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

Abstract

Detecting fractures in radiographic images remains challenging because of subtle crack patterns, complex anatomical structures, and limited annotation data. We present a memory-augmented teacher-student framework built on SAM2 to transfer anatomy-aware knowledge into a lightweight detector. A memory bank stores anatomical fracture prototypes from the teacher and guides the student with uncertainty-aware distillation, improving robustness and consistency in low-data clinical settings.