Rehabilitation Driven Optimized YOLOv11 Model for Medical X-Ray Fracture Detection

Abstract

Highlights: Main Findings A medical X-ray fracture detection model with precise localization based on YOLOv11n is proposed to solve the problems of false localization and poor accuracy in existing models. The improved model, trained with an expanded dataset using data augmentation and enhanced with a Bone-MSCA module and Focal-SIoU loss function, outperforms other mainstream single-object detection models, with significant improvements in detection accuracy, recall rate, F1-Score and mean Average Precision 50. Implication of the Main Findings The proposed model can provide more accurate and reliable fracture detection from X-ray images, which is beneficial for timely and appropriate medical treatment. The techniques used in this research, such as data augmentation, the Bone-MSCA module and Focal-SIoU loss function, can be a reference for other medical image detection research, potentially improving the performance of related models. Accurately identifying fractures from X-ray images is crucial for timely and appropriate medical treatment. However, existing models suffer from problems of false localization and poor accuracy. Therefore, this research proposes a medical X-ray fracture detection model with precise localization based on the You Only Look Once version 11 nano (YOLOv11n) model. Firstly, a data augmentation technique combining random rotation, translation, flipping and content recognition padding is designed to expand the public dataset, alleviating the overfitting risk due to scarce medical imaging data. Secondly, a Bone-Multi-Scale Convolutional Attention (Bone-MSCA) module, designed by combining multi-directional convolution, deformable convolution, edge enhancement and channel attention, is introduced into the backbone network. It can capture fracture area features, explore multi-scale features and enhance attention to spatial details. Finally, the Focal mechanism is combined with Smoothed Intersection over Union (Focal-SIoU) as the loss function to enhance sensitivity to small fracture areas by adjusting sample weights and optimizing direction perception. Experimental results show that the improved model trained with the expanded dataset outperforms other mainstream single-object detection models. Compared with YOLOv11n, its detection accuracy, recall rate, F1-Score and mean Average Precision 50 increase by 4.33%, 0.92%, 2.52% and 1.24%, respectively, reaching 93.56%, 86.29%, 89.78% and 92.88%. Visualization of the results verifies its high accuracy and positioning ability in medical X-ray fracture detection.