DWFAU-Net: A Clinically-Guided Dual-Window Fusion Attention U-Net for Coronary Artery Segmentation in CT Angiography

Coronary artery segmentation in coronary computed tomography angiography (CCTA) is challenging because the vessels are small, low in contrast and anatomically variable. We propose a Dual-Window Fusion Attention U-Net (DWFAU-Net) in which two Hounsfield Unit windows, selected on validation patients, are processed by parallel encoders, fused by learnable 1 × 1 convolutions and decoded by a shared attention-gated decoder. Each window input combines the normalized CT image with a Frangi vesselness map and a Gaussian-smoothed image, and training uses a topology-aware BCE, Soft-Dice and clDice loss. On 16 Rotterdam volumes, under a patient-level protocol and without a region of interest, centerline or seed points, DWFAU-Net achieved a Dice Similarity Coefficient (DSC) of 0.691 and an IoU of 0.580 on unseen patients, with a stable DSC of 0.697 ± 0.012 over five trainings. In five-fold patient-level cross-validation, it achieved the highest slice DSC in folds 1–4 and the lowest mean 95th-percentile Hausdorff distance (30.55 pixels), with overall overlap statistically comparable to single-window and concatenation models. Ablations showed a consistent gain from the Frangi channel (volumetric Dice 0.712 vs. 0.684) and identified a clDice weight of 0.5 as the best of the tested weights, preserving patient-level overlap. The model was robust to ±50 HU contrast-enhancement shifts (DSC change ≤ 0.040) and segmented a patient in 0.31 s. A slice-level split gave a DSC of 0.870, showing how strongly the data-splitting protocol affects reported accuracy.