[MICCAI2026 Workshop Deep-Brea3th] BreastMammo and DenseMammo: Benchmarks for Mammography Domain Generalization

Representative Samples And Density Distributions

Histogram-Based Domain Generalization Pipeline

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.

📄 Paper link:

MICCAI: https://papers.miccai.org/miccai-2026-sat/Deep_Brea3th_021.html

arXiv: https://arxiv.org/abs/2608.10271

📊 Dataset:

BreastMammo: https://osf.io/n4yr2/

DenseMammo: https://osf.io/4azcr/

We provide both 16-bit PNG and DICOM formats. This repository uses the PNG files.

🧠 Pre-trained Weights:

https://huggingface.co/phy710/BreastMammo

📌 Repository Structure

The codebase is organized into modules corresponding to internal dataset benchmarks and external domain generalization evaluations:

├── BreastMammo/
│   ├── density/               # Breast density classification (5-fold CV)
│   └── diagnosis/
│       ├── single/            # Single-view pathology diagnosis (benign vs. malignant)
│       └── two/               # Two-view (CC + MLO) pathology diagnosis
├── DenseMammo/
│   └── density/               # 4-view screening density classification (5-fold CV)
├── external/
│   ├── LUMINA/                # Unseen domain baseline evaluation (No DG)
│   ├── LUMINA-Histogram/      # Proposed foreground-only histogram matching DG
│   ├── LUMINA-DFT/            # Discrete Fourier Transform-based DG baseline
│   └── LUMINA-MixStyle/       # Feature-level MixStyle DG baseline
├── generative/
│   ├── dft.py                 # DFT-based image style synthesis & background masking
│   ├── histogram.py           # Foreground-only histogram matching generation
│   ├── dataset.py             # Data loading for generative alignment
│   └── seed.py                # Reproducibility seed configuration
├── figures/                   # Visualizations for pipeline, samples, and benchmark plots
├── LICENSE
└── README.md

🚀 Usage & Experiments

The trained weights are available at https://huggingface.co/phy710/BreastMammo. If you want to test our trained models, please download them and put the saved folder into the corresponding task.

Internal Benchmarks

Please use folders BreastMammo and DenseMammo for this section.

Training and Testing

In each task, go to the corresponding folder, then run

./main.sh [-model model_name] [-input_size size] [-data_path data_path]

Here, [-input_size] can be 224 or 512, [-model] can be efficientnet_b0, densenet121, resnet50, and swin_t. Other models may be supported but are not tested yet.

For example:

./main.sh -model swin-T -input_size 224 -data_path /dataset/BreastMammo_PNG

You can get the test results by running the command like the following:

python fold_test.py --model --data-path /dataset/BreastMammo_PNG --model swin_t --input-size 224

External Domain Generalization Evaluation

Please use folders "generative" and "external" for this section.

Generative Synthetic Images

You may run histogram.py or dft.py for histogram-based and DFT-based domain generation. You may revise the following code at lines 59--65 in histogram.py and lines 99--105 in dft.py to choose the source and reference dataset.

To generate synthetic BreastMammo images in the DenseMammo domain:

source_root = BreastMammo_root
source = BreastMammo_density(root = BreastMammo_root)
ref = DenseMammo_density(root= DenseMammo_root)

This will generate folders BreastMammo_XXX_YYY, where XXX is Histogram or DFT. YYY is 25, 50, 75, or 100, which stands for α in Eq (1) in our paper, multiplied by 100.

To generate synthetic DenseMammo images in the BreastMammo domain:

source_root = DenseMammo_root
ref = BreastMammo_density(root = BreastMammo_root)
source = DenseMammo_density(root= DenseMammo_root)

This will generate folders DenseMammo_XXX_YYY.

Then you will put these generated folders where you store BreastMammo_PNG and DenseMammo_PNG (i.e., /dataset).

Training and Testing

In each task, go to the corresponding folder, then run

./main.sh -model swin-T -input_size 224 -data_path /dataset/

You can get the test results by running the command like the following:

python fold_test.py --model --data-path /dataset/ --model swin_t --input-size 224

Benchmark

📝 Citation

If you use this dataset in your research, please cite our MICCAI paper:

@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}
}