BreastDivider: A Large-Scale Dataset and Model for Left–Right Breast MRI Segmentation

---

📰 News

• 04/26 – Released BreastDivider Model Pip package for easier use!
• 08/25 – 📦 Dataset V2 released — now 17,956 cases with left/right as well as partial lesion segmentation masks, and over 3000 lesion classification targets
• 08/25 – 🏆 Used in the winning solution of the ODELIA Breast Cancer Classification Challenge
• 07/25 – Released BreastDivider Model and Dataset for public use
• 07/25 – Accepted to MICCAI WOMEN 2025!

---

---

🧠 Introduction

Breast MRI plays a pivotal role in breast cancer detection, diagnosis, and treatment planning. BreastDivider addresses a critical limitation in breast MRI segmentation: the lack of distinction between the left and right breasts in most public datasets and models.

We introduce the first publicly available large-scale dataset with explicit left and right breast segmentation labels, now comprising over 17,000 3D MRI scans. Alongside, we provide a robust nnU-Net–based segmentation model, trained to reliably separate left and right breast regions in clinical MRI data.

This resource serves as a foundation for anatomically aware AI in breast MRI, enabling improved unilateral classification, treatment response evaluation, and post-mastectomy follow-up. It also supports large-scale pretraining for downstream tasks.

---

📂 Dataset and Model

BreastDivider includes:

• 🔹 17,956 3D breast MRI scans with left/right segmentation masks, curated from 7 public datasets: Duke-Breast-Cancer-MRI, MAMA-MIA, Advanced-MRI-Breast-Lesions, EA1141, ODELIA, ISPY1, ISPY2
• 🔹 Lesion annotations:
  • 3021 lesion classification targets
  • 467 lesion segmentation masks
• 🔹 Pretrained nnU-Net model achieving 0.99 Dice in 5-fold cross-validation
• 🔹 Docker container for seamless deployment and inference

📥 Links:

• Dataset: 🤗 BreastDividerDataset
• Model: 🤗 BreastDividerModel
• Docker: DockerHub

---

📂 Dataset Folder Structure

dataset/
├── imagesTr_batch1/
├── imagesTr_batch2/
├── labelsTr_batch1/
├── labelsTr_batch2/
├── lesion_annotations/
│   ├── classification/
│   └── segmentation/

• imagesTr_batch*: Training images in .nii.gz format (split into two batches)
• labelsTr_batch*: Left/right segmentation masks in .nii.gz format (split into two batches)
• lesion_annotations/classification: classification.csv with lesion labels
• lesion_annotations/segmentation: Lesion masks for bilateral images

---

Install

pip install breastdivider

Python Usage

from breastdivider import predict

predict(
    input_path="case_0000.nii.gz",
    output_path="case_seg.nii.gz",
)

For repeated inference, create and reuse a predictor:

from breastdivider import BreastDividerPredictor

predictor = BreastDividerPredictor(device="cuda")
predictor.predict(
    input_path="case_0000.nii.gz",
    output_path="case_seg.nii.gz",
)

CLI Usage

breastdivider predict case_0000.nii.gz case_seg.nii.gz --device cuda

To pre-download the model:

breastdivider download

Input Format

nnunetv2 expects single-channel files named like CASE_0000.nii.gz.

• If you pass a directory, the package forwards it directly to nnunetv2.
• Directory inputs with arbitrary .nii.gz filenames are automatically staged into nnU-Net's CASE_0000.nii.gz naming scheme before prediction.
• If you pass a single .nii.gz file, the package temporarily stages it under the expected *_0000.nii.gz naming scheme before prediction.

Notes

• The underlying model is hosted at ykirchhoff/BreastDividerModel.
• The inference backend is nnunetv2.
• The package is a wrapper around the published model.

---

📄 Citation

If you use this dataset or model in your work, please cite:

@article{rokuss2025breastdivider,
  title     = {Divide and Conquer: A Large-Scale Dataset and Model for Left–Right Breast MRI Segmentation},
  author    = {Rokuss, Maximilian and Hamm, Benjamin and Kirchhoff, Yannick and Maier-Hein, Klaus},
  journal   = {arXiv preprint arXiv:2507.13830},
  year      = {2025}
}

License

Note that while this repository is available under Apache-2.0 license (see LICENSE), the model checkpoint is Creative Commons Attribution Non Commercial Share Alike 4.0!