EM-DETR foundation detector trained across heterogeneous medical datasets
Original title: Toward Foundation Detection Models via Example-Based Grounding in EM-DETR
Authors report about 1 AUC point higher on four CXR findings and 2–3 mAP50 points over single-dataset EM-DETR.
The paper proposes a multi-domain lesion-detection model that combines medical datasets annotated for anatomy and abnormalities. A student-teacher strategy merges heterogeneous label spaces to limit catastrophic forgetting. Building on example-based EM-DETR, training is joint across domain-specific sets. The authors report higher image-level AUC on four chest X-ray findings and leading mAP50 on chest X-rays and mammography.
Unconfirmed: The state-of-the-art and gain figures are the authors' reported results.
- Image-level AUC exceeded prior results by 1 point on four CXR findings.
- Joint training beat single-dataset EM-DETR by 2 to 3 mAP50 points.
- Student-teacher training combines datasets that use different label spaces.
From the original article
1 min read at arxiv.orgDeepLearning (DL)-based Computer-Aided Diagnostic (CAD) systems have shown great promise in supporting clinicians, but their performance heavily depends on the availability of large, well-annotated datasets. In medical lesion detection, the scarcity of comprehensive, stan dardized annotations and the heterogeneity of existing datasets limit the development of robust foundation models.
By Sheethal Bhat · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.