Title Page
Abstract
Contents
Introduction 10
Previous studies 13
Systematic review of previous literature 13
Automatic detection of TMJOA 14
Methods 16
Dataset 17
Manual segmentation and preprocessing 17
Evaluation 20
Results 24
Accuracy evaluation based on DSC 24
Visualization and comparison of 3D models 25
Discussion 28
Conclusion 30
References 31
Appendix 36
Korean abstract 42
Table 1. The Dice similarity coefficients for 26 subjects (test datasets) using 3D U-Net and cascaded 3D U-Net. 24
Table 2. The absolute mean differences and the number of corresponding points for mandibular condyles of 26 subjects (test datasets) using 3D U-Net and cascaded 3D U-Net. 27
Figure 1. Overall study procedure and data distribution. 16
Figure 2. Preprocessing procedure for 3D U-Net and cascaded 3D U-Net. (a) Normalization of the image (b) cut into right and left sides (c) right-side images flipped horizontally. 19
Figure 3. The CBCT image, ground truth (green box) and the predicted labels (blue box) obtained using 3D U-Net for each slice. 22
Figure 4. The CBCT image, ground truth, and the predicted labels obtained using cascaded 3D U-Net for each slice. 23
Figure 5. Difference map between ground truth and prediction in 3D U-Net and cascaded 3D U-Net viewed from medial side. Area of over-segmentation is indicated by red color and under-segmentation by green color.... 26