Artificial intelligence-based method for detecting wrist fractures in children

Comprehensive Summary

This study evaluates whether an AI object detection model can improve the detection of pediatric wrist fractures on Xray images. Using a retrospective dataset of over 20,000 annotated pediatric wrist Xrays from the publicly available GRAZPEDWRI-DX database, the authors developed an enhanced deep learning model Kid-YOLO based on the YOLO11 architecture. The model incorporates architectural modifications and a specialized loss function to better detect subtle and rare fracture patterns. Compared with multiple established object detection models, Kid-YOLO achieved higher precision, recall, and mean average precision, with notable improvements in detecting less common fracture types. The authors note that while performance improved, the study is limited by use of a single dataset and the lack of external or prospective clinical validation.

Outcomes and Implications

This work is important because pediatric wrist fractures are common and can be difficult to diagnose accurately, especially when fracture lines are subtle or imaging resources are limited. Improved AI detection tools can help support clinicians in identifying fractures that are at greater risk of being missed on an initial review. Clinically, the model can be a tool that assists medical professionals rather than a replacement for physician interpretation, especially in resource limited settings. However, the model was trained on a single dataset and does not have explainable AI methods to clarify how predictions are made.