Accurate multi-contrast MRI super-resolution via a dual cross-attention transformer network (DCAMSR)

<div align="center"> <img src="figure/Architecture.png" width="70%"> </div>

News

2023.2.3 here we are 🪧🪧🪧

2023.5.24 Early Accepted By MICCAI2023 🎆🎆🎆

Dependencies

• numpy==1.18.5
• scikit_image==0.16.2
• torchvision==0.8.1
• torch==1.7.0
• runstats==1.8.0
• pytorch_lightning==0.9.0
• h5py==2.10.0
• PyYAML==5.4
• timm
• einops
• python-opencv

Data:

The data used for the image super-resolution task comes from the fastMRI dataset and M4Raw.

The multi-contrast MR images csv file is released in dataset fold. fastMRI csv file comes from MINet.

Within each task folder, the following structure is expected:

    data0/fastmri_knee
    ├── singlecoil_train
    │   ├── xxx.h5
    │   ├── ...
    ├── singlecoil_val
    │   ├── xxx.h5
    │   ├── ...

    data0/M4RawV1.1
    ├── multicoil_train
    │   ├── xxx.h5
    │   ├── ...
    ├── multicoil_val
    │   ├── xxx.h5
    │   ├── ...

Code Usage

Install

pip install -r requirement.txt

Training

cd experimental/DCAMSR
python train.py

Evaluation

cd experimental/DCAMSR
python test.py --mode test --resume xxx

weight release

Acknowledgement

We borrow some codes from MASA and MINet. We thank the authors for their great work.