MeisenMeister

MeisenMeister is a framework for breast cancer classification on DCE-MRI. It is designed to support reproducible multi-stage workflows from dataset fingerprinting and experiment planning to preprocessing, training, benchmarking, and ROI-level inference.

This repository contains trained MeisenMeister model weights for bilateral breast MRI classification into healthy, benign, and malignant classes.

The weights were trained on ODELIA and AMBL data.

🏆 The MeisenMeister framework was used for the winning solution of the MICCAI 2025 ODELIA Breast MRI Challenge on Grand Challenge:

https://odelia2025.grand-challenge.org/

Download From Hugging Face

Download the repository with:

hf download Bubenpo/MeisenMeister --local-dir ./MeisenMeister

Then use the downloaded model path:

./MeisenMeister

If you use the default Hugging Face cache instead, the model will usually be stored under a path like:

~/.cache/huggingface/hub/models--Bubenpo--MeisenMeister/snapshots/<snapshot-id>

Installation

git clone https://github.com/MIC-DKFZ/MeisenMeister.git
cd MeisenMeister
conda create -n meisenmeister python=3.12 -y
conda activate meisenmeister
pip install -e .

Contents

• dataset.json
• mmPlans.json
• fold_0 to fold_6 with model_best.pt
• optionally fold_all/model_best.pt for faster single-checkpoint inference

Input Format

Each case must have 3 channels:

• _0000.nii.gz = pre
• _0001.nii.gz = post1
• _0002.nii.gz = post2

Example:

case001_0000.nii.gz
case001_0001.nii.gz
case001_0002.nii.gz

Usage

This model repository is intended to be used with the mm_predict_from_modelfolder CLI command.

Full ensemble mode:

mm_predict_from_modelfolder \
  ./MeisenMeister \
  -i /path/to/input_images \
  -o /path/to/output \
  -f 0 1 2 3 4 5 6 \
  --checkpoint best

Faster mode using only fold_all:

mm_predict_from_modelfolder \
  ./MeisenMeister \
  -i /path/to/input_images \
  -o /path/to/output \
  -f all \
  --checkpoint best

License

The MeisenMeister source code is licensed under the Apache License 2.0.

Model weights are licensed under CC BY-NC-SA 4.0 due to downstream dataset licensing constraints from the data used for training.

Citation

If you use MeisenMeister in research, please cite:

Hamm, B., Kirchhoff, Y., Rokuss, M., and Maier-Hein, K., MeisenMeister: A Simple Two Stage Pipeline for Breast Cancer Classification on MRI, arXiv:2510.27326 [cs.CV], 2025.

Paper:

https://arxiv.org/pdf/2510.27326

@article{hamm2025meisenmeister,
  title={MeisenMeister: A Simple Two Stage Pipeline for Breast Cancer Classification on MRI},
  author={Hamm, Benjamin and Kirchhoff, Yannick and Rokuss, Maximilian and Maier-Hein, Klaus},
  journal={arXiv preprint arXiv:2510.27326},
  year={2025}
}

This model also relies on the BreastDivider dataset and segmentation work:

Dataset:

https://huggingface.co/datasets/Bubenpo/BreastDividerDataset

@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}
}