Leveraging Pre-trained Models for FF-to-FFPE Histopathological Image Translation

Paper | Models

Prerequisites

• Python >= 3.8
• NVIDIA GPU + CUDA CuDNN
• Torch 2.0.1 + Torchvision 0.15.2

Getting started

• Clone this repository:

git clone git@github.com:QilaiZhang/Diffusion-FFPE.git
cd Diffusion-FFPE
pip install -r requirements.txt

• Install CONCH and place it in ./checkpoints
• Prepare FF and FFPE datasets following AI-FFPE, and split the datasets into training, validation, and test sets.

Training

• Train the Diffusion-FFPE model:

python train.py --train_source_folder [TRAIN_FF_FOLDER] --train_target_folder [TRAIN_FFPE_FOLDER] --valid_source_folder [VALID_FF_FOLDER] --valid_target_folder [VALID_FFPE_FOLDER]

• Resume training from latest checkpoints:

python train.py --train_source_folder [TRAIN_FF_FOLDER] --train_target_folder [TRAIN_FFPE_FOLDER] --valid_source_folder [VALID_FF_FOLDER] --valid_target_folder [VALID_FFPE_FOLDER] --ckpt_path [CHECKPOINTS_FOLDER] --resume

Inference

• Download pre-trained checkpoints and place it in ./checkpoints.
• Generate FFPE images from FF test dataset:

python inference.py --img_path [TEST_FF_FOLDER] --pretrained_path ./checkpoints/model.pkl

Evaluation

• Compute statistics for FFPE test datasets:

python eval.py --data_path [TEST_FFPE_FOLDER] --ref_path [TEST_STATISTICS_PATH] --save-stats

• Compute FID and KID:

python eval.py --data_path [GENERATE_FFPE_FOLDER] --ref_path [TEST_STATISTICS_PATH] --fid --kid

Visualization

References

If our work is useful for your research, please consider citing:

@inproceedings{zhang2024leveraging,
      author={Qilai Zhang and Jiawen Li and Peiran Liao and Jiali Hu and Tian Guan and Anjia Han and Yonghong He},
      title={Leveraging Pre-trained Models for FF-to-FFPE Histopathological Image Translation}, 
      booktitle={arXiv preprint arXiv:2406.18054},
      year={2024}
}

Acknowledgments

Our code is developed based on img2img-turbo, vision-aided-gan, CONCH and clean-fid. Thanks for their awesome work.