Design Rules for Robust Coronary Artery Segmentation: A Systematic Analysis of Dataset Size, Windowing,...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Cardiovascular Imaging

Background:

• Automated segmentation of coronary arteries from computed tomography coronary angiography (CTCA) is crucial for quantitative analysis.
• Segmentation performance is known to be sensitive to dataset characteristics and model configurations.

Purpose of the Study:

• To systematically evaluate the impact of dataset size, CT window width/level (W/L) values, and nnU-Net model architectures on coronary artery segmentation accuracy.
• To assess the robustness of segmentation performance across varying vessel geometries, including curvature and tortuosity.

Main Methods:

• Utilized 1000 annotated CTCA scans from the ImageCAS dataset for training and a 200-scan independent set for testing.
• Investigated the effects of varying training dataset sizes, different CT W/L values (e.g., 800/200 HU, 1300/350 HU), and compared 2D vs. 3D nnU-Net architectures (Full-Resolution, Low-Resolution, Cascade, ensemble).
• Assessed segmentation performance using Dice Similarity Coefficient (DSC) and Intersection over Union (IoU), and evaluated consistency across vessels with different geometric complexities.

Main Results:

• Segmentation accuracy increased with dataset size, plateauing beyond 100 training cases.
• 3D nnU-Net models significantly outperformed the 2D model, with the 3D ensemble achieving the highest accuracy (DSC 0.8337, IoU 0.7178).
• Segmentation performance was consistent across vessels with varying curvature and tortuosity, demonstrating robustness.

Conclusions:

• The nnU-Net framework provides accurate and generalizable coronary artery segmentation.
• Optimizing input parameters (W/L values) and employing 3D ensemble architectures can further enhance segmentation accuracy and robustness.