Transformer-UNet 단계적 네트워크를 통한 조영 증강 CT 영상 기반 대동맥 박리 자동 분할 연구 = Automatic Segmentation of Aortic Dissection in Contrast-Enhanced CT Images Using Transformer-UNet Cascade Network

다국어 초록 (Multilingual Abstract)

Aortic Dissection (AD) is a severe condition caused by a tear in the aortic inner wall, allowing blood to flow between the layers of the aortic wall and potentially leading to life-threatening complications. Managing AD involves imaging, medical treatment, and sometimes surgery. CT scans, which produce high-resolution images quickly, are commonly used for AD diagnosis and prognosis evaluation. Accurate segmentation of the True Lumen (TL), False Lumen (FL), and Thrombosis (TH) is crucial, but manual measurement is time-consuming and variable. To address this, computer vision, machine learning, and deep learning methods have been introduced. Although CNN-based models have played a significant role in medical image analysis, they have limitations in comprehensively understanding anatomical structures. To overcome these limitations, Transformer-based models have been introduced, excelling in extracting global context information but being less effective in capturing local texture details. Therefore, this study proposes a model that combines the strengths of CNNs and Transformers. We designed a two-stage model: the first stage uses a 3D Transformer UNet to learn the aorta’s global information, while the second stage uses a 3D UNet to learn the detailed textures of TL, FL, and TH. Additionally, a multi-scale patch extraction method is applied to effectively capture both the aorta’s global information and detailed textures. This model's two-step approach—using a 3D Transformer UNet for global context and a 3D CNN UNet for local texture—has been validated in ablation studies. The model's performance was evaluated using the dataset from Asan Medical Center and compared with existing models such as nnUNet and nnFormer. Our method achieved Dice Similarity Coefficients (DSC) of 0.917, 0.888, and 0.630 for TL, FL, and TH, respectively, demonstrating the highest segmentation accuracy. The model's robustness and generalizability were further assessed using external datasets, showing potential for improving AD diagnosis and treatment across various clinical settings.

목차 (Table of Contents)

• Abstract ⅰ
• Contents ⅱ
• Contents of Tables ⅲ
• Contents of Figures ⅳ
• 1. Introduction 1
• 2. Materials and Method 5
• 2.1. Datasets 5
• 2.2. CNNs vs Transformer 6
• 2.3. Method 7
• 3. Experiments 16
• 3.1. Performance Comparison with Baseline Models 16
• 3.2. Performance Comparison with Previous Studies 18
• 3.3. Ablation Study 20
• 3.4. Hyperparameters and Development Environments 20
• 4. Discussion 22
• 5. Conclusion 23
• References 24
• Abstract (with Korean) 26
• Contents of Tables
• Table 1. Dataset description 4
• Table 2. Quantitative evaluation compared to State-Of-The-Art models 16
• Table 3. Quantitative evaluation compared to State-Of-The-Art models (confusion volume metric) 16
• Table 4. Indirect quantitative evaluation compared to previous studies on AD segmentation 18
• Table 5. Ablation study 19
• Table 6. Hyperparameters of the Stage 1 and 2 models 21
• Table 7. Development environments 21
• Contents of Figures
• Figure 1. Overview of the proposed method architecture 6
• Figure 2. The stage 1 model architecture and modules 9
• Figure 3. Qualitative evaluation compared to State-Of-The-Art models 15
• Figure 4. The sample case illustrating limitation 22