Improving pediatric hip fracture detection using deep learning: multicenter validation and clinical reader study

To develop and evaluate a deep learning model for automated localization and diagnosis of femoral neck fractures in children under 8 years of age using hip radiographs.

This retrospective multicenter study included 794 hip radiographs from 640 pediatric patients (median age, 4.1 years; 62.5% male) collected from four tertiary hospitals between June 2013 and December 2024. A YOLOv11-based object detection model was trained on 712 radiographs and externally validated on 82 radiographs. Diagnostic performance was measured by area under the receiver operating characteristic curve (AUROC), sensitivity, and specificity. A multi-reader study was conducted using the external test set, where five physicians (two senior radiologists, one junior radiologist, two emergency physicians) interpreted radiographs with and without AI assistance. Statistical analysis included DeLong's test, McNemar tests, and Fleiss' κ.

The model achieved AUROCs of 0.911 (95% CI: 0.864-0.949) on the internal test set and 0.873 (95% CI: 0.792-0.935) on the external test set. Sensitivity and specificity were 84.9% and 85.5% internally, and 80.8% and 91.1% externally. Among junior readers, AI assistance significantly improved diagnostic accuracy (mean ΔAUROC = + 0.083; P = 0.007) and interobserver agreement (κ from 0.49 to 0.61). The model localized fractures in real time with a mean inference time of 56.2 ms.

A YOLOv11-based deep learning model accurately detected femoral neck fractures in children and significantly improved diagnostic accuracy and consistency among less experienced readers. These findings support its integration as a real-time assistive tool in pediatric trauma care.