DisC-Diff

This repository is implemented based on openai/guided-diffusion, with modifications for loss functions and backbone network improvements.

DisC-Diff: Disentangled Conditional Diffusion Model for Multi-Contrast MRI Super-Resolution
Ye Mao*, Lan Jiang *, Xi Chen \, Chao Li,

DisC-Diff is multi-contrast brain MRI super-resolution method designed based on denoising diffusion probabilistic models. Specifically, DisC-Diff leverages a disentangled multi-stream network to exploit complementary information from multi-contrast MRI, improving model interpretation under multiple conditions of multi-contrast inputs. We validated the effectiveness of DisC-Diff on two datasets: the IXI dataset, which contains 578 normal brains, and a clinical dataset with 316 pathological brains.

Dependencies

A conda environment named DisC-Diff can be created and activated by running the following commands:

conda env create -f environment.yaml
conda activate DisC-Diff

Dataset & Pretrained Models

• The processed IXI dataset for training and testing can be downloaded through this link.
• The models pretrained on IXI dataset under x2 & x4 resolution scale can be downloaded through this link.

Model Training

1. Modify the arguments hr_data_dir, lr_data_dir,and other_data_dir in config/config_train.yaml into the paths for your downloaded training T2-HR, T2-LR, and T1-HR data.
2. In train_job.sh, replace the second line into export PYTHONPATH= "Your Repository Path".
3. Run bash train_job.sh.

Model Evaluation

1. Modify the arguments hr_data_dir, lr_data_dir,and other_data_dir in config/config_test.yaml into the paths for your downloaded testing T2-HR, T2-LR, and T1-HR data.
2. In test_job.sh, replace the second line into export PYTHONPATH= "Your Repository Path".
3. Run bash test_job.sh.

Sample Results

BibTeX

@article{mao2023disc,
  title={DisC-Diff: Disentangled Conditional Diffusion Model for Multi-Contrast MRI Super-Resolution},
  author={Mao, Ye and Jiang, Lan and Chen, Xi and Li, Chao},
  journal={arXiv preprint arXiv:2303.13933},
  year={2023}
}