Title Page
영문요약
Contents
1. Introduction 10
1.1. The clinical significance of flatfoot 10
1.2. BMI as a risk factor for flatfoot 13
1.3. Prevalence of flatfoot 14
1.4. Interobserver variance of flatfoot diagnosis using radiograph 14
1.5. Prevalence and geographical distribution in South Korea 14
1.5.1. Nationwide prevalence in South Korea 14
1.5.2. Geographic prevalence in South Korea 15
1.5.3. Fluctuation of geographic prevalence in South Korea 17
1.6. Necessity of research 17
1.7. Related works 17
1.8. Purpose 18
2. Materials and Methods 19
2.1. Ethics approval and consent to participate 19
2.2. Dataset 19
2.2.1. Internal dataset 19
2.2.2. External dataset 24
2.2.3. Number of datasets 24
2.3. Ground truth labeling 30
2.3.1. Definition of landmarks and measuring real angles 30
2.3.2. Definition of semantic segmentation and evaluation methods 33
2.4. Deep learning architecture 35
2.4.1. Deep learning architecture for landmark detection 35
2.4.2. Performance comparison of deep learning models for landmark detection 38
2.4.3. Deep learning architecture for semantic segmentation 39
2.4.4. Regression 41
2.5. Testing of automated angle measurements and observer studies 41
2.5.1. Landmark detection 41
2.5.2. Semantic segmentation 43
2.6. Statistical analysis 43
3. Results 45
3.1. Landmark detection with human observers 45
3.1.1. Flatfoot angle measurements of human observers and the deep learning model (DLM) 45
3.1.2. Diagnostic accuracy 47
3.1.3. Agreement between ground truth and deep learning model (Bland–Altman plot) 49
3.1.4. Intra-observer reliability 51
3.1.5. Inter-observer reliability 51
3.1.6. External validation 53
3.2. Semantic segmentation with human observers 54
3.1.1. Flatfoot angle measurements of human observers and the deep learning model (DLM) 54
3.2.2. Diagnostic accuracy 56
3.2.3. Time taken for flatfoot angle measurement among human observers (GT, OS, GP1 and GP2) 59
3.2.3. External validation 59
3.3. Regression 60
4. Discussion 63
5. Conclusion 69
6. Reference 70
국문요약 74