Deep Learning for Shoulder Fracture Detection
A deep learning-based ensemble system for automated shoulder fracture detection in clinical radiographs, addressing the need for rapid and accurate fracture identification in emergency and orthopedic settings.
Key findings
- Multi-model system trained on 10,000 annotated shoulder X-rays
- Combines Faster R-CNN (ResNet50-FPN, ResNeXt), EfficientDet and RF-DETR, fused with Soft-NMS, WBF and NMW ensembling
- The NMW ensemble reached 95.5% accuracy and an F1-score of 0.9610, outperforming every individual model
- Designed for binary fracture screening and triage, not detailed orthopedic classification
Plain-English summary
- The question
- Shoulder fractures are often missed in busy emergency and high-volume settings — the paper cites studies reporting up to 10% missed. Can an ensemble of detection models reliably flag them on X-ray?
- The data
- 10,000 annotated shoulder X-rays.
- The method
- Several object-detection models — Faster R-CNN (ResNet50-FPN and ResNeXt backbones), EfficientDet and RF-DETR — combined with box- and classification-level ensembling (Soft-NMS, WBF and NMW fusion).
- The result
- The NMW ensemble achieved 95.5% accuracy and an F1-score of 0.9610, beating each individual model on all key metrics, with strong recall and localization precision. The model is limited to fracture/no-fracture detection.
Why it matters for hospitals and radiologists
If up to 10% of shoulder fractures can be missed, a screening model with strong recall gives the reading radiologist a second look on every shoulder film. The authors are explicit that it is built for triage support, not to replace detailed orthopedic classification.
Related at 5C: Bionic, the AI co-radiologist
Citation
Hemanth Kumar M, Karthika M, Saianiruth M, Vasanthakumar Venugopal, Anandakumar D, Revathi Ezhumalai, Charulatha K, Kishore Kumar J, Dayana G, Kalyan Sivasailam, Bargava Subramanian. A Deep Learning-Based Ensemble System for Automated Shoulder Fracture Detection in Clinical Radiographs. arXiv preprint arXiv:2507.13408, 2025. https://doi.org/10.48550/arXiv.2507.13408