[Paper] [Dataset] [Github] [BibTeX]

02-07-2024: Updated files for slice-, volume-, exam-level labels and their paths for T2 and Diffusion sequences in the fastMRI prostate dataset.

Classification: The classification folder contains code for training deep learning models to detect clinically significant prostate cancer. Reconstruction: The reconstruction folder contains code for training deep learning models for reconstructing diffusion MRI images from undersampled k-space.

This repository contains code to facilitate the reconstruction of prostate T2 and DWI (Diffusion-Weighted Imaging) images from raw (k-space) data from the fastMRI Prostate dataset. It includes reconstruction methods along with utilities for pre-processing and post-processing the data.

The package is intended to serve as a starting point for those who want to experiment and develop alternate reconstruction techniques.

The code requires python >= 3.9

Install FastMRI Prostate: clone the repository locally and install with

git clone [email protected]:tarund1996/fastmri_prostate_test.git
cd fastmri_prostate_test
pip install -e .

The repository is centered around the fastmri_prostate package. The following breaks down the basic structure:

fastmri_prostate: Contains a number of basic tools for T2 and DWI reconstruction

• fastmri_prostate.data: Provides data utility functions for accessing raw data fields like kspace, calibration, phase correction, and coil sensitivity maps.
• fastmri.reconstruction.t2: Contains functions required for prostate T2 reconstruction
• fastmri.reconstruction.dwi: Contains functions required for prostate DWI reconstruction

fastmri_prostate_recon.py contains code to read files from the dataset and call the T2 and DWI reconstruction functions for a single h5 file.

fastmri_prostate_tutorial.ipynb walks through an example of loading a h5 file from the fastMRI prostate dataset and reconstructing T2/DW images.

To reconstruct T2/DW images from the fastMRI prostate raw data, users can download the dataset and run fastmri_prostate_recon.py with appropriate arguments, specifying the path to the root of the downloaded dataset, output path to store reconstructions, and the sequence (T2, DWI, or both).

python fastmri_prostate_recon.py \  
    --data_path <path to dataset> \  
    --output_path <path to store recons> \  
    --sequence <t2/dwi/both>

The reconstruction algorithms implemented in this package requires the following hardware:

• A computer with at least 32GB of RAM
• A multi-core CPU

The run time of a single T2 reconstruction takes ~15 minutes while the Diffusion Weighted reconstructions take ~7 minutes on a multi-core CPU Linux machine with 64GB RAM. A bulk of the time is spent in applying GRAPPA weights to the undersampled raw kspace data.

fastMRI_prostate is MIT licensed, as found in LICENSE file

If you use the fastMRI Prostate data or code in your research, please use the following BibTeX entry.

@article{tibrewala2024fastmri,
  title={FastMRI Prostate: A public, biparametric MRI dataset to advance machine learning for prostate cancer imaging},
  author={Tibrewala, Radhika and Dutt, Tarun and Tong, Angela and Ginocchio, Luke and Lattanzi, Riccardo and Keerthivasan, Mahesh B and Baete, Steven H and Chopra, Sumit and Lui, Yvonne W and Sodickson, Daniel K and others},
  journal={Scientific Data},
  volume={11},
  number={1},
  pages={404},
  year={2024},
  publisher={Nature Publishing Group UK London}
}

The code for the GRAPPA technique was based off pygrappa, and ESPIRiT maps provided in the dataset were computed using espirit-python