A lightweight model of bioresorbable vascular scaffold detection and segmentation in optical coherence tomography images

Bioresorbable vascular scaffold (BVS) is a promising alternative to metallic coronary stents, addressing the long-term risks associated with permanent implants. However, their clinical utility depends on rigorous monitoring of deployment accuracy and degradation dynamics. Intravascular optical coherence tomography (IVOCT) is the only modality suitable for this task, but current automated analysis approaches remain imperfect. This paper presents BVS-YOLO, a geometry-aware and lightweight deep learning model specifically developed for all-stage BVS strut detection and segmentation in IVOCT images. The model introduces a modified P2 detection layer to enhance sensitivity to small-scale strut structures and adopts a new Inner-SIoU loss to improve localization precision for circumferential, rotationally symmetric patterns. During the validation, BVS-YOLO outperforms existing models, including YOLOv5, YOLOv8, U-Net, and Mask R-CNN, in both detection and segmentation tasks. It achieves higher Dice and mAP50-95 with only 2.1M parameters, thereby improving post-procedural assessment efficiency and reducing the diagnostic burden on clinicians.