Comparative Analysis of YOLOv11 with Previous YOLO in the Detection of Human Bone Fractures

Accurate and rapid detection of bone fractures is an important challenge in the medical world, particularly in the field of radiology. This study aims to analyze and compare the performance of the YOLOv11 model with several previous versions of YOLO, namely YOLOv5, YOLOv8, and YOLOv10 in the task of detecting human bone fractures on X-ray and MRI images. The dataset used is the Human Bone Fractures Multi-modal Image Dataset (HBFMID) which consists of 641 raw images (510 X-rays and 131 MRIs). The four models were trained using the HBFMID dataset that had gone through a manual augmentation and annotation process, then tested using evaluation metrics such as precision, recall, mAP50, and mAP50-95. The training results showed that YOLOv11 has the most stable and consistent loss curve, with a fast convergence process. In terms of evaluation, YOLOv11 recorded a precision of 99.87%, a recall of 100%, a 99.49% mAP50, and an 84.13% increase in the number of mAP-95s, which generally outperformed other models. In addition, the visual prediction results show that YOLOv11 can detect fracture areas with the right bounding box and a balanced confidence score, without showing symptoms of overconfidence or inconsistency. When compared to approaches from previous studies, YOLOv11 also showed a significant improvement in detection accuracy. Thus, YOLOv11 is rated as the most optimal and reliable model in deep learning-based automatic bone fracture detection. This model has great potential to be applied in medical diagnosis support systems to improve the efficiency and accuracy of digital fracture identification.