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

Automated coronary artery segmentation from computed tomography coronary angiography (CTCA) can support quantitative analysis, yet performance varies with data characteristics and model configuration. This study systematically evaluates how dataset size, computed tomography (CT) window width and level (W/L) values, model architectures, and vessel geometry affect nnU-Net-based coronary artery segmentation.

Using 1000 annotated CTCA scans from the publicly available ImageCAS dataset, we conducted a series of experiments in which we varied the number of training cases, applied different W/L values, compared 2D and multiple 3D nnU-Net architectures (Full-Resolution, Low-Resolution, Cascade, and ensemble), and assessed performance across vessel curvature and tortuosity. All models were evaluated on an independent 200-case test set using Dice similarity coefficient (DSC) and intersection over union (IoU).

Segmentation performance increased with dataset size and plateaued beyond 100 training cases. W/L values of 800/200 Hounsfield units (HU) and 1300/350 HU slightly improved performance over original images. 3D models outperformed the 2D model, with the 3D ensemble achieving the highest accuracy (DSC 0.8337; 95% confidence interval [CI]: 0.8269-0.8405, IoU 0.7178; 95% CI: 0.7080-0.7275). Performance remained consistent across vessels with different curvature and tortuosity, indicating model robustness to geometric complexity.

The nnU-Net framework achieves accurate and generalizable coronary artery segmentation across diverse dataset sizes, W/L values, model architectures, and vessel geometries. Optimizing input settings and architectural strategies further enhances segmentation accuracy.