:1532- 1556. doi:10.1002/mrm.22924. 255. Goldstein, Andrea and FBM. Neuroanatomy, Corpus Callosum. StatPearls Publ. 2017. 256. O’Driscoll, K., & Leach JP. No longer Gage: an iron bar through the head: Early observations of personality change after injury to the prefrontal cortex. Br Med J. 1998;7(23):1673-1674. 257. Damasio, Antonio R. and SWA. The frontal lobes. Clin Neuropsychol. 1993;3:409-460. 258. Alvarez, Julie A. and EE. Executive function and the frontal lobes: a meta-analytic review. Neuropsychol Rev. 2006;16(1):17-42. 259. Yuan, Peng and NR. Prefrontal cortex and executive functions in healthy adults: a meta- analysis of structural neuroimaging studies. Neurosci Biobehav Rev. 2014;42:180-192. 260. Kamagata, K., Motoi, Y., Tomiyama, H., Abe, O., Ito, K., Shimoji, K., Suzuki, M., Hori, M., Nakanishi, A., Sano, T. and Kuwatsuru R. Relationship between cognitive impairment and white-matter alteration in Parkinson’s disease with dementia: tract-based spatial statistics and tract-specific analysis. Eur Radiol. 2013;7(23):1946-1955. 261. Somasundaram K, Kalavathi P. Analysis of Imaging Artifacts in MR Brain Images. Orient 107 J Comput Sci …. 2012. http://computerscijournal.org/dnload/Somasundaram-K-and- Kalavathi-P-/OJCSV05I01P135-141.pdf. 262. Ledig C, Theis L, Husz F, et al. Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network. :4681-4690. 263. Yi X, Walia E, Babyn P. Generative Adversarial Network in Medical Imaging: A Review. 2019. 264. Nazeri K, Ng E, Joseph T, Qureshi FZ, Ebrahimi M. EdgeConnect : Generative Image Inpainting with Adversarial Edge Learning. 265. Teeuwisse WM, Schmid S, Ghariq E, Veer IM, Osch MJP Van. Time-Encoded pseudoContinuous Arterial Spin Labeling : Basic Properties and Timing Strategies for Human Applications. 2014;1722:1712-1722. doi:10.1002/mrm.25083. 266. Wang J, Alsop DC, Song HK, et al. Arterial transit time imaging with flow encoding arterial spin tagging (FEAST). Magn Reson Med. 2003;50(3):599-607. doi:10.1002/mrm.10559. 267. Miller JH. Pediatric Brain MR Perfusion Imaging. J Radiol Nurs. 2017;36(1):31-37. doi:10.1016/j.jradnu.2017.01.001. 268. Welker K, Boxerman J, Kalnin XA, et al. WHITE PAPER ASFNR Recommendations for Clinical Performance of MR Dynamic Susceptibility Contrast Perfusion Imaging of the Brain. 2015:41-51. 269. An H, Lin W. Impact of Intravascular Signal on Quantitative Measures of Cerebral Oxygen Extraction and Blood Volume Under Normo- and Hypercapnic Conditions Using an Asymmetric Spin Echo Approach. 2003;716(May):708-716. doi:10.1002/mrm.10576. 270. Yablonskiy DA, Sukstanskii AL, He X. Blood oxygenation level-dependent ( BOLD ) - based techniques for the quanti fi cation of brain hemodynamic and metabolic properties – theoretical models and experimental approaches. 2013;(March 2012):963-986. doi:10.1002/nbm.2839. 271. Perlmutter JS, Powers WJ, Herscovitch P, Fox PT, Raichle ME. Regional Asymmetries of Cerebral Blood Flow , Blood Volume , and Oxygen Utilization and Extraction in Normal Subjects. 1987:64-67. 272. Jeurissen B, Tournier J, Dhollander T, Connelly A, Sijbers J. NeuroImage Multi-tissue constrained spherical deconvolution for improved analysis of multi-shell diffusion MRI 108 data. 2014;103:411-426. 273. Andersson JLR, Sotiropoulos SN. NeuroImage Non-parametric representation and prediction of single- and multi-shell diffusion-weighted MRI data using Gaussian processes. Neuroimage. 2015;122:166-176. doi:10.1016/j.neuroimage.2015.07.067. 274. Lin Z, Li Y, Su P, et al. Magnetic Resonance in Medicine Non-contrast MR imaging of blood-brain barrier permeability to water. 2018;(January):1507-1520. doi:10.1002/mrm.27141. 275. Strbian D, Durukan A, Pitkonen M, et al. THE BLOOD – BRAIN BARRIER IS CONTINUOUSLY OPEN FOR SEVERAL WEEKS FOLLOWING TRANSIENT FOCAL CEREBRAL ISCHEMIA. 2008;153:175-181. doi:10.1016/j.neuroscience.2008.02.012. 276. Jia Y, Gholipour A, Member S, He Z. A New Sparse Representation Framework for Reconstruction of an Isotropic High Spatial Resolution MR Volume From Orthogonal Anisotropic Resolution Scans. 2017;36(5):1182-1193. 277. Zoran D. From Learning Models of Natural Image Patches to Whole Image Restoration. 278. Dalca A V, Bouman KL, Freeman WT, Rost NS, Sabuncu MR, Golland P. Medical Image Imputation from Image Collections. :1-14.

Abstract (if available)

Abstract Patients with thalassemia intermedia, beta thalassemia major and sickle cell disease (SCD) are known to have higher risk of cerebrovascular disease and stroke. Although modern medical management has lowered overt stroke occurrence in patients with sickle cell disease (SCD), progressive white matter (WM) damage remains common. It is known that cerebral blood flow (CBF) increases to compensate for anemia, but sufficiency of cerebral oxygen delivery, especially in the WM, has not been systematically investigated. We measured cerebral perfusion by arterial spin labeling in SCD patients, non-sickle anemic patients, and race-matched healthy controls. Imaging data from control subjects was used to calculate maps for CBF and oxygen delivery for three study groups and their T-score maps. We found that whole brain and grey matter (GM) oxygen delivery were normal in SCD, but WM oxygen delivery was 35% lower than in controls. Age and hematocrit were the strongest predictors for WM CBF and oxygen delivery in patients with SCD. There was spatial co-localization between regions of low oxygen delivery and white matter hyperintensities on T2 FLAIR imaging. ❧ One of the consequences of impaired hemodynamic balance in SCD is microstructural damage in the WM. In this thesis, we investigated the impact of anemia and sickled hemoglobin separately on the microstructural changes in the 11 major WM tracts. Decreased FA in the genu of CC, left inferior CST and IFO was observed in chronically anemic patients, regardless of anemia subtype, while FA reductions in the CST and SLF were unique to SCD patients. Patients with WMHs had more significant FA abnormalities, which were found in the same areas as patients without WMHs. Slower processing speed and response inhibition skills were observed in SCD patients, consistent with WM involvement. Decreased FA values in the CC significantly correlated with all nine neurocognitive measures, which included processing speed, working memory, and executive functions, suggesting a critical importance for CC fiber integrity in core neurocognitive processes. ❧ Finally, to create a silent infarction “heat map” in the common image space so that multi-modal group analysis is possible, we utilized convolutional neural network (CNN) and achieved improved lesion segmentation performance by combining three CNNs with different patch size in parallel. To register low and anisotropic resolution images to high and isotropic template, we generated high resolution T2 MR scans using GAN. However, using GAN to gap the resolution and image quality discrepancy between research and clinical dataset requires a well-generalized network. ❧ Taking together, measurement of cerebral oxygen delivery and microstructural injury is necessary and imperative to identify an imaging biomarker of stroke, to unveil the onset, progression, and development of brain infarction, and to better understand how cerebral vascular disease would physiologically and hematologically impact the brain.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Susceptibility-weighted MRI for the evaluation of brain oxygenation and brain iron in sickle cell disease

PDF

Quantitative MRI for oxygenation imaging in cerebrovascular diseases

PDF

Estimation of cognitive brain activity in sickle cell disease using functional near-infrared spectroscopy and dynamic systems modeling

PDF

Assessing cerebral oxygen supply and demand in sickle cell anemia using MRI

PDF

Anemia and white matter: a diffusion MRI analysis of the human brain

$Quantitative MRI for the measurement of cerebral oxygen extraction fraction in sickle cell disease$

PDF

Quantitative MRI for the measurement of cerebral oxygen extraction fraction in sickle cell disease

PDF

Cerebral oxygen saturation and cerebrovascular instability in newborn infants with congenital heart disease compared to healthy controls

PDF

Neurological consequences of chronic anemia

PDF

Photoplethysmogram-based biomarker for assessing risk of vaso-occlusive crisis in sickle cell disease: machine learning approaches

PDF

Clinical significance of oxyhemoglobin dissociation curves on the interpretation of pulse oximetry in Sickle Cell disease

PDF

Role of the autonomic nervous system in vasoconstriction responses to mental stress in sickle cell disease: a bioengineering perspective

PDF

Unveiling the white matter microstructure in 22q11.2 deletion syndrome with diffusion magnetic resonance imaging

PDF

Modeling of cardiovascular autonomic control in sickle cell disease

PDF

Alpha globin and vascular reactivity in humans: loss of alpha globin genes in human subjects is associated with improved nitric oxide-mediated vascular perfusion

PDF

A virtual reality exergaming system to enhance brain health in older adults at risk for Alzheimer’s disease

PDF

Contributions of dynamic cerebrovascular function to cognitive decline and dementia: development and validation of a novel neuroimaging approach

PDF

Principal dynamic mode analysis of cerebral hemodynamics for assisting diagnosis of cerebrovascular and neurodegenerative diseases

PDF

Pulsatile and steady flow in central veins and its impact on right heart-brain hemodynamic coupling in health and disease

PDF

Placenta growth factor-miRNAs-lncRNAs axis in the regulation of ET-1 gene involved in pulmonary hypertension in sickle cell disease

PDF

Diffusion MRI of the human brain: signal modeling and quantitative analysis

Asset Metadata

Creator Chai, Yaqiong (author)

Core Title Cerebrovascular disease of white matter in patients with chronic anemia syndrome

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Biomedical Engineering

Publication Date 11/22/2019

Defense Date 07/26/2019

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag anemia,cerebrovascular disease,lesion,OAI-PMH Harvest,sickle cell disease,white matter

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Lepore, Natahsa (committee chair), Wood, John (committee chair), Rajagopalan, Vidya (committee member)

Creator Email chai.qiong@gmail.com,Yaqiong.Chai@loni.usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-238437

Unique identifier UC11673164

Identifier etd-ChaiYaqion-7877.pdf (filename),usctheses-c89-238437 (legacy record id)

Legacy Identifier etd-ChaiYaqion-7877.pdf

Dmrecord 238437

Document Type Dissertation

Rights Chai, Yaqiong

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA