Efficient Operators

This repository contains code for the "Efficient Forward (And Backward!) Models for Image Reconstruction" presentation, given as an educational talk at the 2023 ISMRM Annual Meeting.

At a high level the repository contains code for reconstructing GRASP data using modern PyTorch. The repository uses the GRASP data, which the user is expected to download separately.

Installation

These installation instructions require the use of Anaconda. First, go to the miniconda site and install miniconda. After that, you should be able to run the commands below.

Create a new environment to use for the examples:

conda create -n efficient-operator python=3.10

Activate your new environemnt:

conda activate efficient-operator

It's recommended to install the anaconda package, which has a wide suite of basic tools:

conda install anaconda

You will want to separately update ffmpeg so that the videos display correctly in the notebook:

conda update ffmpeg

After that you can install PyTorch using the conda instructions. Once you have PyTorch, you can install torchkbnufft

pip install torchkbnufft

And finally, install the local efficient operator (effop) package by running the following in the root directory of this repository:

pip install -e .

The pip install -e . command is an editable install, so any changes you make in the local files should be reflected when you use the package.

Data configuration

You'll first need to download the data from the following link:

https://cai2r.net/resources/grasp-matlab-code/

After you've downloaded the file, add it to data_loc.yaml as

path/to/liver_data.mat

replacing the path with the true path on your system.

Running the examples

All of the examples from the presentation are in the notebooks folder in order of slide presentation. The final notebook, 5_full_reconstruction.ibynb is self-contained if you want to skip to the end. Note that the outputs have been wiped from this notebook to comply with the original data agreement against redistribution on the NYU website.

Contributing

If you find any mistakes or bugs in the code, please raise an issue.

If you have any questions, feel free to open a discussion.