FAD-MIL: a weakly supervised fracture detection model based on X-ray images
Feng Xue1, Yuan Zhang2, Wen Zhao3,4
1The First Affiliated Hospital of Xinjiang Medical University, Urumqi, China.
Scientific Reports
|March 18, 2026
Summary
Fracture detection in radiographs is improved by a new weakly supervised learning method, FAD-MIL. This approach enhances diagnostic accuracy and interpretability, especially in settings with limited resources.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Orthopedics
Background:
- Radiograph interpretation for fractures is inconsistent, particularly in resource-limited settings.
- Weakly supervised learning (WSL) offers a scalable alternative to manual annotation but struggles with subtle fracture details.
Purpose of the Study:
- To introduce FAD-MIL (Fracture-Aware Dual-stream Multiple-Instance Learning) to improve WSL for fracture detection.
- To address limitations of existing WSL methods in capturing focal fracture cues.
Main Methods:
- Developed a global-local dual-stream architecture for comprehensive feature extraction.
- Implemented a fracture-aware gating mechanism to re-weight informative regions.
- Utilized Top-K instance selection to focus on discriminative patterns.
Main Results:
- FAD-MIL achieved an AUC of 0.833 on the FracAtlas dataset.
- Outperformed Mean-Pool MIL and Tile-Vote MIL, performing comparably to ABMIL.
- Provided interpretable instance-level attribution and gradient-based heatmaps.
Conclusions:
- FAD-MIL demonstrates superior performance and interpretability in weakly supervised fracture detection.
- The method shows promise for improving radiograph interpretation in challenging clinical environments.
- Further validation with matched non-fracture controls is needed to assess specificity.
