FlowSDF: Flow Matching for Medical Image Segmentation Using Distance Transforms

International Journal of Computer Vision (IJCV), 2025

This repository contains the official implementation of the paper: FlowSDF: Flow Matching for Medical Image Segmentation Using Distance Transforms

Lea Bogensperger, Dominik Narnhofer, Alexander Falk, Konrad Schindler, Thomas Pock

🔧 Getting Started

📁 Data Preprocessing

Preprocess your data following the structure used in generative-segmentation-sdf, particularly the notebook /preprocess_data/precompute_sdf.ipynb. Ensure that precomputed SDF masks and images are saved in a directory containing train.csv and test.csv.

Update the data_path in the config file under the general section to point to this directory.

🚀 Running the Code

Clone the repository (including submodules):

git clone --recurse-submodules git@github.com:leabogensperger/FlowSDF.git

Use main.py to start training via trainer.py. Use sampler.py to generate SDF-based segmentation masks from images listed in test.csv. Make sure to specify the model and checkpoint to be used in the config script.

# 1. Create and activate the environment
conda create -n flowsdf-env python=3.11
conda activate flowsdf-env
pip install -r requirements.txt

# 2. Run the training code for MoNuSeg
python main.py --config "cfg/monuseg.yaml",

# 3. Sampling from this trained model 
python sampler.py --config "cfg/monuseg.yaml"

Citation

If you find our work helpful, please consider citing:

@article{bogensperger2025flowsdf,
  title={FlowSDF: Flow matching for medical image segmentation using distance transforms},
  author={Bogensperger, Lea and Narnhofer, Dominik and Falk, Alexander and Schindler, Konrad and Pock, Thomas},
  journal={International Journal of Computer Vision},
  pages={1--13},
  year={2025},
  publisher={Springer}
}