Deep learning model classifies pediatric femoral neck fractures on x-rays with high accuracy

A YOLO-based deep learning model with wavelet attention achieved AUC 0.94–0.99 for classifying pediatric femoral neck fractures on internal testing. On external testing, 31 orthopedic surgeons’ accuracy improved to 79–87% with model assistance (P<0.001).

• The model was developed and validated on 5,555 proximal femoral growth plates, 1,306 femoral neck fractures (various Delbet-Colonna types), and 257 subtrochanteric fractures.
• With AI assistance, inter-reader agreement (Fleiss kappa) improved across all experience levels, with the largest gain in residents (+27.4%).
• Limitation: Retrospective, single-center design; external reading study may not reflect real-world multi-institutional generalizability.

RadPigeon summaries are original and for information only. They are not clinical advice.