https://atlas.rsna.org/schemas/2025-11/dataset.json

fastMRI: A Publicly Available Raw k-Space and DICOM Dataset of Knee Images for Accelerated MR Image Reconstruction Using Mach...

https://doi.org/10.1148/ryai.2020190007

Yvonne.Lui@nyulangone.org

National Institutes of Health grants R01EB024532 and P41EB017183.

Curation of the dataset was part of a study approved by our local institutional review board.

The fastMRI dataset is the first large-scale public dataset that includes raw k-space data and DICOM data from a clinical population, tailored for MR image reconstruction research using machine learning. It aims to accelerate research, enable large-scale validation, and enhance reproducibility in the field.

The purpose of the fastMRI dataset is to provide the first step toward addressing the lack of a large-scale public dataset that includes raw k-space data for MR image reconstruction. It aims to provide a resource to improve image acquisition and reconstruction itself using machine learning techniques.

The k-space dataset consists of fully sampled data from 1594 consecutive clinical MRI proton density–weighted acquisitions of the knee. The DICOM dataset includes 10012 consecutive clinical knee MRI examinations from 9290 patients, representing a full complement of clinical acquisitions.

The 1594 k-space data examples are partitioned into training, validation, multicoil testing, single-coil testing, and a hold-back set for a challenge. The 10012 DICOM data are not partitioned into training, validation, testing, and challenge sets, but provided as a single dataset for auxiliary training or generalizability testing.

The dataset does not provide diagnostic labeling segmentations, text reports, statistics on the prevalence of pathology, information on metal implants, or demographic information.

The examples in the training and validation set are identical for the single-coil and multicoil k-space datasets. For the challenge and test set, unique examples are provided for the single-coil and multicoil datasets to prevent information sharing.

No examinations were excluded owing to presence of imaging artifacts from motion, pulsatile flow, and so forth, for both k-space and DICOM data.

k-space data were deidentified via conversion to the vendor-neutral International Society for Magnetic Resonance in Medicine (ISMRM) raw data format. DICOM data were deidentified by using the Radiological Society of North America’s clinical trial processor tool. All metadata, as well as the DICOM images themselves, were manually inspected to ensure that no protected health information remained in the dataset.

Not possible due to deidentification of k-space data and DICOM data, manual inspection for PHI, and generation of random integer patient identifiers for DICOM.

The dataset includes pathologic findings at a rate representative of a clinical patient population, as it consists of consecutive examinations.