Comprehensive Summary
This study introduces a new deep learning model based on YOLOv5 (image detection framework) for the classification of distal radius fracture (DRF) radiographs into three standard views: anterior/posterior (AP), oblique (OB), and lateral (LAT). A set of 1,593 DRF radiographs (544 AP, 521 OB, and 540 LAT) from 540 patients was created. All radiographs were labelled using Labellerr (an online annotation software) and verified by an orthopedic surgeon for accuracy. The YOLOv5s model was selected due to its lightweight nature and its ability to make quick, reliable predictions. The model was trained from scratch using the radiograph set and focused on key factors such as precision, recall, and F1 score. The model showed an overall accuracy of 97.3%, with the strongest performance in the LAT view (100%) and the weakest in the OB view (93%). Precision was also consistently high at 96.8%. The final verified model was developed into a mobile application available for physician use.
Outcomes and Implications
This tool can provide essential support to triaging physicians by assisting in initial classification of radiographs. This may improve efficiency in clinical workflows and could reduce time between diagnosis and treatment. The model’s accessibility also demonstrates its relevance in primary care and rural settings, where specialized providers may be limited. Further refinement of the oblique view and integration into more clinical settings is needed to confirm the model’s generalizability. Future work will explore the use of this model in more settings, such as fracture management and support in decision-making.