conference · 2025 · Published
Exemplar Med-DETR: Toward Generalized and Robust Lesion Detection in Mammogram Images and Beyond
What this paper does
Detectors trained on one mammography cohort tend to collapse on another, especially in dense breasts. Exemplar Med-DETR learns class-specific exemplar features inside the model and fuses them with image features through cross-attention, so detection is anchored to prototypes rather than to raw appearance. On the Vietnamese VinDr-Mammo cohort it reports mAP 0.70 for mass and 0.55 for calcification detection — a 16 percentage-point absolute improvement over the previous state of the art — and roughly a twofold gain on an out-of-distribution Chinese cohort. It also transfers to chest X-ray and angiography, improving by 4 and 7 points. I contributed as a co-author.
How
Class-specific exemplar features are derived within the detector and fused with image features via cross-attention, with iterative training that refines the exemplars alongside the detection heads.
Reported results
- VinDr-Mammo: mAP 0.70 for mass detection and 0.55 for calcifications, an absolute improvement of 16 percentage points over the previous state of the art.
- Out-of-distribution CMMD (Chinese) cohort: approximately a twofold gain in lesion detection performance.
- Beyond mammography: mAP 0.25 for mass detection in chest X-ray and 0.37 for stenosis detection in angiography, improvements of 4 and 7 percentage points respectively.
Relevance
Detection performance that does not survive a change of population or scanner is not clinically usable. This paper treats cross-cohort robustness as the headline result rather than a supplementary table.
What it does not show
Absolute mAP on the harder tasks (calcifications, chest X-ray masses) remains modest, and evaluation is retrospective.
Citation
S. Bhat, B. Georgescu, A. Bhandary Panambur, M. Zinnen, T.-T. Nguyen, A. Mansoor, K. K. Elbarbary, S. Bayer, F.-C. Ghesu, S. Grbic, A. Maier. Exemplar Med-DETR: Toward Generalized and Robust Lesion Detection in Mammogram Images and Beyond. MICCAI 2025, pp. 205–215. doi:10.1007/978-3-032-04978-0_20
View original abstract
Lesion detection in medical images has to survive dense anatomy, domain shift between populations, and differences between imaging modalities. Exemplar Med-DETR derives class-specific exemplar features inside the detector and fuses them with image features through cross-attention, trained iteratively, so that detection is anchored to learned prototypes rather than to appearance alone. The method is evaluated on mammography for mass and calcification detection, on an out-of-distribution external cohort, and on chest X-ray and angiography.
