HoechstGAN

Project page | Paper

This repository contains the official code for the paper:

HoechstGAN: Virtual Lymphocyte Staining Using Generative Adversarial Networks
Georg Wölflein, In Hwa Um, David J. Harrison and Ognjen Arandjelović
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Jan 2023.

If you find this code useful, please consider citing:

@inproceedings{hoechstgan,
  title = {HoechstGAN: Virtual Lymphocyte Staining Using Generative Adversarial Networks},
  author = {W\"{o}lflein, Georg and Um, In Hwa and Harrison, David J and Arandjelovi\'{c}, Ognjen},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  month = {January},
  year = {2023},
  pages = {4997--5007}
}

@article{hoechstgan_dataset,
  title = {Whole-Slide Images and Patches of Clear Cell Renal Cell Carcinoma Tissue Sections Counterstained with {Hoechst} 33342, {CD3}, and {CD8} Using Multiple Immunofluorescence},
  author = {W\"{o}lflein, Georg and Um, In Hwa and Harrison, David J and Arandjelovi\'{c}, Ognjen},
  journal = {Data},
  volume = {8},
  year = {2023},
  month = {February},
  number = {2},
  article-number = {40}
}

Data

We have made our dataset publicly available on BioImage Archive, alongside an accompanying article in the Data journal that contains a detailed description of the dataset.

Installing

We provide a Dockerfile as well as a docker-compose.yml file that builds the Docker container and mounts the code (i.e. this repository) as a volume.

docker-compose up -d --build

Then, in the hoechstgan container, run

./install.sh

Code

The structure of the code is inspired by the pytorch implementation of pix2pix, but has been heavily modified. We use hydra for configuration management and wandb for tracking experiments.

Most of the relevant implementation is concentrated in the following files: hoechstgan.py, networks.py, and composites.py.

Training

The table below lists the commands used to train each of the models evaluated in Table 2 of the paper.

Model	Command
HoechstGAN-MCD	poetry run python train.py +experiment=hoechstgan_composite discriminator.type=joint generator.composites.0.train.schedule=sigmoid generator.composites.0.train.args.from_epoch=8 generator.composites.0.train.args.to_epoch=12 dataset.max_size=250000 norm=instance
HoechstGAN-MC	poetry run python train.py +experiment=hoechstgan_composite generator.composites.0.train.schedule=sigmoid generator.composites.0.train.args.from_epoch=8 generator.composites.0.train.args.to_epoch=12 dataset.max_size=250000
HoechstGAN-MD	poetry run python train.py +experiment=hoechstgan discriminator.type=joint dataset.max_size=250000
HoechstGAN-M	poetry run python train.py +experiment=hoechstgan dataset.max_size=250000
HoechstGAN-D	poetry run python train.py +experiment=hoechstgan_basic discriminator.type=joint dataset.max_size=250000
pix2pix	poetry run python train.py +experiment=cy3 dataset.max_size=250000 and poetry run python train.py +experiment=cy5 dataset.max_size=250000
Regression-MCD	poetry run python train.py +experiment=hoechstgan_composite discriminator.type=joint generator.composites.0.train.schedule=sigmoid generator.composites.0.train.args.from_epoch=8 generator.composites.0.train.args.to_epoch=12 dataset.max_size=250000 norm=instance gan=regression

Some other useful command-line options are: gpus, dataset.num_threads, and dataset.batch_size. All of the options are described in the config.yaml file.

Testing

To test a model, obtain its wandb ID and run

poetry run python test.py {wandb_id}