Skip to content

StanfordMIMI/deep-feature-mr-recon

Repository files navigation

Using Deep Feature Distances for Evaluating MR Image Reconstruction Quality

Deep Feature Distances Methods

Datasets

All reader study MR reconstructions and radiologist reader study scores can be downloaded here.

Code

Set-up

Create a conda environment for this project:

conda create -n dfd_env python=3.9
conda activate dfd_env

Install torch

conda install pytorch=2.1.1 torchvision=0.16.1 cudatoolkit=10.1 -c pytorch

Finally install dependencies from the requirements.txt file

pip install -r requirements.txt

Basic Usage

To compute metrics on the MR reconstruction reader study dataset, run the following command:

cd deep-feature-mr-recon # Navigate to your cloned repo
python deep-feature-mr-recon/reader_study_metrics.py --img_dir [path_to_image_folder] --results_dir [path_to_save_results]

The Jupyter Notebook ReaderStudy_vs_Metrics.ipynb can then be used to analyze correlations between the computed metrics and radiologist reader study scores.

Advanced Usage

The deep-feature-mr-recon project is built on top of meddlr, a config-driven an ML framework built to simplify medical image reconstruction and analysis problems. Deep Feature Metrics such as LPIPS and SSFD have been incorporated into meddlr to use as both an evaluation and optimization metric for any MR reconstruction task. Refer to the meddlr documentation for more details.

Citation

@inproceedings{adamson2023using,
  title={Using Deep Feature Distances for Evaluating MR Image Reconstruction Quality},
  author={Adamson, Philip M and others},
  booktitle={NeurIPS 2023 Workshop on Deep Learning and Inverse Problems},
  year={2023}
}

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published