A high-precision object detection algorithm based on improved YOLOv5s for medical x-ray physical feature reconstruction

Background. Precise identification of occult fractures in hand x-ray imaging remains challenging due to low physical contrast and overlapping anatomical structures. This often leads to missed fine-grained fracture details, compromising emergency nursing decisions and patient safety. Methods. To address this, we propose an enhanced fracture detection framework based on an improved YOLOv5s architecture tailored for medical physical imaging. First, we integrate a content-aware dynamic upsampling (DySample) operator to replace conventional interpolation, preserving fine-grained edge features through point-wise dynamic sampling. Second, we design the C3-EMSCP module, which leverages efficient multi-scale convolution plus (EMSConvP) to capture complex fracture morphologies across diverse spatial scales. Results. Experimental results on a dataset of 4654 hand radiographs demonstrate that the proposed model achieves a mean average precision (mAP@0.5) of 71.1% and a recall of 66.3%. It outperforms the baseline YOLOv5s by 3.3% in mAP@0.5 and surpasses the state-of-the-art YOLOv11n by 9.1% in recall and 6.8% in mAP@0.5. Significance. By generating an intuitive “Pain Risk Map,” this system provides a high-precision visualization tool for auxiliary diagnosis, effectively bridging the gap between automated detection and clinical decision-making. It enables nurses to accurately localize fracture sites, avoid painful areas during procedures, and implement patient-centered painless care.