Task Specialized Bone Fracture Detection with Attention Models and Knowledge Distillation
Presented Research PaperDeveloped a Task-Specialized Knowledge Distillation (TSKD) framework for automated bone fracture detection from X-ray images, published at IEEE ICICV 2026. Trained three anatomically specialized teacher models for hand, leg, and hip fractures, then distilled their combined expertise into a single lightweight student model with 95% fewer parameters. Incorporated a dual-kernel multi-head CNN with a spatial attention mechanism to capture fracture patterns at multiple scales while suppressing irrelevant background noise. The framework achieved 93.81% test accuracy, outperforming conventional single-teacher and ensemble distillation baselines.
