This directory contains code for compressed sensing baselines. The baselines are based on the following paper:

ESPIRiT—an eigenvalue approach to autocalibrating parallel MRI: Where SENSE meets GRAPPA (M. Uecker et al., 2013)

which was used as a baseline model in

fastMRI: An Open Dataset and Benchmarks for Accelerated MRI ({J. Zbontar*, F. Knoll*, A. Sriram*} et al., 2018)

The implementation uses the BART toolkit. To install BART, please follow the installation instructions.

Once BART is installed, set the TOOLBOX_PATH environment variable to point to the location where the repo was cloned and PYTHONPATH to the python wrapper for BART:

export TOOLBOX_PATH=/path/to/bart
export PYTHONPATH=${TOOLBOX_PATH}/python:${PYTHONPATH}

where /path/to/bart is the path to the cloned BART repository, not your OS installed BART program.

To run the reconstruction algorithm on the validation data, run:

python run_bart.py \
    --challenge CHALLENGE \
    --data_path DATA \
    --output_path reconstructions_val \
    --reg_wt 0.01 \
    --mask_type MASK_TYPE \
    --split val

where CHALLENGE is either singlecoil or multicoil. And MASK_TYPE is either random (for knee) or equispaced (for brain). The outputs are saved in a directory called reconstructions_val. To evaluate the results, run:

python fastmri/evaluate.py \
    --target-path TARGET_DATA \
    --predictions-path reconstructions_val \
    --challenge CHALLENGE

To apply the reconstruction algorithm to the test data, run:

python run_bart.py \
    --challenge CHALLENGE \
    --data_path DATA \
    --output_path reconstructions_test \
    --split test

The outputs will be saved to reconstructions_test directory which can be uploaded for submission.

Note: for the 2020 Brain Challenge we have opted to not include compressed sensing as a FAIR/NYU baseline for the leaderboard. The 2020 Challenge uses equispaced masks, which are not supported by compressed sensing theory.