Anatomy-Aware Distillation with Memory-Augmented SAM2 for Fracture Detection

Abstract

Fracture detection in radiographic imaging remains challenging due to subtle fracture patterns, complex anatomical structures, and limited annotated data. This study proposes a memory-augmented knowledge distillation framework that combines a SAM2 teacher model with a lightweight student detector. A memory bank stores anatomy-aware fracture embeddings, and the student learns consistent anatomical reasoning through alignment with relevant prototypes. We further reduce false positives by selectively transferring high-confidence knowledge using uncertainty-based weighting. Experiments on KHATL and FracAtlas show improved detection accuracy and robustness, demonstrating effectiveness for medical fracture detection.