Cardiovascular MRI (CMR) is a non-invasive imaging modality that provides excellent soft-tissue contrast without the use of ionizing radiation. The limited efficiency of MRI data acquisition and physiological motions necessitate data undersampling. Recovering diagnostic quality CMR images from highly undersampled data has been an active area of research, and several data acquisition and processing methods have been proposed to accelerate CMR. The availability of data to objectively evaluate and compare different reconstruction methods could expedite innovation and promote clinical translation of these methods. In this work, we introduce an open-access dataset, called OCMR, that provides multi-coil k-space data from 53 fully sampled cardiac cine series (comprising 81 slices) and 212 real-time, prospectively undersampled cardiac cine series (comprising 842 slices). The fully sampled data are intended for quantitative comparison and evaluation of image reconstruction methods. The free-breathing, prospectively undersampled data are intended to qualitatively evaluate the performance and generalizability of the reconstruction methods. The data were collected on three Siemens Magnetom scanners: Prisma (3T), Avanto (1.5T) and Sola (1.5T). The OCMR dataset is comprised of HDF5 files, with data in each file following the ISMRMRD format. Each file in the dataset is assigned eight attributes, which allows selecting a subset of the dataset.
For more details, visit www.ocmr.info -- a dedicated website for the OCMR dataset.
Before downloading the data, read terms and conditions at www.ocmr.info/download/. If you agree to all the listed terms and conditions, you may proceed with downloading the data. Listed below are the instructions to read OCMR data into Matlab and Python.
Download tar archive (ocmr_cine.tar.gz) by visiting https://ocmr.s3.amazonaws.com/data/ocmr_cine.tar.gz. This single file contains the entire dataset.
Download Matlab wrapper (read_ocmr.m) and an example Matlab script (example_ocmr.m) from https://github.com/MRIOSU/OCMR/tree/master/Matlab.
Download ISMRMRD libraries from https://github.com/ismrmrd/ismrmrd/tree/master/matlab/%2Bismrmrd. Note, only the subfolder ‘/+ismrmrd’ is required.
Place read_ocmr.m, example_ocmr.m, and the entire ‘/+ismrmrd’ subfolder in one folder. Execute example_ocmr.m in Matlab and select the data file to be read; it will generate a nine-dimensional array (kData) for the k-space data and a structure (param) that captures acquisition parameters.
Note, the file listing the attributes of each data file can be found here: https://ocmr.s3.amazonaws.com/ocmr_data_attributes.csv.
Download read_ocmr.py and example_ocmr.ipynb from https://github.com/MRIOSU/OCMR/tree/master/Python.
Execute example_ocmr.ipynb; it will generate a nine-dimensional array (kData) for the k-space data and a structure (param) that captures acquisition parameters. Note, this step eliminates the need to explicitly download the large tar file, ocmr_cine.tar.gz.
By adjusting the filters in example_ocmr.ipynb, one could download selected files based on the attributes. Note, the file listing the attributes of each data file can be found here: https://ocmr.s3.amazonaws.com/ocmr_data_attributes.csv.
Once a data file is read into Matlab or Python, it yields the k-space array, kData, and a structure, param. The kData array has nine dimensions: [kx, ky, kz, coil, phase, set, slice, rep, avg], which represent frequency encoding, first phase encoding, second phase encoding, coil, phase (time), set (velocity encoding), slice, repetition, and number of averages, respectively. For example, a cine stack with frequency encoding size 160, phase encoding size 120, number of coils 18, number of frames 60, number of slices 10 will generate kData with these dimensions: 160x120x1x18x60x1x10x1x1. The second output, param, provides pertinent acquisition parameters. For example, param.FOV, param.TRes, param.flipAngle_deg, param.sequence_type specify field-of-view, temporal resolution, flip angle, and the type of sequence, respectively.
All ISMRMRD data included in OCMR have been de-identified, where Protected Health Information (PHI) as well as scan date and location have been removed. All images have been manually inspected to ensure that identifying facial features are not included.
The current version (v1.0) of OCMR only includes cardiac cine data. We intend to progressively expand the size of cardiac cine data and include additional CMR applications, including 2D phase-contrast MRI.
For help with downloading and reading the data, contact Chong Chen (chen.7211@osu.edu); for feedback or questions about the OCMR dataset, contact the principal investigator, Rizwan Ahmad (ahmad.46@osu.edu)