DDM²: Self-Supervised Diffusion MRI Denoising with Generative Diffusion Models, ICLR 2023

Paper: https://arxiv.org/pdf/2302.03018.pdf

Dependencies

Please clone our environment using the following command:

conda env create -f environment.yml  
conda activate ddm2

Usage

Data

For fair evaluations, we used the data provided in the DIPY library. One can easily access their provided data (e.g. Sherbrooke and Stanford HARDI) by using their official loading script:

hardi_fname, hardi_bval_fname, hardi_bvec_fname = get_fnames('stanford_hardi')
data, affine = load_nifti(hardi_fname)

Configs

Different experiments are controlled by configuration files, which are in config/.

We have provided default training configurations for reproducing our experiments. Users are required to change the path vairables to their own directory/data before running any experiments. More detailed guidances are provided as inline comments in the config files.

Train

The training of DDM² contains three sequential stages. For each stage, a corresponding config file (or an update of the original config file) need to be passed as a coommand line arg.

1. To train our Stage I:
   python3 train_noise_model.py -p train -c config/hardi_150.json
   or alternatively, modify run_stage1.sh and run:
   ./run_stage1.sh

2. After Stage I training completed, the path to the checkpoint of the noise model need to be specific at 'resume_state' of the 'noise_model' section in corresponding config file. Additionally, a file path (.txt) needs to be specified at 'initial_stage_file' in the 'noise_model' section. This file will be recorded with the matched states in Stage II.

3. To process our Stage II:
   python3 match_state.py -p train -c config/hardi_150.json
   or alternatively, modify run_stage2.sh and run:
   ./run_stage2.sh

4. After Stage II finished, the state file (a '.txt' file, generated in the previous step) needs to be specified at 'stage2_file' variable in the last line of each config file. This step is neccesary for the following steps and inference.

5. To train our Stage III:
   python3 train_diff_model.py -p train -c config/hardi_150.json
   or alternatively, modify run_stage3.sh and run:
   ./run_stage3.sh

6. Validation results along with checkpoints will be saved in the /experiments folder.

Inference (Denoise)

One can use the previously trained Stage III model to denoise a MRI dataset through:
python denoise.py -c config/hardi.json
or alternatively, modify denoise.sh and run:
./denoise.sh

The --save flag can be used to save the denoised reusults into a single '.nii.gz' file:
python denoise.py -c config/hardi.json --save

Quantitative Metrics Calulation

With the denoised Stanford HARDI dataset, please follow the instructions in quantitative_metrics.ipynb to calculate SNR and CNR scores.

This notebook is derived from this DIPY script. Please respect their license of usage.

Citation

If you find this repo useful in your work or research, please cite:

@inproceedings{xiangddm,
  title={DDM $\^{} 2$: Self-Supervised Diffusion MRI Denoising with Generative Diffusion Models},
  author={Xiang, Tiange and Yurt, Mahmut and Syed, Ali B and Setsompop, Kawin and Chaudhari, Akshay},
  booktitle={The Eleventh International Conference on Learning Representations}
}