fastMRI (knee and brain)
fastMRI: raw k-space and DICOM dataset of knee and brain MRI
Raw MRI scanner data (k-space) of knee and brain scans, released to train and compare machine learning methods that reconstruct images from accelerated scans. 1,594 knee and 6,970 brain scans from NYU Langone, plus about 10,000 clinical knee exams as DICOM.
Overview
fastMRI is a joint project of NYU Langone Health and Facebook AI Research (now Meta AI) that released one of the first large collections of raw MRI scanner data. It was built to train and compare machine-learning methods that reconstruct images from undersampled k-space, so that MRI exams can be made faster, and it powered the fastMRI reconstruction challenges.
Composition
The knee part holds fully sampled multi-coil k-space from 1,594 clinical knee scans, each a coronal proton-density volume with or without fat suppression, plus an emulated single-coil version. It is split into 973 training, 199 validation and 118 multi-coil test volumes, with further single-coil test and held-back challenge volumes; test and challenge volumes come undersampled and without ground truth. A separate DICOM collection adds 10,012 consecutive clinical knee exams from 9,290 patients with up to five standard sequences (coronal PD with and without fat suppression, sagittal PD, sagittal and axial fat-suppressed T2). The brain part holds raw k-space for 6,970 axial scans: T1-weighted with and without contrast agent, T2-weighted and FLAIR. Not every brain exam has every contrast. Prostate and breast sets released later are not covered here.
Acquisition
All raw data come from Siemens scanners at NYU: for the knee three 3 T systems (Skyra, Prisma, Biograph mMR) and a 1.5 T Aera with a 15-channel knee coil and a clinical 2D turbo spin echo protocol; for the brain 11 magnets at five locations, at 1.5 T (Avanto, Aera) and 3 T (Prisma, Skyra, Biograph, Tim Trio). Raw data were converted to the vendor-neutral ISMRMRD format and shipped as one HDF5 file per volume. The DICOM images come from a wider range of scanners and are mostly reconstructions of accelerated acquisitions.
Annotations
No clinical labels are included. The reconstruction targets are root-sum-of-squares images of the fully sampled data.
Known limitations
- Brain k-space slices near and below the orbits were zeroed for de-identification, and only axial 2D brain data were released.
- All raw data come from one vendor and one health system.
- No age, sex or diagnosis information is published for the cohort.
- Counts differ slightly between the paper and the website (for example 6,970 vs 7,002 brain scans).
Cohort
Aggregate numbers from the sources below. Bars are relative to the largest value.
Anatomy
- Knee 9,290
Description of the Dataset, DICOM Dataset in Knoll et al. 2020, Radiology - Artificial Intelligence (PMC6996599)
Split by anatomy
ScansReported cross table. A dot marks a cell the source does not give.
| Knee | |
|---|---|
| Training | 973 |
| Validation | 199 |
| Test | 118 |
Table 4 (p. 10) in Zbontar et al., fastMRI - An Open Dataset and Benchmarks for Accelerated MRI (arXiv:1811.08839v2, pages of that version)
Anatomy by field strength
ScansReported cross table. A dot marks a cell the source does not give.
| 3 T | 1.5 T | |
|---|---|---|
| Brain | 3,969 | 3,001 |
| Knee | · | 695 |
Table 2 (p. 8), Table 3 (p. 8) in Zbontar et al., fastMRI - An Open Dataset and Benchmarks for Accelerated MRI (arXiv:1811.08839v2, pages of that version)
Anatomy by contrast / sequence
ScansReported cross table. A dot marks a cell the source does not give.
| Proton density weighted | |
|---|---|
| Knee | 1,594 |
Section 4.2 (p. 7) in Zbontar et al., fastMRI - An Open Dataset and Benchmarks for Accelerated MRI (arXiv:1811.08839v2, pages of that version)
Other reported numbers 8
License and access
Our reading of the license, not legal advice. Before you use the data, read the original license and confirm that your use is allowed. We take no responsibility for how you use a dataset. Full disclaimer
Sign a data use agreement, often reviewed by the provider
NYU Langone Health fastMRI Dataset Sharing Agreement
NYU grants a royalty-free license for internal research or education only, after an online application with an electronic signature. No selling, commercial exploitation or passing on of the data or of files derived from it, except to people under your direct supervision and in cited academic publications. Copies must be destroyed when the work is done.
The website says each sub-dataset has its own Data Sharing Agreement; the agreement text shown on the page is the one summarised here. The prostate and breast sub-datasets are not part of this entry.
What you can do
- No
- Not stated
- Conditional
- Yes
- Conditional
What you can share
- No
- No
- Not stated
What you must do
- Yes
- Share alike No
- Yes
- Conditional
- Manuscript review No
- Release code No
- Return results No
- Yes
Limits
- Yes
- Conditional
Citation
Knoll F, Zbontar J, Sriram A, et al. fastMRI: A Publicly Available Raw k-Space and DICOM Dataset of Knee Images for Accelerated MR Image Reconstruction Using Machine Learning. Radiol Artif Intell 2(1):e190007 (2020). doi:10.1148/ryai.2020190007; and Zbontar J, Knoll F, Sriram A, et al. fastMRI: An Open Dataset and Benchmarks for Accelerated MRI. arXiv:1811.08839.
Sources
Every number on this page comes from one of these documents. Each chart names the table or page it is taken from. The raw numbers are in stats.csv.