Exemplar Med-DETR: Toward Generalized and Robust Lesion Detection in Mammogram Images and Beyond

Scientific summary

Exemplar Med-DETR uses learned exemplar representations to improve lesion localization under dense anatomy, domain shift, and cross-modality variation.

Abstract

Exemplar Med-DETR is a multi-modal contrastive detector for robust lesion detection across mammography, chest X-ray, and angiography.

Why it matters

The study addresses a central translational barrier in medical detection: maintaining lesion-level performance across populations, scanners, and imaging modalities.

Contribution

The paper introduces exemplar-guided feature-based detection with cross-attention and iterative training.

Method overview

Class-specific exemplar features are derived inside the model and fused with image features through cross-attention for detection.

Key findings

• The arXiv abstract reports mAP50 0.70 for mammography mass detection and 0.55 for calcifications.
• It reports a 16 percentage-point absolute improvement over previous state of the art on Vietnamese dense breast mammograms.
• A radiologist-supported evaluation on 100 out-of-distribution Chinese mammograms showed a twofold lesion-detection gain.

Citation

TODO: Add BibTeX.