Real-Time Elbow Fracture Detection on Mobile Devices: Performance and Limitations

This study investigates the feasibility of a smartphone application that utilises the YOLOv11 object detection model to diagnose elbow fractures from X-ray images, motivated by the poor clinician performance in diagnosing these injuries. The investigation included training a YOLOv11 model on a labelled dataset of elbow fractures and deploying it on a mobile application. The application was capable of running inference on images loaded from the photo library, taking photographs for inference and running live inference using an image stream from the phone’s camera. The model achieved an average mAP@50 of 69.3% and an F1 score of 92.7% on radiograph scans, exhibiting overall poor results on both tests. Specifically, F1 scores ranged from 31% to 60.3% in the camera tests and from 28.8% to 43.1% in the live inference tests. The results suggest that for fracture detection to be reliably used with the phone’s camera, a diverse and high-quality dataset that accounts for different viewing conditions is required.