Paper Overview
Research Area: Computer Vision Authors: Sheethal Bhat, Bogdan Georgescu, Awais Mansoor, Florin C. Ghesu, Sasa Grbic, Andreas Maier Published: 2026-11-10 arXiv: 2611.14518
Abstract (translated from Chinese summary)
Deep learning-based computer-aided diagnostic (CAD) systems show great promise, but their performance heavily depends on large-scale, finely annotated data. In medical lesion detection, comprehensive and standardized annotations are scarce and existing datasets are heterogeneous, limiting the development of robust foundation models. This paper proposes a novel approach to building a multi-domain lesion detection foundation model by integrating datasets with heterogeneous annotation spaces.
The authors first propose a student-teacher training strategy to effectively combine heterogeneous label spaces, mitigating catastrophic forgetting and improving feature learning. They then adapt their previously proposed example-based EM-DETR framework so it can jointly learn from multiple domain-specific datasets, enabling fine-grained, example-driven detection across modalities.
Key Results
- Surpasses prior state-of-the-art by over 1 percentage point in image-level AUC on four key chest X-ray signs
- Achieves leading mAP50 on both chest X-ray (CXR) and mammography
- Outperforms single-dataset EM-DETR by 2-3 percentage points
Conclusion
This approach paves the way for scalable foundation models in medical imaging that leverage heterogeneous data while maintaining robust, generalizable performance.
---
*Auto-collected on 2026-11-12*