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

k-Space Dataset

Fully sampled k-space data from 1594 consecutive clinical MRI proton density–weighted acquisitions of the knee in the coronal plane with and without frequency-selective fat saturation are included. The measurement identifiers in the k-space data were generated to be random integers. No examinations were excluded owing to presence of imaging artifacts from motion, pulsatile flow, and so forth. No contrast agents were injected at any of these examinations. Scans were performed on three clinical 3-T systems (Siemens Magnetom Skyra, Prisma, and Biograph-mMR) and one clinical 1.5-T system (Siemens Magnetom Aera) using clinical multichannel receive coils. Cartesian 2D turbo spin-echo sequences that are part of the routine clinical protocol at our institution were used. The publicly available software package Yarra (10) was used to gather k-space data from the MRI scanners. Example images from reference reconstructions are shown in Figure 1. The data are provided together with metadata that allow reconstruction of images by means of a simple inverse Fourier transform. In particular, the individual k-space lines are already correctly sorted according to their position in the acquisition trajectory. No further preprocessing steps were performed on the data. An overview of the most relevant metadata fields is given in Table 2. For a complete list of the metadata that is included in the ISMRM raw data format, we refer the reader to Inati et al (9). The article that describes the k-space data format is also accompanied by an online code repository that provides tools to load and reconstruct the data for most commonly used programming languages and computing environments (C/C++, Matlab, Python). k-Space data from the fastMRI dataset can be processed with any of these code resources. Vendor-specific metadata about the pulse sequences used for data acquisition are not included. Because the data were acquired with a multichannel receive array coil, a proper combination of the individual coil images is a necessary step in the image reconstruction process. The most straightforward approach, which is also commonly used in clinical MRI protocols, is to use a sum-of-squares combination of the individual coil images. Image reconstruction of accelerated acquisitions via parallel imaging requires an additional calibration step to obtain coil sensitivity information. This can be done either explicitly by obtaining maps of the coil sensitivity profiles (11,12) or by estimating convolution kernels in k-space (13,14). Coil sensitivity profiles are not included in the database for two reasons. First, we want to avoid any bias toward a particular method for coil sensitivity estimation. The most common strategies for parallel imaging are described in the references cited above and several open-source software implementations for them are available online. Second, this would double the size of the dataset.

Figure 1:

Download as PowerPoint

Table 2: Overview of Selected Metadata Fields That Are Included Together with the Raw k-Space Data

We also provide simulated single-coil k-space data derived from the acquired multicoil k-space data using an “emulated single-coil” combination algorithm (15). The rationale for providing simulated single-coil data—even though reconstruction from multicoil data is expected to be more precise and closer to most clinical acquisition and reconstruction pipelines—is threefold: (a) To lower the barrier of entry for researchers who may not be familiar with MRI data, since the use of a single coil removes a layer of complexity, (b) to include a task that is relevant for the single-coil MRI machines still in use throughout the world, and (c) to separate out the aspects of reconstruction related to compressed sensing rather than parallel imaging. The 1594 k-space data examples are partitioned into the following six components: (a) training—coronal proton density weighted (484 examinations, average of 36 images) and coronal proton density weighted with fat suppression (489 examinations, average of 36 images); (b) validation—coronal proton density weighted (100 examinations, average of 36 images) and coronal proton density weighted with fat suppression (99 examinations, average of 36 images); (c) multicoil testing—coronal proton density weighted (59 examinations, average of 36 images) and coronal proton density weighted with fat suppression (59 examinations, average of 36 images); (d) single-coil testing—coronal proton density weighted (54 examinations, average of 36 images) and coronal proton density weighted with fat suppression (54 examinations, average of 36 images).

The remaining 196 examinations are held back for a planned image reconstruction challenge. The training and validation datasets may be used to fit model parameters and to optimize hyperparameter values. The test dataset is used to compare the results across different approaches. Evaluation on the test set is accomplished by uploading results to the public leaderboard at https://fastmri.org/. The first official challenge associated with the dataset is forthcoming.

The examples in the training and validation set are identical for the single-coil and multicoil datasets. For the challenge and test set, unique examples are provided for the single-coil and the multicoil dataset. This ensures that information cannot be shared between the two challenges at the test stage.

Examples in the test and challenge sets contain undersampled k-space data. The undersampling is performed by retrospectively masking k-space lines from a fully sampled acquisition. k-Space lines are omitted only in the phase-encoding direction to simulate physically realizable accelerations in 2D data acquisitions. The same undersampling mask is applied to all slices of an example. To provide diverse undersampling patterns across the datasets, the undersampling mask is chosen randomly for each example, subject to constraints on the number of fully sampled central lines and the overall undersampling factor. Figure 2 shows details on the undersampling procedure.

Figure 2:

Download as PowerPoint

DICOM Dataset

In addition to the k-space data, fastMRI also includes 10 012 consecutive DICOM image datasets from 9290 patients undergoing clinical knee MRI examinations with a full complement of clinical acquisitions represented, including a range of tissue contrasts and different planes of imaging. The total number of examinations and average number of images per sequence are: (a) coronal proton density weighted: 9947 examinations, average of 33 images; (b) coronal proton density weighted with fat suppression: 10 192 examinations, average of 31 images; (c) axial T2 weighted with fat suppression: 9640 examinations, average of 33 images; (d) sagittal proton density weighted: 10 491 examinations, average of 31 images; and (e) sagittal T2 weighted with fat suppression: 7311 examinations, average of 29 images.

This also includes contrast agent−enhanced examinations. Both pre- and postcontrast images are included. There is no overlap between the examinations in the DICOM dataset and the k-space dataset. No examinations were excluded owing to the presence of imaging artifacts from motion, pulsatile flow, and so forth. The DICOM data are produced using a wide range of scanners within our institution. They are not partitioned into training, validation, testing, and challenge sets. Instead, they are provided as a single dataset, for example, for the purpose of auxiliary training or to test generalizability of techniques developed. The DICOM patient identifiers were generated to be random integers.