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

Scientific summary

MM-DETR models paired mammography views jointly, using cross-view information rather than independent single-image detection.

Abstract

MM-DETR is a multi-view, multi-modal mammography detector that uses paired CC and MLO views to emulate part of the radiologist reading workflow.

Why it matters

Multi-view fusion better reflects clinical mammography interpretation, where lesion evidence is assessed across paired projections.

Contribution

The workshop paper adds bidirectional cross-attention fusion to combine craniocaudal and mediolateral oblique views for mass detection.

Method overview

A DETR-style detector integrates information from paired mammography views through a cross-attention fusion module before lesion prediction.

Key findings

• Springer reports mass detection mAP of 0.654 on VinDR-Mammo.
• The reported result outperforms Mammo-CLIP mAP 0.580 by an absolute 12.8% margin.
• The paper reports a 5.9% lower false-negative rate in DENSITY C cases versus a single-view baseline.

Citation

TODO: Add BibTeX.