Title Page

Contents

Abstract 10

Chapter 1. Introduction 12

1.1. Introduction 12

1.2. Research Aims 15

Chapter 2. Classification of Headache Disorders 17

2.1. Introduction 17

2.2. Materials and Methods 19

2.2.1. Subjects 19

2.2.2. Headache clinic registry 19

2.2.3. Stacked classifier model 28

2.2.4. Feature selection 30

2.2.5. Classifier performance evaluation 30

2.2.6. Comparison with other methods 31

2.2.7. Classification of secondary headache disorders 31

2.3. Results 32

2.3.1. Selected features 32

2.3.2. Classifier performances of the primary headache disorder 34

2.3.3. Comparison of feature selection methods 37

2.3.4. Comparison of binary classifiers 38

2.3.5. Classifier performances of secondary headache disorder 39

2.4. Discussion 41

Chapter 3. Segmentation-guided Prediction of Overall Survival in Brain Tumors 46

3.1. Introduction 46

3.2. Related Works 49

3.2.1. Medical Image Segmentation 49

3.2.2. Brain Tumor Segmentation 49

3.2.3. Overall Survival Prediction 50

3.2.4. Segmentation Guided Regression 50

3.3. Methods 52

3.3.1. Model Architecture 52

3.3.2. Training Objectives 56

3.4. Experiments 58

3.4.1. Datasets 58

3.4.3. Implementation details 66

3.5. Results 70

3.5.1. Brain Tumor Segmentation 70

3.5.2. Overall survival regression 72

3.5.3. Ablation Study 75

3.6. Discussion 82

Chapter 4. Segmentation of Cerebral Microbleeds 88

4.1. Introduction 88

4.2. Methods 92

4.2.1. Datasets 92

4.2.2. Pre-training with Masked Image Modeling 94

4.2.3. Cerebral Microbleed Segmentation 96

4.2.4. Ensemble Modeling 96

4.3. Experiments 97

4.3.1. Data Preprocessing 97

4.3.2. Implementation Details 98

4.3.3. Evaluation Metrics 99

4.4. Results 100

4.4.1. Cerebral Microbleed Segmentation 100

4.4.2. Ablation Study 100

4.5. Discussion 103

Chapter 5. Conclusion 105

References 109

논문요약 127

Table 2.1. The 75 questions used in our study. 21

Table 2.2. Distribution of primary headache subtypes. Values are reported as mean with standard deviation in parenthesis. P-values were obtained from... 26

Table 2.3. Selected features from different layers of the classifier. The features listed in the right column positively correlated with the target subtype... 33

Table 2.4. Classifier performance of both cohorts. 35

Table 2.5. Confusion matrix for the training cohort. 36

Table 2.6. Confusion matrix for the test cohort. 37

Table 2.7. Comparison of the proposed method and other feature selection methods in the test cohort in terms of classifier performance. 38

Table 2.8. Comparison of the proposed method with other feature selection methods in the test cohort in terms of classifier performance. 38

Table 2.9. Confusion matrix with secondary headache for the training cohort. 39

Table 2.10. Confusion matrix with secondary headache for the test cohort. 40

Table 3.1. Distribution of patients with high-grade glioma used in this study. 62

Table 3.2. Comparison of Dice scores on the brain tumor segmentation using the BraTS 2017 dataset. NEC/NET: Necrotic and non-enhancing tumor. 70

Table 3.3. Comparison of performance on the overall survival prediction on the BraTS 2017 dataset. 73

Table 3.4. Comparison of performance on the overall survival prediction on the UCSF-PDGM dataset. 73

Table 3.5. Ablation analysis of pre-training with low-grade glioma patients, focal loss, and contrastive loss on the BraTS 2017 dataset. The UNETR... 75

Table 3.6. Comparison of T1ce reconstruction performance for enhancing tumor region on the BraTS 2017 dataset. T1ce: Contrast-enhanced weighted-... 86

Table 4.1. Quantitative results of CMB segmentation and CMB with brain tissue segmentation tasks. CMB: Cerebral Microbleed. 101

Table 4.2. Ablation analysis of ensemble configurations in CMB with brain tissue segmentation task. MIM: masked image modeling, ViT: Vision... 101

Figure 2.1. Distribution of patients used in the study. (a) Number of subjects for each primary headache subtypes in the training cohort. (b) Number of... 27

Figure 2.2. Structure of the stacked classifier model. 29

Figure 3.1. Overview of our segmentation-guided overall survival regression method. Both the brain tumor segmentation branch and the overall... 53

Figure 3.2. An example of high-grade glioma patients on the BraTS 2017 dataset. The columns represent MRI slices of T1, contrast-enhanced T1, T2,... 60

Figure 3.3. An example of low-grade glioma patients on the BraTS 2017 dataset. The columns represent MRI slices of T1, contrast-enhanced T1, T2,... 61

Figure 3.4. An example of the WHO CNS grade IV patients in the UCSF-PDGM training cohort. The columns represent MRI slices of T1, contrast-... 64

Figure 3.5. An example of WHO CNS grade IV patients in the UCSF-PDGM test cohort. The columns represent MRI slices of T1, contrast-enhanced T1,... 65

Figure 3.6. An overall training strategy of our proposed network. (a) In the pre-training phase, low-grade glioma patients without survival information... 69

Figure 3.7. Qualitative results of our proposed method and baseline methods in an axial view, coronal view, and sagittal view. The columns are an input slice... 71

Figure 3.8. Scatter plot between the predicted overall survival and the actual overall survival of patients on the BraTS 2017 dataset with (a) UNETR... 74

Figure 3.9. Comparison of Spearman's rank correlation coefficient for overall survival regression with different backbones on the BraTS 2017 dataset for... 78

Figure 3.10. Comparison of Harrell's concordance index for overall survival regression with different backbones on the BraTS 2017 dataset for each tumor... 79

Figure 3.11. Comparison of Spearman's rank correlation coefficient for overall survival regression with different training configurations on the BraTS... 80

Figure 3.12. Comparison of Harrell's concordance index for overall survival regression with different training configurations on the BraTS 2017 dataset for... 81

Figure 3.13. Kaplan-Meier survival plot comparing each brain tumor subtype. 87

Figure 4.1. Visual representations of patients with cerebral microbleeds(CMB). The columns represent the input gradient echo (GRE) MRI slice and its... 93

Figure 4.2. An overall block diagram of our pre-training framework for CMB segmentation. Left: Randomly masked patches, denoted as hatched square, are 95

Figure 4.3. Qualitative comparison of our methods with different baselines in the test set. The first column shows an input gradient echo MRI slice in an axial... 102