Artificial Intelligence Enhanced Chest CT Modeling: Airway Labeling, Disease Classification, Image Enhancement, and Synthesis = 인공 지능 강화 흉부 CT 모델링: 기도 라벨링, 질병 분류, 이미지 향상 및 합성

다국어 초록 (Multilingual Abstract)

Medical image analysis plays a vital role in modern healthcare by enabling the early identification, diagnosis, and management of various diseases. This thesis investigates advanced machine learning methods for medical image classification and generation, with a particular focus on quantitative computed tomography (QCT) images. The overarching goal is to enhance the accuracy, interpretability, and accessibility of medical imaging analysis, thereby supporting more reliable clinical decision-making. The first contribution introduces BranchLabelNet, a dividing-and-grouping multi- label classification framework for automated airway branch labeling. By systematically dividing-and-grouping related labels, BranchLabelNet improves classification accuracy while mitigating challenges posed by inter-label dependencies. Experimental results demonstrate its superiority over conventional multi-label methods, particularly in anatomically complex regions. The second contribution explores the use of Kolmogorov- Arnold networks (KANs) for binary classification of QCT data. Leveraging learnable activation functions, KAN achieves state-of-the-art classification accuracy while offering improved interpretability compared to standard neural networks. Performance evaluations across metrics such as accuracy, precision, recall, F1-score, specificity, and Matthews correlation coefficient confirm its robustness and clinical relevance. The third contribution investigates a generative adversarial network (GAN)-based super-resolution framework, multi-level densely connected super-resolution network with GAN (mDCSRN-GAN), for enhancing the resolution of lung CT images. The proposed model successfully reconstructs high-resolution images from low-resolution inputs, preserving key structural and functional features. This enables reliable quantitative analysis while potentially reducing radiation exposure, demonstrating the practical value of super- resolution methods in pulmonary imaging. The final contribution presents an initial exploration of diffusion models for synthesizing functional residual capacity (FRC) CT images from total lung capacity (TLC) scans. An adapted self-consistent recursive diffusion bridge (SelfRDB) framework, augmented with multi-scale feature extraction, spatial attention, and composite loss functions, achieved strong quantitative performance and preserved global anatomical structures. While challenges remain in capturing fine- grained details, this work establishes a foundation for future advancements in diffusion- based thoracic image translation. Collectively, the findings of this thesis highlight the potential of advanced machine learning techniques, including multi-label classification, interpretable neural networks, super-resolution, and generative modeling, to advance QCT analysis. These approaches contribute toward improved disease detection, risk assessment, and clinical decision support, while also paving the way for future research in machine learning-driven medical imaging.

목차 (Table of Contents)

• CHAPTER 1 INTRODUCTION 1
• 1.1 Background 1
• 1.1.1 Respiratory diseases 1
• 1.1.2 Quantitative computed tomography 2
• 1.1.3 Airway branch labeling approaches 3
• 1.1.4 Quantitative computed tomography-based imaging for respiratory risk assessment 4
• 1.1.5 Advanced neural network architectures 5
• 1.1.6 Generative models 8
• 1.2 Thesis objective, overview, and proposed research 11
• CHAPTER 2 A MULTI-LABEL CLASSIFICATION METHOD FOR HUMAN AIRWAY LABELING USING A DIVIDE-AND-GROUP STRATEGY 16
• 2.1 Introduction 16
• 2.2 Materials and methodologies 17
• 2.2.1 Generation of training and testing dataset 19
• 2.2.2 Addressing class imbalance with Tomek Links 21
• 2.2.3 Classification architecture 23
• 2.3 Results 30
• 2.3.1 Generation of n-ary trees 30
• 2.3.2 A model for one-step multiclass classification 30
• 2.3.3 Applying Tomek Links under-sampling to address class imbalance 31
• 2.3.4 Hierarchical binary/ternary classification using an ensemble approach 32
• 2.3.5 Post-label adjustment process to prevent duplicate branch names 33
• 2.3.6 Compared with pre-trained model 34
• 2.4 Discussion 35
• 2.5 Conclusion 38
• CHAPTER 3 KOLMOGOROV-ARNOLD NETWORKS FOR CEMENT DUST- EXPOSED PATIENTS CLASSIFICATION USING COMPUTED TOMOGRAPHY IMAGES 40
• 3.1 Introduction 40
• 3.2 Materials and methodologies 42
• 3.2.1 Dataset description and features extraction 44
• 3.2.2 Kolmogorov-Arnold networks 47
• 3.2.3 Quantitative computed tomography imaging data with Kolmogorov-Arnold network 49
• 3.2.4 Evaluation metrics 51
• 3.2.5 Analysis of feature importance 52
• 3.3 Results 52
• 3.3.1 Performance of the Kolmogorov-Arnold network 52
• 3.3.2 Comparison of model performance and efficiency 53
• 3.3.3 Reliability, explainability, and feature contributions 56
• 3.4 Discussion 61
• 3.5 Conclusion 66
• CHAPTER 4 AN APPLICATION OF A GAN-BASED SUPER-RESOLUTION MODEL FOR LUNG CT IMAGING 67
• 4.1 Introduction 67
• 4.2 Methods 68
• 4.2.1 Data acquisition. 69
• 4.2.2 Three-dimensional multi-level super-resolution network architecture 70
• 4.2.3 Quantitative computed tomography-based airway structural and parenchymal functional variables 74
• 4.2.4 Statistical analysis 75
• 4.3 Results 76
• 4.3.1 Three-dimensional lung computed tomography imaging super-resolutionmodel 76
• 4.3.2 Statistical analysis results. 80
• 4.4 Discussion 88
• 4.5 Conclusion 91
• CHAPTER 5 AN INITIAL EXPLORATION OF A DIFFUSION MODEL FOR FUNCTIONAL RESIDUAL CAPACITY CT IMAGE SYNTHESIS 92
• 5.1 Introduction 92
• 5.2 Methods 94
• 5.2.1 Self-consistent recursive diffusion bridge model 97
• 5.2.2 Adaptation of self-consistent recursive diffusion bridge model for lung image translation 98
• 5.2.3 Bayesian hyperparameter optimization 109
• 5.2.4 Dataset and evaluation metrics 111
• 5.3 Results 113
• 5.3.1 Bayesian optimization results 113
• 5.3.2 Model performance 114
• 5.3.3 Model comparison. 116
• 5.4 Discussion 118
• 5.5 Conclusion 119
• CHAPTER 6 SUMMARY 121
• 6.1 Summary 121
• 6.2 Future studies 121
• 6.3 Publications 122
• REFERENCES 124
• APPENDICES 145
• Appendix A. Supplementary materials of BranchLabelNet 145
• A.1 Supporting tables 145
• A.2 Supporting figures 147
• Appendix B. Supplementary materials of KAN model 153
• Appendix C. Chest CT image processing overview 153
• Appendix D. SHAP-based feature importance analysis 158