Detection of whole body bone fractures based on improved YOLOv7

In the field of medical imaging, early and accurate detection of bone fractures can significantly improve the treatment effect of patients. In this study, we conduct a comparative study of one-stage and two-stage deep learning architectures, with a particular focus on their ability to autonomously and accurately localize four different fracture morphologies in the whole body: angle fractures, normal fractures, line fractures, and messed-up angle fractures. Using well-annotated datasets, we explore the capabilities of the frontier models, with a special focus on the YOLO variants (v4, v5, v7, v8 and our improved v7 model), SSD, Faster-RCNN, and Mask-RCNN. To further improve the detection accuracy, we introduce an Enhanced Intersection of Unions (EIoU) loss function to refine the positional differences between the predicted bounding box and the ground truth bounding box. We measure the performance of the models by precision, recall, mAP, and IoU metrics. Our analysis illuminates the strengths and limitations of each model for bone fracture detection and highlights the advances made by integrating the attention mechanism into YOLOv7. Most notably, our customized YOLOv7-ATT model incorporating the attention mechanism significantly outperforms the baseline metrics of the pre-trained model, achieving a mAP of 80.2%. It exhibits excellent generalization on the FracAtlas dataset, achieving a mAP of 86.2%, which is significantly better than the other models. This study provides researchers with a foundational resource aimed at optimizing and deploying deep learning models for fracture detection in clinical settings.