DCD-VMUNet: A Dynamic Multi-Scale and Geometry-Aware Network for Coronary Vessel and Stenosis Segmentation in X-Ray Angiography

Accurate segmentation of coronary vessels and stenotic regions in X-ray coronary angiography remains challenging because of low vessel-to-background contrast, vessel overlap, thin branches, and the small, ambiguous appearance of stenotic lesions. In this study, we propose DCD-VMUNet, a structure-aware framework for automated coronary vessel and stenosis segmentation. Built on VM-UNet, the model incorporates a dynamic multi-scale kernel module, a centerline head, and a direction head. The dynamic module adaptively integrates fine boundary details with broader contextual cues, while the auxiliary heads introduce geometric supervision to enhance structural continuity and orientation awareness. DCD-VMUNet was independently trained and evaluated for binary vessel and stenosis segmentation on the ARCADE dataset. It was compared with YOLOv11-X, DeepLab v3+, UNet, UNet++, TransUNet, MALUNet, and VM-UNet using pixel-wise metrics and, for vessel segmentation, the centerline-aware clDice metric. For vessel segmentation, DCD-VMUNet achieved the best performance, with IoU, precision, recall, and F1-score of 0.6838, 0.8418, 0.7846, and 0.8122. For stenosis segmentation, it also ranked first across all metrics, reaching 0.4042, 0.5826, 0.5688, and 0.5757, respectively. Ablation studies confirmed the complementary effects of multi-scale feature extraction and geometric supervision. Additional evaluation on the independent ICA dataset further supported the vessel segmentation performance of DCD-VMUNet beyond the ARCADE benchmark. These results indicate that DCD-VMUNet improves coronary vessel delineation and pixel-level stenosis localization, supporting automated coronary angiographic assessment.