Enhanced YOLOv8 with Hybrid Attention Gating Block for Intertrochanteric Fracture Detection in Complex X-Ray Images

Timely and accurate detection of intertrochanteric fractures on pelvic radiographs may support prompt clinical assessment and treatment planning. However, in emergency clinical settings, standard object detection models frequently suffer from high false-alarm rates due to complex anatomical structures and normal cortical bone overlaps. To address this clinical limitation, this application-focused study proposes an enhanced deep learning framework named YOLOv8s-HAGB, tailored specifically for automated intertrochanteric fracture screening. We integrate a customized hybrid attention gating block (HAGB), which combines dilated local feature extraction, pooled contextual modeling, and context-aware spatial gating to modulate feature responses in complex pelvic radiographs. The framework was evaluated on a clinical dataset comprising 694 unilateral radiographs from 694 unique individuals, including 458 participants with confirmed intertrochanteric fractures and 236 fracture-negative controls, using stratified fivefold cross-validation with fixed fold assignments. The complete fivefold cross-validation procedure was independently repeated five times using different random initializations. Within each repetition, the metrics were averaged across the five validation folds, and the final results are reported as the mean ± standard deviation across the five repetition-level averages. Under the repeated fixed-fold cross-validation protocol, YOLOv8s-HAGB achieved a precision of 0.893 ± 0.008 and an mAP@50 of 0.866 ± 0.011, with higher point estimates than the evaluated YOLOv8s, YOLOv9s, YOLOv10s, SE, GAM, and CBAM baselines. These internally validated results support further evaluation of the framework on independent external cohorts and in prospective clinical settings.