Accurate detection of pediatric fractures in radiographs remains challenging due to subtle visual cues and the high prevalence of false-positive detections produced by automated systems. To address this limitation, we propose a lightweight region-of-interest (Region of Interest) adjudication framework that operates as a second-stage verification module to refine detector-generated candidates. The proposed framework integrates iterative hard-negative mining with confidence-aware score fusion to suppress anatomically confounding regions such as growth plates and overlapping structures. Unlike end-to-end detection approaches, the method is designed to function as a modular post-detection refinement stage, enabling improved decision reliability without modifying the underlying detector architecture. Each candidate Region of Interest is evaluated using a compact adjudication network conditioned on detector confidence, and final predictions are obtained through a calibrated fusion strategy. The framework is evaluated on the publicly available GRAZPEDWRI-DX pediatric radiograph dataset using patient-level disjoint training, validation, and held-out test splits to ensure unbiased performance estimation. Experimental results demonstrate that the proposed approach reduces false-positive detections while maintaining high sensitivity. At the selected operating point, the method achieves an F1-score of 0.88 and mAP@0.5 of 0.887, outperforming the detector-only baseline under identical evaluation conditions. In addition, gradient-based activation mapping (Grad-CAM) is employed to provide Region of Interest-level visual explanations, supporting interpretability of adjudication decisions. The proposed framework maintains low computational overhead, making it suitable for integration into real-world clinical workflows as a decision-support component.
Aravinda et al. (Mon,) studied this question.