Unifying Adult and Pediatric Domains: Multi-Dataset Deep Learning for Fracture Classification, Detection, and Segmentation

Le résumé fourni par la source

Fracture detection from radiographs is critical across both adult and pediatric populations, yet existing public datasets are isolated by age group and anatomical site. This study systematically evaluates deep learning (DL) models (ResNet, YOLOv8, YOLOv11, Faster R-CNN, and DeepLabV3) for fracture classification, detection, and segmentation using (i) FracAtlas, an adult multi-organ fracture dataset, and (ii) GRAZPEDWRI-DX, a pediatric wrist dataset. Training models independently on FracAtlas, followed by a unified training pipeline that harmonizes label spaces across both datasets, enabling models to consume adult and pediatric radiographs as a single input stream. Using unified training improves crossdomain generalization, reduces overfitting, and yields consistent gains in classification AUC (+3.8%), detection mAP@0.5 (+4.2%), and segmentation Dice (+2.7%) on held-out test sets. Beyond performance, multi-dataset training narrows the gap between pediatric and adult domains, providing more reliable predictions in clinically challenging cases. Moreover, the study also suggest that integrating heterogeneous datasets offers a scalable pathway toward foundation models for musculoskeletal imaging.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.