EfficientNetV2-Small with PPO fine-tuning yields 98% sensitivity for pediatric fracture classification.

A model combining EfficientNetV2-Small and PPO achieved 95.9% AUC and 98.4% sensitivity for pediatric fracture detection on radiographs (n=1221, single center). Outperformed baseline CNNs. Prototype; needs prospective validation.

• Retrospective evaluation on 1221 pediatric appendicular radiographs from a single hospital, annotated by one pediatric radiologist.
• The model outperformed MobileNetV2, ResNet50, Xception, DenseNet121, and base EfficientNetV2-Small in accuracy, AUC, and sensitivity (exact comparator figures not provided).
• Major limitations: single-center, no external validation, no reported specificity, and reference standard from a single reader; prospective multi-center study required.

RadPigeon summaries are original and for information only. They are not clinical advice.