Distal radius fracture detection: a comparison of computer vision algorithms

Distal radius fractures (DRFs) are among the most common skeletal injuries encountered in emergency and orthopedic settings, accounting for a significant portion of wrist-related trauma cases across all age groups. Despite their prevalence, subtle fractures are frequently missed, particularly by less experienced clinicians, due to variations in anatomical presentation and radiographic complexity. Early and accurate detection is essential for ensuring proper treatment and avoiding long-term functional impairment. This study presents a deep learning-based system for the automatic detection and localization of DRFs in wrist radiographs, developed with the goal of supporting both clinical workflows and medical education. Accordingly, several state-of-the-art object detection and segmentation models are evaluated, including Faster R-CNN, YOLOv8, YOLOv11, Mask R-CNN, and EfficientNet. Each model is trained and validated using labeled datasets, with performance compared across precision, recall, and F1-score metrics. Based on these findings, EfficientNet emerged as the best-performing model and is subsequently deployed within a custom-designed Python application. The prototype enables users to upload radiographs, view segmentation-based predictions, and submit diagnostic feedback, making it suitable for use in both clinical and training environments. To validate the system's diagnostic potential, an independent evaluation is conducted using 40 radiographs annotated by orthopedic surgeons. The EfficientNet model is compared against five sixth-year medical interns, each of whom manually annotated the same cases. The results show that the model outperforms the interns across all major metrics, achieving a precision of 0.89, recall of 0.98, and F1-score of 0.93. It also correctly identifies all fractures in lateral-view images, which is a view type that posed notable challenges for the human participants, while producing results in under one second per image.