Title Page
Abstract
Contents
Chapter 1. Introduction 24
Chapter 2. Chest X-ray Generation by Matching-pair Synthesis 29
2.1. Introduction 29
2.2. Related Works 32
2.2.1. Generative Adversarial Networks (GANs) 32
2.2.2. Anomaly Detection 33
2.2.3. Thoracic Disease Detection 34
2.3. Method 34
2.3.1. Matching-Pair Synthesis 36
2.3.2. Abnormal-to-Normal Translation 39
2.3.3. Residual-Saliency Abnormality Detection 40
2.3.4. Normal-to-Abnormal Translation for Data Augmentation 42
2.4. Experiments 42
2.4.1. Datasets 42
2.4.2. Implementation Details 43
2.4.3. Evaluation Metrics 46
2.4.4. Comparative Evaluation 47
2.4.5. Discussion 50
2.5. Conclusion 55
Chapter 3. Chest X-ray Generation via Disentangled Contrastive Learning 56
3.1. Introduction 56
3.2. Related Work 59
3.2.1. Latent Space Disentanglement 60
3.2.2. Latent Space Encoding 60
3.2.3. CXR Image Generation 61
3.3. Proposed Method 63
3.3.1. Matching-Pair Synthesis 63
3.3.2. Disentangled Contrastive Learning 65
3.3.3. Learning Disentangled Generation 68
3.3.4. Controlled Generation 70
3.4. Experiments 70
3.4.1. Datasets 70
3.4.2. Implementation Details 71
3.4.3. Comparative Study 78
3.4.4. Expert Evaluation 80
3.4.5. Abnormality Detection 84
3.4.6. Abnormality Classification 84
3.5. Discussion 86
3.5.1. Effect of Auxiliary Classifier 87
3.5.2. Disentangled Latent Embeddings 89
3.5.3. Limitations 90
3.6. Conclusion 90
Chapter 4. Chest X-ray Generation with Anatomical Priors 91
4.1. Introduction 91
4.2. Related Work 95
4.2.1. Anomaly Detection 95
4.2.2. Thoracic Disease Detection 95
4.2.3. Generative Models 96
4.3. Method 97
4.3.1. Matching-Pair Synthesis 97
4.3.2. Generation with Anatomical Priors 98
4.3.3. Residue-Guided Detection 101
4.4. Experiments 104
4.4.1. Datasets 104
4.4.2. Training Details 105
4.4.3. Comparative Study 106
4.5. Discussion 111
4.5.1. Ablation Study 111
4.5.2. Anatomical Priors with Diffusion Models 111
4.5.3. Limitations and Future Work 115
4.6. Conclusion 117
Chapter 5. Conclusion 118
5.1. Summary 118
5.2. Future Work 119
5.2.1. Normal Chest X-ray Generation without Region-level Annotations 120
5.2.2. Robust Anatomical Semantic Estimation 120
Bibliography 121
요약 137
Table 2.1. Comparison of pneumonia detection results using different train- ing augmentation configurations in terms of average precision (mAP) used in... 50
Table 2.2. Comparison of performance with various detection frameworks in terms of conventional average precision (AP). AP: mean average... 51
Table 2.3. Evaluation on vinbigDataset in terms of mean average precision (mAP). +residual denotes a detector model with residual-saliency guidance. 51
Table 2.4. Statistical analysis on RSNA dataset. 53
Table 2.5. Evaluation of different training strategy configurations on RSNA. Gen. represents a translator model and Det. represents a detection model. 53
Table 3.1. Evaluation of the synthesized image quality. 80
Table 3.2. Comparison of pneumonia detection. 85
Table 3.3. Comparison of thoracic disease detection. 86
Table 3.4. Evaluation of the generated image quality by radiologists. 86
Table 3.5. Abnormality classification on different datasets. 87
Table 4.1. Evaluation of the synthesized image quality. 109
Table 4.2. Evaluation of residual difference (RD) scores. 111
Table 4.3. Comparison of pneumonia detection. 112
Table 4.4. Comparison of thoracic disease detection. 113
Table 4.5. Ablation study on different guidance components. 113
Table 4.6. GANs against diffusion models. 116
Figure 1.1. Challenges in analyzing chest radiography. 25
Figure 1.2. The research direction of this dissertation. Starting from image conditioned CXR generation by leveraging synthetic matching-pairs in Chapter... 26
Figure 2.1. Manual diagnosis vs. automatic diagnosis. 30
Figure 2.2. Proposed framework overview. 35
Figure 2.3. Matching-pair CXR Synthesis. 37
Figure 2.4. Residual-Saliency Abnormality Detection. 40
Figure 2.5. Generated radio-realistic chest X-ray images. 48
Figure 2.6. Constructed matching-pair CXRs. 49
Figure 2.7. Effects of residual-saliency on detection. 52
Figure 2.8. Limitations. 54
Figure 3.1. This paper presents a GAN framework that disentangles the latent vector z into two distinct attribute vectors: disease-specific and person-specific... 58
Figure 3.2. Overview of our framework. 62
Figure 3.3. Framework for learning the disentangled latent space and controllable generator. 66
Figure 3.4. Relations between different pairs of CXRs for contrastive learning. 67
Figure 3.5. Architecture of the generator and the discriminator for matching-pair synthesis. 71
Figure 3.6. Architecture of multi-scale encoder. 72
Figure 3.7. Examples of generated matching-pairs. 74
Figure 3.8. Controlled abnormal-to-normal synthesis. 75
Figure 3.9. Controlled abnormal-to-normal synthesis. 76
Figure 3.10. Controlled abnormal-to-normal synthesis. 77
Figure 3.11. Diversity of targeted CXR generations. 81
Figure 3.12. Diversity of targeted CXR generations. 82
Figure 3.13. Diversity of targeted CXR generations. 83
Figure 3.14. Study on the effect of the auxiliary classifier. 88
Figure 3.15. t-SNE embeddings of disentangled representations. 89
Figure 4.1. AntGAN generates normal CXR images from abnormal CXRs by conditioning on anatomical priors, aligning them with the anatomical structures... 93
Figure 4.2. Anatomy-conditioned generation. 99
Figure 4.3. Residue-guided detection. 103
Figure 4.4. Residual difference (RD) score. 107
Figure 4.5. Comparison of CXR generation. 110
Figure 4.6. Comparison with Diffusion model. 115