BreastMammo and DenseMammo: Benchmarks for Mammography Domain Generalization

MICCAI 2026 Workshop Deep-Brea3th (Best Paper Award)
Hongyi Pan1, Gorkem Durak1, Halil Ertugrul Aktas1, Andrea Mia Bejar1, Mustafa Ege Seker2,
Nebile Alibeyoglu3, Rumeysa Guclu3, Rana Gunoz Comert Bozkurt3, Sibel Ozkan Gurdal4, Neslihan Cabioglu3,
Beyza Ozcinar3, Ravza Yilmaz3, Vahit Ozmen5, Erkin Aribal6, Sukru Mehmet Erturk3,
Yalda Zafari7, Mohamed Mabrok7, Kayhan Batmanghelich8, Mohammad Yaqub9, Ziyue Xu10, Ulas Bagci1
1Northwestern University  |  2University of Wisconsin-Madison  |  3Istanbul University  |  4Namik Kemal University  |  5Istanbul Florence Nightingale Hospital
6Acibadem Mehmet Ali Aydinlar University  |  7Qatar University  |  8Boston University  |  9MBZUAI  |  10NVIDIA
MICCAI arXiv Code ✿ OSF (BreastMammo) ✿ OSF (DenseMammo) Hugging Face (Weights)

BreastMammo & DenseMammo introduce multi-view mammography benchmarks for breast density assessment and pathology diagnosis across disparate clinical domains. We propose a foreground-only histogram matching protocol that decouples breast tissue style from anatomy, achieving a peak AUC of 98.32% on screening data and significantly outperforming MixStyle and Fourier-based domain generalization frameworks on unseen clinical sites (TNMammo & LUMINA).

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.

Representative Samples And Density Distributions

Representative Samples and Density Distributions

Histogram-Based Domain Generalization Pipeline

Multi-View Domain Generalization and Classification Pipeline

Internal Benchmark

Internal Benchmark Evaluation Results

External Domain Generalization Evaluation

External Generalization Evaluation on TNMammo and LUMINA

BibTeX Citation

@InProceedings{PanHon_BreastMammo_MICCAISAT2026,
        author = { Pan, Hongyi AND Durak, Gorkem AND Aktas, Halil Ertugrul AND Bejar, Andrea M. AND Seker, Mustafa Ege AND Alibeyoglu, Nebile AND Guclu, Rumeysa AND Bozkurt, Rana Gunoz Comert AND Gurdal, Sibel Ozkan AND Cabioglu, Neslihan AND Ozcinar, Beyza AND Yilmaz, Ravza AND Ozmen, Vahit AND Aribal, Erkin AND Erturk, Sukru Mehmet AND Zafari, Yalda AND Mabrok, Mohamed AND Batmanghelich, Kayhan AND Yaqub, Mohammad AND Xu, Ziyue AND Bagci, Ulas},
        title = { { BreastMammo and DenseMammo: Benchmarks for Mammography Domain Generalization } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17256},
        month = {pending},
        page = {pending}
}
}