<![CDATA[Bone fracture detection in X-ray images remains challenging because fracture lines often appear as low-contrast, subtle, and visually similar patterns across fracture types. Although YOLO-based detectors have been widely used for medical object detection, the effect of contrast enhancement is commonly evaluated on a single architecture, making it unclear whether preprocessing benefits are consistent across different YOLO generations. This study investigates the architecture-dependent effect of Contrast Limited Adaptive Histogram Equalization (CLAHE) on YOLOv5, YOLOv8, YOLOv10, and YOLOv11 for bone fracture detection. A total of 1,539 annotated X-ray images were prepared in YOLO bounding-box format and evaluated under two scenarios: original images and CLAHE-enhanced images. Model performance was assessed using precision, recall, mAP50, and mAP50-95, followed by paired architecture-level comparison using paired t-test, Wilcoxon signed-rank test, and bootstrap confidence intervals. The results show that CLAHE does not uniformly improve all detection metrics. YOLOv8 without CLAHE achieved the strongest mAP50 and recall, whereas YOLOv11 with CLAHE produced the highest mAP50-95, indicating better localization precision under stricter IoU thresholds. The statistical comparison suggests a positive but exploratory improvement in mAP50-95 after CLAHE, while other metrics showed no significant architecture-level difference. These findings demonstrate that image enhancement effectiveness is architecture-dependent and should be selected according to the feature extraction and localization characteristics of the detector rather than applied as a universal preprocessing step.]]>
Copyrights © 2026
