Performance analysis of deep learning based aortic segmentation from 4D flow MRI across different architectures and training strategies

다국어 초록 (Multilingual Abstract)

Deep learning based aortic segmentation has become an essential component for enabling fully automated hemodynamic analysis using 4D flow MRI. However, its clinical application remains limited due to variability in segmentation quality, differences in data distributions across institutions, and challenges in accurately capturing vascular boundaries in low resolution and noise sensitive images. To address these limitations, this study systematically evaluates three representative segmentation network architectures, including UNet, diffusion models, and transformer models, and further introduces two coarse to fine hybrid frameworks, C2D-Diff3D and C2B-Trans3D, which are designed to improve boundary accuracy and overall robustness. A dataset of 550 subjects, consisting of 78 healthy subjects and 472 patients with various cardiovascular conditions, was used for training and evaluation. Model performance was assessed on an independent test set using dice similarity coefficient(DSC), intersection over union(IoU), accuracy, and hausdorff distance(HD). Among the evaluated architectures, the C2D-Diff3D framework achieved the most consistent and accurate segmentation across all aortic regions, outperforming both the UNet and transformer baselines. Additional experiments on external institutional datasets revealed performance degradation caused by differences in image contrast and intensity characteristics. This issue was effectively reduced through a transfer learning approach that incorporated a small number of institution specific training samples. In summary, this study presents a comprehensive comparison of multiple segmentation architectures and proposes practical hybrid and domain adaptation strategies that enhance the clinical applicability of automated aortic segmentation. The findings emphasize the value of multi stage refinement and adaptation to institutional data characteristics in achieving reliable and generalizable segmentation for 4D flow MRI.

목차 (Table of Contents)

• I. Introduction 1
• 1.1 Research background 1
• 1.1.1 Clinical significance of hemodynamic analysis 1
• 1.1.2 Background of 4D flow MRI 2
• 1.1.3 Current workflow hemodynamic analysis in 4D flow MRI 4
• 1.1.4 Data acquisition preprocessing 4
• 1.1.5 Postprocessing 5
• 1.2 Research objectives 6
• 1.2.1 Limitation and need for fully automated analysis 6
• II. Methods 9
• 2.1 Dataset and preprocessing 9
• 2.1.1 Dataset composition 9
• 2.1.2 Aortic partitioning segmentation and image preprocessing pipeline 10
• 2.2 Base segmentation models 11
• 2.2.1 3D UNet baseline model 11
• 2.3 Two-stage hybrid training framework 20
• 2.3.1 Corase-to-Diffusion model (C2D-Diff3D) 20
• 2.3.2 Corase-to-Bottlenek Transformer model (C2B-Trans3D) 23
• 2.4 Continuous adaptation framework 26
• 2.4.1 Transfer learning for domain shift 26
• III. Results 28
• 3.1 Quantitative segmentation metrics on test dataset 28
• 3.1.1 UNet 29
• 3.1.2 Diffusion 31
• 3.1.3 Transformer 33
• 3.1.4 C2D-Diff3D 35
• 3.1.5 C2B-Trans3D 37
• 3.2 Unseen data test on best model 39
• 3.3 Evaluation after Transfer learning 41
• IV. Discussion 42
• 4.1 Automatic segmentation for clinical environment 42
• 4.2 Limitations 42
• 4.3 Future works 43
• V. Conclusion 44
• References 45
• Summary in Korean 48