Title Page
Abstract
List of Abbreviations
Contents
Ⅰ. INTRODUCTION 12
1.1. Study background 12
1.2. Study Motivation 16
1.3. Main contributions 16
Ⅱ. RELATED WORKS 17
2.1. Convolutional neural network 17
2.2. Gradient explanation techniques 19
Ⅲ. PROPOSED METHOD 21
3.1. Explainable Deep Learning approach 21
3.2. Dataset 23
3.3. Data augmentation 24
3.4. Classification of type of the disease based on CNN 25
3.5. Visualization of classification with XAI 26
Ⅳ. IMPLEMENTATION AND EXPERIMENTAL RESULTS 29
4.1. Implementation Environment 29
4.2. Experiment and analysis 30
4.2.1. Experimental results 30
4.2.2. Quantitative results 31
Ⅴ. CONCLUSION AND FUTURE DIRECTION 35
5.1. Conclusion 35
5.2. Future direction 36
References 37
Table 4.1. Implementation Environment 29
Table 4.2. Detailed information on training results of CNN models 32
Table 4.3. Data augmentation method performance 34
Figure1.1. Normal condition and abnormal condition of the endoscopic image with three classes. 13
Figure2.1. General architecture of CNN network 18
Figure2.2. Visual explanation of Grad-Cam output 19
Figure3.1. The structure of the main method 22
Figure3.2. Dataset classes: normal z-line, b normal cecum, c normal polorus, d) esophagitis, e) dyed, and lifter polyps f) dyed dissection margins g) polyps and... 23
Figure3.3. Data augmentation methods and the result of augmentation 25
Figure4.1. The output of the explainable heat map for dyed-p and polyp. 30
Figure4.2. Train accuracy of CNN models for the backbone of Explainable AI 31
Figure4.3. Data augmentation Training results 33