Dataset contents

This page documents, source by source, what kind of data each dataset actually contains — anatomy, contrasts/modalities, sampling, coil configuration, field strength, the on-disk file format and array layout, and how many files of each kind the catalog exposes.

For the download/credential workflow and the reconstruction pipeline see Usage; for licensing see Licensing & legal.

Every source is normalised to a RawAcquisitionData by load_raw. The temporal / parametric axis (cine frames, mapping weightings, …) is always mapped to ISMRMRD contrasts; the slice axis to ISMRMRD slices; repetition stays 0 unless noted.

MRIDATA — mridata.org

What a file contains. One complete raw acquisition in ISMRMRD form: the readout profiles plus the scanner's XML header (encoding/recon matrix, FOV, TE/TR, receiver channel count, …). The 3-D FSE knee volumes are the workhorse: a single fully-sampled Cartesian volume, typically 8–15 coils, tens of phase-encode partitions.

Metadata. The live scrape of mridata.org/list carries per-card fields — vendor, field strength, channel count (→ entry.receiver_channels), matrix size (→ entry.acquisition_dim), plus TE/TR, institution, protocol, download count under entry.extra. mridata.org's own anatomy labels go well beyond this package's curated DICOM Body Part Examined subset (hip, shoulder, spine, phantom scans, …); anything not recognised becomes entry.anatomy == :other. The committed data/mridata_index.toml is the offline fallback used only when the scrape fails entirely.

Any mridata UUID can also be passed to dataset directly — a minimal entry is synthesised from the UUID.

OCMR_SOURCE — Open Cardiac MRI k-space

What a file contains. One multi-coil cardiac cine acquisition. Coil counts are not in OCMR's metadata — they live only inside the ISMRMRD file — so entry.receiver_channels is nothing until the file is loaded.

Coded metadata. OCMR's attributes CSV uses short codes. The ones with a DICOM anchor are decoded onto core fields; the rest stay in entry.extra (DICOM-keyword named — see extra_schema(OCMR_SOURCE)):

Field strength and fully/under-sampled status are also encoded in the file name (fs_0001_1_5T.h5, us_0014_3T.h5) and used as a fallback.

CMRXRECON2024 — MICCAI 2024 cardiac reconstruction challenge

Array layout. Each .mat holds one k-space array shaped (kx, ky, coils, slices, frames) — 4-D (kx, ky, coils, slices) when there is no temporal axis (BlackBlood). Variable name: kspace_full (FullSample) or kus (undersampled tasks). Access the raw arrays with MRITestData.load_mat.

Series. DICOM has no single "modality" attribute for this — see Taxonomy — so each series decomposes onto several core fields. The file name also encodes the series (e.g. cine_sax, t1map, aorta_tra):

¹ entry.sequence: "modified Look-Locker inversion recovery (fast low angle shot readout)" for T1map, "T2-prepared fast low angle shot" for T2map — confirmed against Wang et al. 2025 (see Taxonomy references): "the modified Look-Locker inversion recovery-fast low angle shot sequence was used for T1 mapping" / "the T2-prepared-fast low angle shot sequence was used for T2 mapping", both "with SAX view". T2 mapping is FLASH-based (spoiled gradient echo), not balanced SSFP.

² Sequence and view are confirmed the same way ("the turbo spin-echo sequence was used for black-blood under breath hold", "black-blood with SAX view"); the paper does not state a T1 vs T2 weighting for it (no TE/TR given), so contrast stays genuinely :unknown — that is a gap in the source, not an unverified guess.

Views. SAX (short-axis) — a stack of parallel slices across the left ventricle (base → apex on the slice axis). LAX (long-axis) — the slice axis instead holds the 2-chamber / 3-chamber / 4-chamber views. LVOT (left ventricular outflow tract, orientation == :lvot) is the third cine view.

File counts in the committed map (data/cmrxrecon2024_map.csv, all MultiCoil):

(Validation/Test acquisition parameters — FOV, TR/TE, flip angle, matrix — are not published upstream, so entry.extra["reconstruction_fov_mm"], ["acquisition_matrix"], ["repetition_time_ms"], ["echo_time_ms"] and ["flip_angle_deg"] are populated only for the Training subjects.)

CMRXRECON300 — revised CMRxRecon-2023 k-space

Reconstruction implication. load_raw reads the true acquired-line pattern from the zero-fill pattern of Recon_ks, so the resulting RawAcquisitionData is correctly marked undersampled. A plain inverse FFT aliases; an artifact-free image needs parallel imaging (ESPIRiT / CG-SENSE) using the paired ACS. The ACS lines are written into the same ISMRMRD file, flagged ACQ_IS_PARALLEL_CALIBRATION and centred in the phase-encode extent; the calib file's coordinates are in entry.locator["calib_path"] (a transport detail, not DICOM metadata — kept out of entry.extra).

Series — every subject has all four:

Sets (committed member maps data/cmrxrecon300_<set>_map.csv; each catalog entry pairs one _ks file with its _calib file). Counts below are entries (i.e. subject × modality):

(A few Test subjects are missing a modality, hence the uneven counts.)

CMRxRecon-300 has no upstream index: its catalog is built from the committed per-set member maps, so refresh_index has nothing to fetch and simply reports them.

USC_SPEECH — USC SPAN 75-speaker real-time speech rtMRI

Reconstruction implication. Because the trajectory is non-Cartesian, the loaded RawAcquisitionData's params["trajectory"] is not "cartesian". Build a non-Cartesian AcquisitionData (with the trajectory + density compensation) — an inverse FFT is not applicable.

What is cataloged. Only the 2drt (2-D real-time) sagittal pharynx/larynx raw k-space. Each entry is one subject × stimulus × repetition (entry.subject_id, entry.repetition); ids look like sub001/2drt/01_vcv1_r1. Committed map: data/usc_speech_map.csv, 75 subjects, 2371 files.

Stimuli (entry.extra["protocol_name"]) — 21 per subject:

M4RAW — low-field brain, multi-contrast, multi-repetition

Array layout. kspace on disk is (slices, coils, freq, phase); HDF5.jl reads it reversed and the loader permutes to canonical (kx, ky, coils, slices, 1). reconstruction_rss is the per-repetition RSS magnitude ground truth (not used by the converter). Multi-repetition data is the point of M4Raw: averaging or learning across repetitions to overcome the low SNR of 0.3 T.

Contrasts (entry.contrast/entry.sequence, with entry.repetition 1, 2, …) and file counts in the committed map (data/m4raw_map.csv, 2030 files):

¹ GRE was added in the M4RawV1.5 release, after the original Scientific Data paper (which describes only T1w/T2w TSE and FLAIR); its T1-weighting is confirmed by the M4RawV1.5 release notes and dataset card ("T1w Gradient echo (GRE) data") on github.com/mylyu/M4Raw and Zenodo record 8056074.

Sets (entry.split): :train (1024), :val (240), :test (400); the GRE archive carries no train/val/test split (entry.split === nothing). All splits are fully sampled.

FASTMRI — NYU / FAIR fastMRI

Individual scans are extracted from the archives without downloading them whole. Until the maintainer populates data/fastmri_map.csv, list_datasets(FASTMRI) is empty — see fastMRI: form-gated credentials.

Array layout. kspace on disk is (slices, coils, kx, ky) (multicoil) or (slices, kx, ky) (knee singlecoil); prostate is 5-D (slices, coils, kx, ky, averages) and the averages (interleaved shots) are summed. A 1-D mask (0/1 over ky) is present for test-split data (prospectively undersampled for the challenge) and absent for fully-sampled train/val data; when absent, acquired lines are detected from non-zero k-space energy.

Per-anatomy contents (committed map data/fastmri_map.csv)

The map's series_variant column is the archive filename's middle token — never a coil count: singlecoil/multicoil for knee/brain (→ entry.coil_data), or the sequence type T2/DIFF for prostate (→ entry.contrast/entry.sequence).

Test-split sampling (bug fix). Knee and brain test-split files are prospectively undersampled for the challenge (they ship a mask); only train/val is genuinely fully sampled. entry.fully_sampled reflects this per-file (split !== :test), not just per-anatomy — a change from the pre-refactor catalog, which incorrectly marked every knee/brain test entry as fully sampled regardless of split.

File counts:

(fastMRI ships only train/val publicly for prostate and breast; the brain/knee test splits are prospectively undersampled, and brain additionally ships *_test_full fully-sampled archives.)

Quick cross-source summary