Multistain-CycleGAN: Multi-domain stain normalization with a single model

Pytorch implementation of MultiStain-CycleGAN described in "Multi-domain stain normalization for digital pathology: A cycle-consistent adversarial network approach for whole slide images". MultiStain-CycleGAN is capable of many-to-one stain translation.

This repo is based on the works of Jun-Yan Zhu and Taesung Park et al..

Installation

• Clone this repo with:

git clone https://github.com/marjohe/multistain_cyclegan.git

• Install the requirements:
  • Pip users: pip install -r requirements.txt
  • Conda users: conda env create -f environment.yml

Minimal working example of the normalization process:

We ship a working generator trained on the CAMELYON17 dataset as described in the paper. We trained the generator to translate images to the target domain center_0

• Download the model weights from here
• Copy them into resources/models
• Run normalize.py to normalize and plot the example images of every of the 5 centers of the CAMELYON17 dataset.

Train your own model

• Create a root directory containing 2 directories named trainA and trainB with images of different medical centers / slides

• Pass a list of arguments to train.py and run it(--dataroot and --name are required)

  • dataroot: path/to/train_dir containing trainA and trainB
  • name: experiment name

• For visualization of the train process a visdom server is required:

  • To install visdom: pip install visdom
  • To start the server: python -m visdom.server
  • Click http://localhost:8097 to view the results
  • Visdom can be disabled by passing --display_id 0

• For troubleshooting look here: Frequently Asked Questions

Tiling your own slides

We recommend using our publicly available pipeline for tiling whole slide images that can be found here: WSI Preprocessing

Citation

If you use this work for your research please cite:

 @misc{hetz2023multidomain,
      title={Multi-domain stain normalization for digital pathology: A cycle-consistent adversarial network for whole slide images}, 
      author={Martin J. Hetz and Tabea-Clara Bucher and Titus J. Brinker},
      year={2023},
      eprint={2301.09431},
      archivePrefix={arXiv},
      primaryClass={eess.IV}
}