MRI super-resolution reconstruction using efficient diffusion probabilistic model with residual shifting

🔥🔥Res-SRDiff is a deep learning framework designed to robustly restore high-resolution pelvic T2w MRI and ultra-high field brain T1 maps using an efficient probabilistic diffusion model.

• Our paper on Physics in Medicine and Biology: MRI super-resolution reconstruction using efficient diffusion probabilistic model with residual shifting ❤️

🔍 Diffusion Process

The following diagram illustrates the diffusion process used in this project:

Getting Started

Prerequisites

• Python (>=3.12)
• PyTorch (>=2.5)
• NVIDIA CUDA (for GPU acceleration)
• Additional dependencies as listed in requirements.txt

Installation

1. Clone the repository:

   git clone https://github.com/mosaf/Res-SRDiff.git
   cd Res-SRDiff
   

2. Install dependencies:

   conda env update --file environment.yml --prune
   
   

📦 Pretrained Weights

You can download the pretrained model weights below:

• 🔗 Download Res-SRDiff weights for Prostate (Google Drive)

After downloading, place the .pt files into a folder named out_dir:

mkdir -p out_dir
mv path_to_downloaded_weights.pt out_dir/

⚠️ Note: The provided weights are for research use only and were trained on public datasets.

Running the Code

To run the project, modify the parameters in the main.py file and execute the main.py script:

python main.py

⚙️ Model Hyper-parameters

The diagram below visualizes the key hyper-parameters used in this model:

📚 Citation

If you find Res-SRDiff useful for your research or project, please consider citing our work:

@article{10.1088/1361-6560/ade049,
	author={Safari, Mojtaba and Wang, Shansong and Eidex, Zach and Li, Qiang and Qiu, Richard L J and Middlebrooks, Erik H and Yu, David S and Yang, Xiaofeng},
	title={MRI super-resolution reconstruction using efficient diffusion probabilistic model with residual shifting},
	journal={Physics in Medicine \& Biology},
	url={http://iopscience.iop.org/article/10.1088/1361-6560/ade049},
	doi={https://doi.org/10.1088/1361-6560/ade049},
	year={2025}
}

Acknowledgments

• This project is based on ResShift ❤️.