ocmr-preproc-tf

This is a pipeline in Tensorflow framework to preprocess the OCMR (Open-Access Multi-Coil k-Space Dataset for Cardiovascular Magnetic Resonance Imaging) dataset into the .tfrecord files.

You can attach the details of the OCMR dataset by the following materials.

Paper:

Chen, Chong, et al. "OCMR (v1. 0)--Open-Access Multi-Coil k-Space Dataset for Cardiovascular Magnetic Resonance Imaging." arXiv preprint arXiv:2008.03410 (2020).

GitHub:

MRIOSU/OCMR: OCMR (Open-Access Repository for Multi-Coil k-space Data for Cardiovascular Magnetic Resonance Imaging) (github.com)

Homepage:

https://ocmr.info/

We note that the num of fully sampled data of OCMR is 204, and we divide the data into 124 for training; 40 for validation; and 40 for test, see devide_info.md for the definition of num and more info. The ratio of train, validation and test is about 7:2:2.

We use ESPIRiT method to estimate the coil sensitivity maps of the multi-coil images and then merge them into single-coil images.

We perform the data augmentation for the training data, i.e., crop the single-coil data into 128×128×16 (x,y,t) with the step of 32×32×8. Eventually, we get 1848 training images.

1. get the preprocessed tfrecord file

You can download the preprocessed tfrecord files via my onedrive

2. Run the code

run main-bart.py after the following steps done. u may change the dirs by yourself.

2.1 Download the data

click HERE to download the original full data

u can get more info in the OCMR Homepage https://ocmr.info/, also the download link.

2.2 divide the data into train/val/test

see devide_info.md

2.3 env & requirements

Linux ( suggested )

• if you are using Docker, you can easily pull the image and create a container via

  docker pull yhaoz/tf:2.9.0-bart
  

• else

  • install bart

    We use bart to estimate the coil sensitivity maps (csm) to combine the multi-coil image into a single-coil image

    For more info,

    GitHub: mrirecon/bart: BART: Toolbox for Computational Magnetic Resonance Imaging (github.com)

    Homepage: BART Toolbox (mrirecon.github.io)

  • pip install -r requirements.txt

    ismrmrd==1.12.5
    matplotlib==3.5.2
    numpy==1.22.3
    tensorflow==2.9.0
    

  • u may also need to install the ismrmrd-python-tools manually

    GitHub: ismrmrd/ismrmrd-python-tools: ISMRMRD Python Toolbox (github.com)

    git clone https://github.com/ismrmrd/ismrmrd-python-tools.git
    pip install ismrmrd-python-tools-master
    

windows ( NOT suggested )

Except bart, u should install other requirements, requiements.txt & ismrmrd-python-tools. Then, u should finetune the code main-bart.py

• remove the bart corresponding codes

• comment the line using bart calculate the coil sensitivity maps

  # csm = bart(1, 'ecalib -m1', k[..., 0]) 

• use the espirit implementation in python via

  csm = espirit_csm(k[..., 0])

  ESPIRiT is a method to estimate the coil sensitivity maps (csm)

  For more info,

  Paper: Uecker, Martin, et al. "ESPIRiT—an eigenvalue approach to autocalibrating parallel MRI: where SENSE meets GRAPPA." Magnetic resonance in medicine 71.3 (2014): 990-1001.

  GitHub:

  mikgroup/espirit-python: Python ESPIRiT implementation. (github.com)

• Note that using bart to calculate the csm takes 1 second while using the python implementation of espirit takes more than 100 seconds

3. Thanks

We thank the following GitHub repos referred to which we finally complete the preprocessing code.

wenqihuang/LS-Net-Dynamic-MRI: This the code repository for paper "Deep Low-rank plus Sparse Network for Dynamic MR Imaging". (github.com)

Keziwen/SLR-Net: Code for our work: "Learned Low-rank Priors in Dynamic MR Imaging" (github.com)

mrirecon/bart: BART: Toolbox for Computational Magnetic Resonance Imaging (github.com)

MRIOSU/OCMR: OCMR (Open-Access Repository for Multi-Coil k-space Data for Cardiovascular Magnetic Resonance Imaging) (github.com)

mikgroup/espirit-python: Python ESPIRiT implementation. (github.com)

4. License

This repo is with the GPL-3.0 license.