SynDiff

Official PyTorch implementation of SynDiff described in the paper.

Muzaffer Özbey*, Onat Dalmaz*, Salman UH Dar, Hasan A Bedel, Şaban Özturk, Alper Güngör, Tolga Çukur, "Unsupervised Medical Image Translation With Adversarial Diffusion Models," in IEEE Transactions on Medical Imaging, vol. 42, no. 12, pp. 3524-3539, Dec. 2023, doi: 10.1109/TMI.2023.3290149.

*: equal contribution

Dependencies

python>=3.6.9
torch>=1.7.1
torchvision>=0.8.2
cuda=>11.2
ninja
python3.x-dev (apt install, x should match your python3 version, ex: 3.8)

Installation

• Clone this repo:

git clone https://github.com/icon-lab/SynDiff
cd SynDiff

Dataset

You should structure your aligned dataset in the following way:

input_path/
  ├── data_train_contrast1.mat
  ├── data_train_contrast2.mat
  ├── data_val_contrast1.mat
  ├── data_val_contrast2.mat
  ├── data_test_contrast1.mat
  ├── data_test_contrast2.mat

where .mat files has shape of (#images, width, height) and image values are between 0 and 1.0.

Sample Data

Sample toy data can also found under 'SynDiff_sample_data' folder of the repository.

Train

python3 train.py --image_size 256 --exp exp_syndiff --num_channels 2 --num_channels_dae 64 --ch_mult 1 1 2 2 4 4 --num_timesteps 4 --num_res_blocks 2 --batch_size 1 --contrast1 T1 --contrast2 T2 --num_epoch 500 --ngf 64 --embedding_type positional --use_ema --ema_decay 0.999 --r1_gamma 1. --z_emb_dim 256 --lr_d 1e-4 --lr_g 1.6e-4 --lazy_reg 10 --num_process_per_node 1 --save_content --local_rank 0 --input_path /input/path/for/data --output_path /output/for/results

Pretrained Models

We have released pretrained diffusive generators for T1->PD and PD->T1 tasks in IXI and T1->T2 and T2->T1 tasks in BRATS datasets. You can save these weights in relevant checkpoints folder and perform inference.

Test

python test.py --image_size 256 --exp exp_syndiff --num_channels 2 --num_channels_dae 64 --ch_mult 1 1 2 2 4 4 --num_timesteps 4 --num_res_blocks 2 --batch_size 1 --embedding_type positional  --z_emb_dim 256 --contrast1 T1  --contrast2 T2 --which_epoch 50 --gpu_chose 0 --input_path /input/path/for/data --output_path /output/for/results

Citation

Preliminary versions of SynDiff are presented in NeurIPS Medical Imaging Meets and IEEE ISBI 2023. You are encouraged to modify/distribute this code. However, please acknowledge this code and cite the paper appropriately.

@ARTICLE{ozbey_dalmaz_syndiff_2024,
  author={Özbey, Muzaffer and Dalmaz, Onat and Dar, Salman U. H. and Bedel, Hasan A. and Özturk, Şaban and Güngör, Alper and Çukur, Tolga},
  journal={IEEE Transactions on Medical Imaging}, 
  title={Unsupervised Medical Image Translation With Adversarial Diffusion Models}, 
  year={2023},
  volume={42},
  number={12},
  pages={3524-3539},
  keywords={Biological system modeling;Computational modeling;Training;Generative adversarial networks;Image synthesis;Task analysis;Generators;Medical image translation;synthesis;unsupervised;unpaired;adversarial;diffusion;generative},
  doi={10.1109/TMI.2023.3290149}}

For any questions, comments and contributions, please contact Muzaffer Özbey (muzafferozbey94[at]gmail.com) or Onat Dalmaz (onat[at]stanford.edu)

(c) ICON Lab 2023

Acknowledgements

This code uses libraries from, pGAN, StyleGAN-2, and DD-GAN repositories.