Taxonomy

DatasetEntry's field names, value vocabularies and units are anchored in the DICOM standard wherever DICOM has an attribute for the concept. A handful of fields have no DICOM equivalent; those are documented extensions, each with a one-line justification. This page is the reference for both, plus the ISMRMRD trajectory crosswalk and the external sources consulted while designing it.

Use dicom_tag/dicom_keyword to look the mapping up programmatically:

julia> dicom_tag(:field_strength)
(0x0018, 0x0087, "MagneticFieldStrength")

julia> dicom_keyword(:receiver_channels)  # an extension — no DICOM tag

Core fields → DICOM attributes

source, id, approx_size_bytes, sha256, url, extra and locator are transport or identity fields, not imaging metadata, and are not DICOM-anchored.

extra keys → DICOM attributes

Populated per source (see extra_schema for what each source actually carries):

Modality (0008,0060) is MR for every entry in the package — a constant, stored nowhere. That is also why the old free-text modality key was retired: the name is taken by DICOM and means something else (see TAXONOMY_EXTENSIONS and the per-source decomposition into contrast/cardiac_sync/orientation/etc.).

The seven extensions

Each exists because DICOM has no attribute for the concept:

• receiver_channels — DICOM enumerates coil elements in Multi-Coil Definition Sequence (0018,9045) with Multi-Coil Element Used (0018,9048); there is no count attribute. Anchored on ISMRMRD acquisitionSystemInformation/receiverChannels.
• split — ML-corpus partition, not an imaging concept.
• cohort — no DICOM research-subject-class attribute.
• undersampling_pattern — Parallel Acquisition Technique (0018,9078) names the reconstruction (SENSE/SMASH/GRAPPA), not the sampling mask. VISTA, kt-Gaussian, kt-radial, Poisson-disc have no term.
• orientation vocabulary — View Code Sequence (0054,0220) is the right container and takes SNOMED codes, but DICOM publishes no Context Group for cardiac MR views (the cardiac view CIDs are echocardiography). Local symbols for now; SNOMED coding is a later refinement, and codes must not be invented.
• quantitative — DICOM's anchor is the Parametric Map Storage SOP class, an object type rather than an attribute. BIDS T1map/T2map suffix is the practical equivalent.
• trajectory vocabulary — Geometry of k-Space Traversal (0018,9032) offers only RECTILINEAR/RADIAL/SPIRAL; EPI lives in Scanning Sequence (EP) and golden-angle is a radial ordering with no term. ISMRMRD's enum is strictly more expressive and is the package's output format (crosswalk below).
• acceleration — Parallel Reduction Factor In-plane (0018,9069) is in-plane and parallel-imaging-specific. acceleration is net R, defined at the source's native frame binning; kept in extra["parallel_reduction_factor_in_plane"] when the DICOM value is separately known.

fully_sampled is not an extension: it is Percent Sampling (0018,0093) == 100. The boolean is kept for query ergonomics.

ISMRMRD trajectoryType crosswalk

External references consulted

Standards:

• DICOM PS3.3 §C.8.13.3, MR Image Description Macro — Acquisition Contrast (0008,9209) defined terms, Complex Image Component (0008,9208). https://dicom.nema.org/medical/dicom/current/output/chtml/part03/sect_C.8.13.3.html
• DICOM PS3.3 §C.8.13.4, MR Pulse Sequence Module — MR Acquisition Type (0018,0023), Echo Pulse Sequence (0018,9008), Phase Contrast (0018,9014), Spectrally Selected Suppression (0018,9025), Segmented k-Space Traversal (0018,9033), Coverage of k-Space (0018,9094), Geometry of k-Space Traversal (0018,9032). https://dicom.nema.org/medical/dicom/current/output/chtml/part03/sect_C.8.13.4.html
• DICOM PS3.3 §C.7.6.18, Physiological Synchronization — Cardiac Synchronization Technique (0018,9037) defined terms, Cardiac Signal Source (0018,9085). https://dicom.nema.org/medical/dicom/current/output/chtml/part03/sect_C.7.6.18.html
• DICOM PS3.16 Annex L, Correspondence of Anatomic Region Codes and Body Part Examined — verified terms HEART, AORTA, BRAIN, KNEE, BREAST, PROSTATE, PHARYNXLARYNX, LARYNX, TONGUE, NECK. https://dicom.nema.org/medical/dicom/current/output/chtml/part16/chapter_L.html
• DICOM Standard Browser (Innolitics), Acquisition Contrast (0008,9209). https://dicom.innolitics.com/ciods/enhanced-mr-image/enhanced-mr-image/00089209
• DICOM Standard Browser (Innolitics), Percent Sampling (0018,0093) — "the fraction of acquisition matrix lines acquired, expressed as a percent"; the anchor for fully_sampled. https://dicom.innolitics.com/ciods/mr-image/mr-image/00180093
• DICOM Standard Browser (Innolitics), MR Receive Coil Sequence (0018,9042) and Multi-Coil Definition Sequence (0018,9045) — confirms there is no channel-count attribute, only element enumeration. https://dicom.innolitics.com/ciods/enhanced-mr-color-image/enhanced-mr-color-image-multi-frame-functional-groups/52009229/00189042
• BIDS, Magnetic Resonance Imaging data — suffixes T1w/T2w/PDw/FLAIR/dwi, parametric T1map/T2map, entities acq-/ce-/echo-/flip-/inv-/part-. https://bids-specification.readthedocs.io/en/stable/modality-specific-files/magnetic-resonance-imaging-data.html
• BIDS BEP001 — the community DICOM crosswalk for Parallel Acquisition Technique (0018,9078) and Parallel Reduction Factor In-plane (0018,9069). https://github.com/bids-standard/bep001/blob/master/src/04-modality-specific-files/01-magnetic-resonance-imaging-data.md
• ISMRMRD schema ismrmrd.xsd — trajectoryType enumeration (cartesian, epi, radial, goldenangle, spiral, other), sequenceParameters, acquisitionSystemInformation, measurementInformation. https://raw.githubusercontent.com/ismrmrd/ismrmrd/master/schema/ismrmrd.xsd

Dataset publications (source of the per-source protocol facts — field strength, coil counts, sequence names, sampling — that are not present in any committed CSV):

• fastMRI Breast — 288 spokes, base readout matrix 512, 83 partitions → 192 slices, 16-channel breast coil, 3 T MAGNETOM TimTrio, golden-angle radial VIBE, partial Fourier 6/8. Nyquist spokes ≈ (π/2)·512 ≈ 804, hence net R ≈ 2.8. https://pmc.ncbi.nlm.nih.gov/articles/PMC11791504/ · https://arxiv.org/abs/2406.05270
• USC 75-speaker speech rtMRI (Lim et al., Scientific Data 2021) — "The 13 spiral interleaves, when collected together, fulfil the Nyquist sampling rate"; TR 6.004 ms, TE 0.8 ms, FOV 200×200 mm, slice 6 mm, matrix 84×84, flip angle 15°, 8-channel upper-airway array, 1.5 T GE Signa Excite. https://www.nature.com/articles/s41597-021-00976-x
• M4Raw (Lyu et al., Scientific Data 2023) — 0.3 T Oper-0.3 (Ningbo Xingaoyi), four-channel head coil, 18 axial slices, 5 mm thick, 0.94×1.23 mm in-plane, T1w/T2w TSE and FLAIR, 183 volunteers. https://www.nature.com/articles/s41597-023-02181-4
• M4Raw GitHub repository / Zenodo record 8056074 — the M4RawV1.5 release notes and dataset card, source for the GRE contrast added after the original paper ("T1w Gradient echo (GRE) data for all 183 subjects are released"). https://github.com/mylyu/M4Raw
• CMRxRecon2024 dataset paper (Wang et al., Radiology: Artificial Intelligence 2025) — source for the T1/T2 mapping and BlackBlood pulse sequences and views: "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"; "the turbo spin-echo sequence was used for black-blood under breath hold", "black-blood with SAX view"; "'TrueFISP' sequence for cine, phase-contrast (flow2d), and tagging". https://pubs.rsna.org/doi/10.1148/ryai.240443 · https://pmc.ncbi.nlm.nih.gov/articles/PMC11950877/

Notes on a few source-specific field values verified against those publications:

• CMRxRecon T1/T2 mapping — SAX view for both T1map and T2map. T2 mapping is FLASH-based (spoiled gradient echo), not balanced SSFP.
• CMRxRecon BlackBlood — the paper gives the sequence (turbo spin echo) and view (SAX) but no T1/T2 weighting (no TE/TR), so contrast = :unknown reflects a genuine gap in the source rather than a guess.
• M4Raw "T1 GRE" — T1-weighted per the M4RawV1.5 dataset card, which post-dates the original Scientific Data paper.