/MRSRGAN

This repo is an official implementation for the paper"Super-Resolution for Ultra High-Field MR Images" from MIDL 2022

Primary LanguagePythonMIT LicenseMIT

MRSRGAN

Notice there's no pre-trained weight available for this model, but the weight could be provided per request via email.

Pre-processing

  • Standardize your images by mean=0, std=1

  • Crop your 3D MRI images into 64*64*64 cubes by running:

    python mains/utils/crop_nifti.py /data/path/to/your/images "your subjectes prefix string(for wildcard search)"

Run the training:

WGAN-GP

python mains/MRSRGAN_WGAN_GP.py --path /your/data/crop/path --val_path /your/validation_data/crop/path

Res10-GAN

This script works the same way as is in WGAN-GP

After inferencing

After the training you will get weights for your patches MRI cubes, then run the following after you've inferenced LR pachtes to assemble them up:

python mains/utils/assemble_crop_v3.py --path /path/to/the/inferred_patches/ --subj "wildcard_name" --scale int(upscale_factor)

Citation

@inproceedings{
wang2022superresolution,
title={Super-Resolution for Ultra High-Field {MR} Images},
author={Qi Wang and Julius Steiglechner and Tobias Lindig and Benjamin Bender and Klaus Scheffler and Gabriele Lohmann},
booktitle={Medical Imaging with Deep Learning},
year={2022},
url={https://openreview.net/forum?id=EFiFV2MSNEB}
}