Robustness evaluation of YOLO-based fracture detection in low-quality radiographs

Detecting fractures on radiographs remains challenging because low contrast, noise, anatomical overlap, and projection may obscure subtle abnormalities. We evaluated CLAHE preprocessing, input resolution, and P2 feature integration in a YOLO-based framework using FracAtlas. Single-run ablations were followed by three-seed robustness and size-stratified analyses and a single-institution evaluation of 993 hip radiographs collected at the authors’ institution. In the matched single-run head ablation, CLAHE-preprocessed YOLOv8m and YOLOv8m-P2 both achieved an mAP@0.5 of approximately 0.68 at 960 × 960 pixels. Across three prespecified seeds, mean mAP@0.5 was 0.59 ± 0.03 and 0.58 ± 0.07, respectively, with no consistent P2 advantage. Recall in the smallest internal bounding-box quartile was identical (0.40 ± 0.05). The fixed primary checkpoint achieved an institutional-cohort mAP@0.5 of 0.72 and recall of 0.73. However, median normalized target area was approximately sixfold larger than in FracAtlas, and the cohorts differed in anatomical scope, class distribution, acquisition, and annotation. These findings are cohort-specific and do not demonstrate improved generalizability. P2 provided no consistent benefit, while the single-run CLAHE and resolution findings remain exploratory.