A Path Towards Clinical Adaptation of Accelerated MRI

Accelerated MRI reconstructs images of clinical anatomies from sparsely sampled signal data to reduce patient scan times. While recent works have leveraged deep learning to accomplish this task, such approaches have often only been explored in simulated environments where there is no signal corruption or resource limitations. In this work, we explore the following augmentations to neural network MRI image reconstructors to enhance their clinical relevancy. Namely,

1. (Detecting Signal Corruption) We propose a neural network-based approach for detecting sources of image artifacts during signal preprocessing.
2. (Acceleration Factor Balancing) We demonstrate that training reconstructors on MR signal data with variable acceleration factors can improve their average performance at any time step of a clinical patient scan.
3. (Learning Multiple Anatomies) We offer a new loss function to overcome catastrophic forgetting when models learn to reconstruct MR images of multiple anatomies and orientations.
4. (Phantom Pre-Training) We propose a framework using simulated phantom data to leverage transfer learning for learning to reconstruct other anatomies with limited clinically acquired datasets and compute capabilities.

Installation

To install and run our code, first clone the accMRI repository.

git clone https://github.com/michael-s-yao/accMRI
cd accMRI

Next, create a virtual environment and install the relevant dependencies.

python -m venv env
source env/bin/activate
pip install -r requirements.txt

Download and organize the fastMRI single- and multi- coil data from the fastMRI website. The datasets may be organized as follows:

└── data
    ├── knee       
        ├── knee_multicoil_train
        ├── knee_multicoil_val
        ├── knee_multicoil_test
        ├── knee_singlecoil_train
        ├── knee_singlecoil_val
        ├── knee_singlecoil_test
    ├── brain      
        ├── brain_multicoil_train
        ├── brain_multicoil_val
        ├── brain_multicoil_test
        ├── brain_singlecoil_train
        ├── brain_singlecoil_val
        ├── brain_singlecoil_test

Organization

The disciminator directory contains our implementation for the k-space signal corruption detector, while the reconstructor directory contains our implementation for our acceleration factor balancing, sequential learning, and transfer learning experiments. Generally speaking, main.py trains a model while infer.py allows you to run inference using a trained model. In training models using reconstructor/main.py, we highlight the following arguments:

• (Acceleration Factor Balancing) In our implementation, a model will be trained on variably accelerated data by default. You can set a fixed acceleration factor(s) to use for training using the --fixed_acceleration argument.
• (Learning Multiple Anatomies) In sequentially learning a subsequent reconstruction task, you can specify the EWC weighting parameter in the --ewc argument, and a path to the checkpoint file from the previous learning task in the --ewc_state_dict argument.
• (Phantom Pre-Training) To pretrain a model using simulated phantom data, use the --tl flag and specify the number of coils to simulate using the --num_coils argument (default 15 coils). You can then finetune the pre-trained model by passing it in to the --ckpt_path argument in a subsequent run of main.py.

For additional details, please run python main.py --help (or python infer.py --help) within the relevant directory.

Contact

Questions and comments are welcome. Suggestions can be submitted through Github issues. Contact information is linked below.

Michael Yao

Michael Hansen (Corresponding Author)

Citation

@InProceedings{pmlr-v193-yao22a,
  title={A Path Towards Clinical Adaptation of Accelerated {MRI}},
  author={Yao, Michael S. and Hansen, Michael S.},
  booktitle={Proceedings of the 2nd Machine Learning for Health symposium},
  pages={489-511},
  year={2022},
  volume={193},
  series={Proceedings of Machine Learning Research},
  publisher={PMLR},
  pdf={https://proceedings.mlr.press/v193/yao22a/yao22a.pdf},
  url={https://proceedings.mlr.press/v193/yao22a.html},
}

License

This repository is MIT licensed (see LICENSE). Portions of the code in this repository is adapted from the fastMRI repository as part of a collaborative research project between Facebook AI Research (FAIR) and NYU Langone Health.