VEA-SegUNet: Edge-Enhanced Multi-Scale Network with F2-Optimization for Robust Coronary Artery Segmentation

Accurate diagnosis of coronary artery disease (CAD) requires precise vessel segmentation in coronary angiography. However, challenges such as low contrast, imaging artifacts, and vessel overlap often result in segmentation discontinuities and hinder the identification of small vessels. To overcome these issues, we propose VEA-SegUNet, a novel vessel segmentation method for X-ray coronary angiography. Our approach incorporates three key innovations. First, the vessel enhancement module (VEA) leverages unsupervised edge detection priors to emphasize vessel boundaries. Second, a multi-scale deformable convolutional attention module is embedded within the U-Net encoder to capture complex vascular structures at different scales, thereby improving vessel continuity and small vessel detection. Third, F2-score optimization prioritizes vessel connectivity in segmentation topology. We validate the F2-score as a superior metric to the traditional F1-score for coronary segmentation and incorporate F2 loss into the framework. Extensive experiments on the DCA1, CHUAC, and XCA datasets demonstrate that VEA-SegUNet outperforms six state-of-the-art U-shaped architectures on several metrics. It achieves an F1-score of 79.1%, an F2-score of 85.1%, a recall of 89.3%, an IoU of 65.1%, an accuracy of 97.7%, and an AUC of 98.8%, while maintaining a very competitive computational efficiency. These results confirm the effectiveness and practicality of VEA-SegUNet for coronary artery segmentation.