MM-DETR: Emulating the Diagnostic Clinical Workflow in Multi-view Multi-modal Mammography Mass Detection

What this paper does

A radiologist never reads one mammographic projection in isolation — the craniocaudal and mediolateral oblique views are compared before a finding is called. Most detectors ignore that. MM-DETR adds bidirectional cross-attention between the paired views inside a DETR-style detector, so evidence in one projection can support or suppress a detection in the other. On VinDr-Mammo it reports mass-detection mAP 0.654 against 0.580 for Mammo-CLIP — 7.4 points higher in absolute terms, which the paper states as a 12.8% relative margin — and a 5.9% lower false-negative rate in the dense-breast (density C) cases where single-view reading struggles most. I co-supervised and co-authored this work.

How

A DETR-style detector with a bidirectional cross-attention fusion module that exchanges information between the paired craniocaudal and mediolateral oblique feature maps before the lesion prediction heads.

Reported results

• VinDr-Mammo mass detection mAP 0.654, against 0.580 for Mammo-CLIP: +7.4 points absolute, reported in the paper as a 12.8% relative margin.
• 5.9% reduction in false negatives for density C (heterogeneously dense) cases relative to a single-view baseline.

Relevance

Cross-view fusion is not an architectural flourish here; it mirrors how the reading is actually done, and the improvement is largest in exactly the subgroup where mammographic sensitivity is known to fall.

What it does not show

Single public cohort, mass detection only, and no prospective or reader evaluation.

Citation