A Reinforcement learning environment to facilitate research on active MRI acquisition. The goal of active-mri-acquisition is to provide a convenient gym-like interface to test the use of reinforcement learning and planning algorithms for subject-specific acquisition sequences of MRI scans.

This repository also contains scripts to replicate the experiments performed in:

Luis Pineda, Sumana Basu, Adriana Romero, Roberto Calandra, Michal Drozdzal, "Active MR k-space Sampling with Reinforcement Learning". MICCAI 2020.

The data to produce the plot in Figure 4 can also be found at this link. Once extracted, the folder structure should be easy to understand, corresponding to acceleration case, policy, and metric, and one file per folder. The file has a dictionary on which doing loaded_dict[time_step]["all"][image_idx] gives you the results for the policy at the time step and image index indicated.

fastMRI repository

fastMRI dataset

active-mri-acquisition is a Python 3.7+ library. We suggest creating a new Python environment for this project, for example by running

    conda create --name activemri python=3.7

Also, make sure your Python environment has PyTorch installed with the appropriate CUDA configuration for your system (we have tested it using CUDA 9.2 and 10.1).

Then, to install active-mri-acquisition, clone this repository, then run

    cd active-mri-acquisition
    pip install -e .

If you also want the developer tools for contributing, run

    pip install -e ".[dev]"

To test your installation, run

    python -m pytest tests/core

To run the fastMRI environments, you need to configure a couple of things. If you try to run any of the default environments for the first time (for example, see our intro notebook), you will see a message asking you to add some entries to the defaults.json file. This file will be created automatically the first time you run it, located at $USER_HOME/.activemri/defaults.json. It will look like this:

{
  "data_location": "",
  "saved_models_dir": ""
}

To run the environments, you need to fill these two entries. Entry "data_location" must point to the root folder in which you will store the fastMRI dataset (for instructions on how to download the dataset, please visit https://fastmri.med.nyu.edu/). Entry "saved_models_dir" indicates the folder where the environment will look for the checkpoints of reconstruction models.

For instructions on how to run the environment, evaluating baselines, and adding your own reconstruction models, please see our documentation.

To run evaluation of the algorithms considered in the paper, please take a look at the example scripts.

@inproceedings{pineda2020activemri,
    author = {Luis Pineda and Sumana Basu and Adriana Romero and Roberto Calandra and Michal Drozdzal},
    title = {{Active MR k-space Sampling with Reinforcement Learning}},
    booktitle = {{International Conference on Medical Image Computing and Computer-Assisted
Intervention}},
    year = {2020},
    publisher="Springer International Publishing",
}

active-mri-acquisition is MIT licensed, as found in the LICENSE file.