FDB

Official PyTorch implementation of FDB as described in the paper

Mirza, M. Usama, Onat Dalmaz, Hasan A. Bedel, Gokberk Elmas, Yilmaz Korkmaz, Alper Gungor, Salman UH Dar, Kader K. Oguz, and Tolga Çukur. "Learning fourier-constrained diffusion bridges for mri reconstruction." IEEE Transactions on Medical Imaging (2026).

Dependencies

python==3.8.13
blobfile==2.0.2
h5py==3.9.0
imageio==2.22.1
mpi4py==3.1.4
numpy==1.24.4
Pillow==10.0.0
torch==2.0.1

Installation

• Clone this repo:

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

Train

For Single-Coil

python train.py --data_dir /path_to_data/ --log_interval 5000 --save_dir 'model_singlecoil' --save_interval 5000 --image_size 256 --num_channels 128 --num_res_blocks 3 --learn_sigma False --dropout 0.3 --diffusion_steps 1000 --lr 1e-4 --batch_size 1 --lr_anneal_steps 100000 --undersampling_rate 2 --data_type 'singlecoil'

For Multi-Coil

python train.py --data_dir /path_to_data/ --log_interval 5000 --save_dir 'model_multicoil' --save_interval 5000 --image_size 384 --num_channels 128 --num_res_blocks 3 --learn_sigma False --dropout 0.3 --diffusion_steps 1000 --lr 1e-4 --batch_size 1 --lr_anneal_steps 15000 --undersampling_rate 2 --data_type 'multicoil'

Inference

For Single-Coil

python sample.py --model_path model_singlecoil/ema_0.9999_100000.pt --data_path /path_to_data/ --image_size 256 --num_channels 128 --num_res_blocks 3 --learn_sigma False --dropout 0.3 --diffusion_steps 1500 --save_path results_singlecoil --num_samples 1 --batch_size 1 --data_type 'singlecoil' --R 4 --contrast 'T1'

For Multi-Coil

python sample.py --model_path model_multicoil/ema_0.9999_015000.pt --data_path /path_to_data/ --image_size 384 --num_channels 128 --num_res_blocks 3 --learn_sigma False --dropout 0.3 --diffusion_steps 1750 --save_path results_multicoil --num_samples 1 --batch_size 1 --data_type 'multicoil' --R 8 --contrast 'FLAIR'

Citation

You are encouraged to modify/distribute this code. However, please acknowledge this code and cite the paper appropriately.

@article{mirza2026learning,
  title={Learning fourier-constrained diffusion bridges for mri reconstruction},
  author={Mirza, M Usama and Dalmaz, Onat and Bedel, Hasan A and Elmas, Gokberk and Korkmaz, Yilmaz and Gungor, Alper and Dar, Salman UH and Oguz, Kader K and {\c{C}}ukur, Tolga},
  journal={IEEE Transactions on Medical Imaging},
  year={2026},
  publisher={IEEE}
}

For any questions, comments and contributions, please contact Usama Mirza (usama.mirza.819[at]gmail.com )

(c) ICON Lab 2023

Acknowledgements

This code uses libraries from DiffuseRecon and Improved DDPM repositories.