Abstract
Coronary artery disease (CAD) is a highly prevalent cardiovascular disease and one of the leading causes of death worldwide. The accurate segmentation of coronary arteries from CT angiography (CTA) images is essential for the diagnosis and treatment of coronary artery disease. However, due to small vessel diameters, large morphological variations, low contrast, and motion artifacts, conventional segmentation methods, including classical image processing (such as region growing and level sets) and early deep learning models with limited receptive fields, are unsatisfactory. We propose SADiff, a hybrid framework that integrates a dilated attention network (DAN) for ROI extraction, a diffusion-based subnet for noise suppression in low-contrast regions, and a striped attention network (SAN) to refine tubular structures affected by morphological variations. Experiments on the public ImageCAS dataset show that it has a Dice score of 83.48% and a Hausdorff distance of 19.43 mm, which is 6.57% higher than U-Net3D in terms of Dice. The cross-dataset validation on the private ImageLaPP dataset verifies its generalizability with a Dice score of 79.42%. This comprehensive evaluation demonstrates that SADiff provides a more efficient and versatile method for coronary segmentation and shows great potential for improving the diagnosis and treatment of CAD.