Accurate segmentation of coronary arteries from X-ray angiographic images is crucial for the diagnosis and interventional planning of coronary heart disease. Although deep learning has achieved remarkable progress in coronary vessel segmentation, existing methods still suffer from limited annotated data, overfitting in small-sample scenarios, and traditional loss functions that fail to balance class imbalance and vascular topological continuity. In this paper, we propose a U-Net model integrating contrastive learning and multi-loss optimization (CLML-UNet) for coronary artery segmentation. By leveraging contrastive learning on unlabeled images, the model effectively learns robust feature representations, thereby reducing dependence on limited annotated data. Furthermore, a composite loss function combining Dice Loss, Focal Loss, and clDice Loss is designed to jointly optimize class balance, hard sample mining, and vascular topology preservation during training. Experiments on a public dataset show that CLML-UNet outperforms eight mainstream baselines on AP, IoU, Dice, clDice, and HD95, with the largest gains on clDice and HD95, the two metrics most relevant to vessel continuity and boundary accuracy. The results indicate that CLML-UNet captures both main trunks and fine branches reliably, offering a practical tool for clinical coronary analysis and an effective solution for medical image segmentation under limited annotations.
Chen et al. (Sat,) studied this question.