BreastMammo and DenseMammo: Benchmarks for Mammography Domain Generalization

2026-08-10

AI Summary

This study addresses the severe domain shift in mammography caused by equipment from different vendors, which significantly hinders the cross-site generalization of AI models. To tackle this issue, the authors introduce two new datasets, BreastMammo and DenseMammo, and propose a foreground-specific histogram matching protocol tailored for mammographic images. Integrated with a Swin Transformer backbone, this approach establishes the first domain generalization benchmark for breast density classification. Evaluated via five-fold cross-validation and external testing on datasets such as TNMammo and LUMINA, the method achieves an internal AUC of 98.32%, substantially outperforming existing techniques like MixStyle and discrete Fourier transform–based methods. The results demonstrate its effectiveness in mitigating domain shifts arising from clinical source variations and enhancing model robustness across domains.

Abstract

Breast density classification is a critical component of breast cancer risk assessment, yet AI models often struggle to generalize across clinical sites due to vendor-specific acquisition styles. In this work, we introduce two new datasets, BreastMammo and DenseMammo, to facilitate robust multi-view mammography research. We propose a domain generalization framework that utilizes a foreground-only histogram matching protocol to resolve the domain shift issue arising from disparate clinical sources. Internal evaluation using a 5-fold cross-validation protocol demonstrates the efficacy of our approach, with the Swin Transformer backbone achieving a peak AUC of 98.32% for density classification. External evaluation on the TNMammo and LUMINA datasets demonstrates that the proposed approach consistently reduces domain shift, significantly outperforming prominent domain generalization paradigms, including MixStyle and Discrete-Fourier-Transform-based frameworks.

Problem

• domain generalization
• mammography
• breast density classification
• domain shift
• clinical site variation

Innovation

• domain generalization
• mammography
• histogram matching
• breast density classification
• Swin Transformer