Improving Coronary Artery Stenosis Segmentation Performance of YOLO Models Through a Dataset Expansion Strategy Based on Local Focused Augmentation.

To address the problems of the small size of coronary artery stenosis regions, strong background interference, and limited high-quality annotated data in X-ray coronary angiography images, a dataset expansion strategy based on local focused augmentation is proposed. The method uses existing annotations of stenotic regions to crop local images centered on the lesion regions and adds them to the original training set, thereby increasing the proportion of lesion regions in the training samples without requiring additional manual annotation. Experiments were conducted using the YOLOv8 and YOLO11 model series, and the applicability of the proposed method was evaluated under different learning rates and batch sizes. The results show that local focused augmentation improves the F1 score under most parameter settings. Further validation on the public ARCADE dataset showed that the precision, recall, and F1 scores of YOLOv8m-seg, YOLOv9c-seg, YOLOv11m-seg, and YOLO26m-seg were all improved, with the F1 score of YOLOv8m-seg increasing from 0.3868 to 0.4045. The proposed method can enhance the learning of local features of coronary artery stenosis and improve the segmentation performance of small stenotic regions without modifying the network architecture.