Rethinking Class Reweighting and Data Augmentation for Long-Tailed Detection in Pediatric Wrist Radiographs

Pediatric wrist X-ray analysis faces a severe long-tailed class imbalance challenge: the GRAZPEDWRI-DX dataset exhibits a 2389:1 ratio between the most common (text, n=19,112) and rarest (foreignbody, n=8) classes. While class-balanced loss reweighting and class-aware data augmentation have been extensively studied for long-tailed image classification, their effectiveness in dense object detection under such extreme imbalance remains underexplored. We present a systematic evaluation of two strategies on GRAZPEDWRI-DX across YOLOv8s and YOLOv11s: (1) inverse-frequency-power class reweighting (cls_pw) on the BCE loss, and (2) Class-Aware Mosaic (CA-Mosaic), a novel augmentation enforcing rare-class exposure during mosaic composition. Twenty-five experiments are conducted, including a 3-seed reproducibility study. Four key findings emerge. First, mild reweighting (cls_pw=0.1) consistently improves test mAP@0.5 from 0.597 to 0.623 on YOLOv8s and 0.599 to 0.622 on YOLOv11s, while aggressive reweighting (cls_pw≥0.75) degrades performance. Second, CA-Mosaic shows strong intensity-dependent, class-selective benefits: bonelesion detection improves stably from 0.124 to 0.412 mAP@0.5 (+232%) at CA-200 on YOLOv11s; boneanomaly improves +44% on average (3-seed) at CA-50, with seed variance reflecting the small test-set size (22 instances). Third, single-strategy CA-Mosaic at high intensity matches combined cls_pw+CA-Mosaic for rare-class detection (both 0.412 for bonelesion), while combined strategies generally degrade overall performance through head-class over-suppression. Fourth, these results suggest that single, well-tuned data-level interventions are preferable to stacked loss+data strategies in extreme long-tailed medical detection. We release code, trained models, and complete experimental logs to facilitate further research.