Self-Supervised Generative Style Transfer for One-Shot Medical Image Segmentation

This repository contains the Pytorch implementation of the proposed method Self-Supervised Generative Style Transfer for One-Shot Medical ImageSegmentation , which has been recently accepted at WACV 2022.

Dependencies

We prefer to have a separate anaconda environment and the following packages to be installed.

1. Python == 3.7
2. tensorflow-mkl == 1.15
3. pytorch == 1.6.0
4. torchvision == 0.7.0
5. pytorch-msssim == 0.2.1
6. medpy == 0.4.0
7. rasterfairy == 1.0.6
8. visdom

Training Modes

The implementaion of our method is available in the folder OURS.

1. Train FlowModel without Appearance Model.

python train.py --ngpus 1  --batch_size 4 --checkpoints_dir_pretrained ./candi_checkpoints_pretrained --dataroot ../CANDIShare_clean_gz --train_mode ae --nepochs 10

2. Train StyleEncoder

python train.py --ngpus 1 --batch_size 16 --checkpoints_dir_pretrained ./candi_checkpoints_pretrained --dataroot ../CANDIShare_clean_gz --train_mode style_moco --nepochs 10

3. Train Appearance Model

python train.py --ngpus 1 --batch_size 1 --checkpoints_dir_pretrained ./candi_checkpoints_pretrained --dataroot ../CANDIShare_clean_gz --train_mode appearance_only --nepochs 10

4. Train Adversarial Autoencoder Flow

python train.py --ngpus 1 --batch_size 2 --checkpoints_dir_pretrained ./candi_checkpoints_pretrained --train_mode aae --nepochs 100

5. Train End to End

python train.py --ngpus 1 --batch_size 1 --checkpoints_dir ./candi_checkpoints --checkpoints_dir_pretrained ./candi_checkpoints_pretrained --dataroot ../CANDIShare_clean_gz --train_mode end_to_end --nepochs 10

For training on OASIS dataset, please change the --dataroot argument to OASIS_clean and --nepochs 1.

Training Steps

1. First train Unet based flow model by running 1. from Train Modes. This will be used to generate images of same styles for training the style encoder.

2. Pre-train style-encoder by running 2. from Train Modes. This will pre-train our style encoder using volumetric contrastive loss.

3. Train end-to-end by running 5. from Train Modes. This will train Appearance Model, Style Encoder and Flow Model end to end using pre-trained Style Encoder. set --use_pretrain to False for training Style Encoder from scratch.

4. Generate Flow Fields in the folder ../FlowFields using trained end-to-end model by running the following command:
   python generate_flow.py

5. Train Flow Adversarial Autoencoder by running 4. from Train Modes.

6. Generate image segmentation pairs using python generate_fake_data.py.

7. Train 3D Unet on the generated image segmentation dataset using the code provided in folder UNET and the following command:

python train.py --exp <NAME OF THE EXPERIMENT> --dataset_name CANDI_generated --dataset_path <PATH TO GENERATED DATASET>

Schematic description of the training phase

Evaluation Script

All evaluation scripts used to generate plots and compute dice score are included in the folder evaluations. To run a particular evaluation, run the following command provinding corresponding opt from the file run_evaluations.py:
python run_evaluations.py <opt>

Pre-trained Models

All pre trained models and datasets can be obtained from here. Please unzip the trained models inside the directory submission_id_675/code/OURS.

Citation

You can find the Self-Supervised Generative Style Transfer for One-Shot Medical Image Segmentation paper at http://arxiv.org/abs/2110.02117

If you find this work useful, please cite the paper:

@misc{tomar2021selfsupervised,
    title={Self-Supervised Generative Style Transfer for One-Shot Medical Image Segmentation},
    author={Devavrat Tomar and Behzad Bozorgtabar and Manana Lortkipanidze and Guillaume Vray and Mohammad Saeed Rad and Jean-Philippe Thiran},
    year={2021},
    eprint={2110.02117},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Licence

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.