Robust Physics-based Deep MRI Reconstruction Via Diffusion Purification

The paper can be found here: https://arxiv.org/pdf/2309.05794

Brief explanation of the inference procedure

The primary workflow consists of two key stages: first, the purification process, implemented using the adv_purification code. Second is passing through the MoDL pre-trained model.

Setting up the code:

Install the required dependencies: mkdir weights wget -O weights/checkpoint_95.pth https://www.dropbox.com/s/27gtxkmh2dlkho9/checkpoint_95.pth?dl=0

create env and activate

conda create -n name python=3.8 conda activate name

install dependencies

conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch

Usage Download the dataset can be found from Dropbox: Data avaliable on https://www.dropbox.com/scl/fi/801dxovhbkp2bkl2krz5x/NEW_KSPACE.zip?rlkey=4u3b32f6c4pfujsv3kp7z5bdk&st=hwe9thrv&dl=0

Open and run the adv_purification.py to have the purification result for the initial stage and evaluate the model on image restoration tasks based on the pretrained MoDL model and run th test case result.

The pretraind MoDL is also on the dropbox https://www.dropbox.com/scl/fi/xnlrcexczb8yr3neshj1b/DIDN_lambda1_3000_images_trained_2.pt?rlkey=icrqens0ltzvtjusxyzuck89s&st=315c00yp&dl=0

The pretrained diffusion model can also be found in the dropbox link https://www.dropbox.com/s/27gtxkmh2dlkho9/checkpoint_95.pth?dl=0. Please note the pre-trained model was adopted from: https://arxiv.org/pdf/2110.05243

if you want to gain the faster MRI reconstruction purification, we will recommond you to follow DDS https://github.com/hyungjin-chung/DDS .The pretraind diffusion model for DDS can be found in the same github page

Directory Structure:

models/: Contains model architecture code which have the DIDN network and the score based MRI model network

autoattack/: Implementation of the auto attack algorithom.

utils/: Utility functions for the project.