This repository provides code for reproducing the results of the paper: Vision Transformers Enable Fast and Robust Accelerated MRI, by Kang Lin and Reinhard Heckel.
The code has been tested for the environment in requirements.txt, and builds on the code from fastMRI, ConViT, and timm.
The experiments from the paper were performed using the fastMRI dataset and the ImageNet dataset.
First, install PyTorch for your operating system and CUDA setup from the PyTorch website.
Then, install all other dependencies from requirements.txt. This can be done, for example, by running
pip install -r requirements.txt
from the directory where you saved requirements.txt. Alternatively, you may run
pip install fastmri
pip install timm
to obtain the dependencies.
The code for reproducing the paper results are provided as Jupyter notebooks: fastmri_training.ipynb and imagenet_pretrain.ipynb.
The notebook fastmri_training.ipynb handles model training, fine-tuning and evaluation on the fastMRI dataset.
The notebook imagenet_pretrain.ipynb provides the code for pre-training our models on the ImageNet dataset.
You may adjust the hyperparamters according to the descriptions in the paper. Also note that in both notebooks the data directory path has to be specified at the marked places.
In the experiments, we also used a simulated single-coil brain dataset, which has been simulated in the same fashion as fastMRI's single-coil knee dataset. The code to reproduce this dataset is provided in simulate_singlecoil_from_multicoil.ipynb.
@inproceedings{
lin2022vision,
title={Vision Transformers Enable Fast and Robust Accelerated {MRI}},
author={Kang Lin and Reinhard Heckel},
booktitle={Medical Imaging with Deep Learning},
year={2022},
url={https://openreview.net/forum?id=cNX6LASbv6}
}
This repository is Apache 2.0 licensed.