HVSMR-2.0

Raw mirror of HVSMR-2.0, a 3D cardiovascular MR dataset for whole-heart segmentation in congenital heart disease, from the authors' figshare collection (10.6084/m9.figshare.c.7074755.v2; Pace et al., Scientific Data 11:721, 2024).

60 CMR scans from 60 patients (ages <1–52 y, mean 11.6), acquired at Boston Children's Hospital on a 1.5 T Philips Achieva: axial SSFP, free-breathing with respiratory navigator, prospective ECG gating. This is a static whole-heart acquisition, not cine. Many but not all patients received IV gadolinium, so contrast is mixed by design. Near-isotropic voxels (avg 0.73×0.73×0.81 mm). All volumes are PIR-oriented.

Classes — 8 foreground structures, strictly disjoint

Label	Structure	Label	Structure
0	background	5	AO — aorta
1	LV — left ventricle	6	PA — pulmonary artery
2	RV — right ventricle	7	SVC — superior vena cava
3	LA — left atrium	8	IVC — inferior vena cava
4	RA — right atrium

One integer per voxel — no nesting, no overlap. Verified across all 180 masks: the value union is exactly {0..8}.

There is no myocardium label. HVSMR 2016 segmented blood pool + ventricular myocardium; 2.0 deliberately dropped myocardium in favour of the 8 chambers and great vessels. Do not expect the two label schemes to be compatible.

13 of 60 cases have empty classes — this is anatomy, not annotation error

• RA (4) absent in 10 cases: pat 16, 20, 30, 39, 45, 47, 49, 50, 56, 58 — these are exactly the ten cases flagged CommonAtrium in hvsmr_clinical.csv (verified), where the common atrium is labelled LA.
• RV (2) absent in 6 cases: pat 50, 52, 56, 57, 58, 59 — single-ventricle cases, where the single ventricle is labelled LV. Five of the six are flagged SingleVentricle in the clinical CSV; the sixth is pat59, the one case with no clinical row at all.
• Both absent in 3: pat 50, 56, 58. The remaining 47 cases carry all 8.

The empty classes are therefore genuine anatomy, not missing annotation.

Macro-averaged Dice must skip classes absent from the ground truth rather than score them 0. present_classes / absent_classes in train.jsonl give this per case.

Three variants — pick one, do not mix

Directory	Image	Intensities	Size
orig/	full volume, cropped at the chin for de-identification	raw (~[0, 7500])	1.85 GB
cropped/	tight crop around the heart, same resolution	raw	0.47 GB
cropped_norm/	same crop as cropped	normalized (blood pool→0.8, lung→0.07, linear transfer; ~[−0.1, 3.3])	1.22 GB

The upstream paper states: "We do not recommend mixing data between the orig, cropped and cropped_norm directories." orig is the canonical unmodified variant and is what image/mask in train.jsonl point at.

Images are float64 except pat 17, 18, 19 and 30, which are uint16 in orig and cropped. Masks are int16 throughout.

⚠️ Filename trap. In cropped_norm/ the masks keep the _cropped_seg stem, not _cropped_norm_seg:

cropped_norm/pat0_cropped_norm.nii.gz          # image
cropped_norm/pat0_cropped_seg.nii.gz           # mask   <- not _cropped_norm_seg
cropped_norm/pat0_cropped_seg_endpoints.nii.gz # optional zone

A loader that derives the mask as {image_stem}_seg.nii.gz silently fails on this variant. Use the explicit paths in train.jsonl. The cropped_norm masks are byte-identical to the cropped masks (verified 60/60); upstream ships them duplicated and this mirror preserves that.

_endpoints files are a don't-care zone, not a second mask

Each case ships a third volume, *_seg_endpoints.nii.gz, using the same integers 1–8 but populating vessel labels only — measured across all 60 cases the values are {3 (LA/pulmonary veins), 5 (AO), 6 (PA), 7 (SVC), 8 (IVC)}, never LV, RV or RA.

It is not a subset of the ground truth. Only 22.8%–66.6% (mean 41.8%) of optional-zone voxels fall inside GT foreground — the zone straddles the vessel boundary, spanning from the minimum required extent out to the maximum tolerated extent. It is sizeable: 3.4%–15.3% (mean 8.3%) of GT foreground volume.

The official evaluation protocol: "the optional zone segmentations should be subtracted from both the ground truth and the predicted segmentation (in one-hot representations) before computing a segmentation score, so that only the required regions are compared." Skipping this step over-penalizes vessel predictions. Per the paper's Table 2: the aorta may optionally continue to the bottom of the image; the distal ~25% of each PA branch is optional; the superior ~25% of the SVC is optional; pulmonary veins may be shorter.

⚠️ Benchmark leakage — pat0–19 ARE the HVSMR 2016 challenge scans

From the paper: "The first 20 images of HVSMR 2.0 come from the original HVSMR dataset, which was held as a challenge at MICCAI 2016. However, the segmentations we now provide for these images are completely different."

HVSMR-2.0 includes all 20 HVSMR 2016 scans — both the 10 training and the 10 test cases. Same pixels; only the annotations are new. Any model pretrained, tuned or benchmarked on HVSMR 2016 leaks into pat0–19.

The hvsmr2016 column in train.jsonl (from upstream hvsmr_technical.csv) preserves the crosswalk: train = pat0–9, test = pat10–19, no = pat20–59. Filter hvsmr2016 == "no" for a clean 40-case subset.

No overlap with ACDC, MMWHS, M&Ms or CMRxMotion — different institutions, populations and sequences.

Splits

There is no official train/val/test split. The paper: "Users can split the 60 cases into training, validation and testing datasets at their discretion." Every row in train.jsonl is therefore split: "train".

Two reference splits are preserved as columns for comparability:

• pace_media2022 — the split used by Pace et al., Medical Image Analysis 80:102469 (2022): 4-fold CV (1–4, 12 cases each) plus 12 held-out test cases (pat 20, 27, 28, 31, 38, 40, 42, 46, 48, 49, 56, 59). This is the comparable choice if you need a published split.
• hvsmr2016 — the 2016 challenge partition (see leakage warning above).

Structure

orig/pat{0..59}_orig.nii.gz                     # image
orig/pat{0..59}_orig_seg.nii.gz                 # 8-class ground truth
orig/pat{0..59}_orig_seg_endpoints.nii.gz       # optional (don't-care) zone
cropped/pat{n}_cropped[_seg][_endpoints].nii.gz
cropped_norm/pat{n}_cropped_norm.nii.gz + pat{n}_cropped_seg[_endpoints].nii.gz
hvsmr_clinical.csv    # upstream, verbatim
hvsmr_technical.csv   # upstream, verbatim
train.jsonl           # 60 rows: paths for all 3 variants + shape/spacing/metadata

train.jsonl per row: sample_id, patient_id, variants (paths + shape + spacing + dtype for each of the three), convenience image/mask/endpoints pointing at orig, present_classes/absent_classes, hvsmr2016, pace_media2022, acquisition parameters (te_ms, tr_ms, flip_angle_deg, bandwidth_hz), and age_years / severity / diagnoses from the clinical CSV.

Upstream CSV quirks (both files are mirrored verbatim; train.jsonl is cleaned)

• hvsmr_technical.csv has a UTF-8 BOM on the Pat header and 77 trailing blank rows (137 lines, 60 real). Read with encoding="utf-8-sig" and drop blank rows.
• hvsmr_clinical.csv has 59 rows — pat59 is missing (verified against the upstream md5, so it is an upstream gap, not a corrupt download). pat59 is a single-ventricle case whose diagnosis row simply does not exist; its has_clinical_row is false.
• The paper refers to these files as hvsmr2_clinical.csv / hvsmr2_technical.csv; the shipped filenames have no "2".

The parquet under data/ is a browsing preview only (one axial slice per case — the one with the most GT foreground — rendered as image / mask / class overlay / optional-zone panel); train on the raw NIfTI volumes.

Orientation

All 180 volumes are stored PIR, so in the native array the anatomical axes are:

Array axis	Direction	Fixing it yields
0	P (anterior→posterior)	coronal
1	I (superior→inferior)	axial
2	R (left→right)	sagittal

A pipeline that reorients to RAS first (e.g. MONAI Orientationd(axcodes="RAS")) puts superior–inferior on axis 2, after which the conventional "slice along the last axis" gives axial slices. Slicing the native array along its last axis gives sagittal slices instead.

Annotation provenance

One consensus annotation per case — no multi-rater or auto-vs-manual tiers. An ensemble of four 3D U-Nets (trained on the 20 HVSMR 2016 whole-heart labels) produced initial labels, merged with manually placed 3D contours separating the 8 structures in SlicerHeart, then corrected in 3D Slicer (~4–8 h/case). QC after every step, review by an associate professor of pediatric radiology/cardiology, then final expert review of all 60 by a pediatric cardiologist. The 20 HVSMR 2016 cases were re-reviewed after the 40 new ones so annotation style stays consistent.

License & ethics

CC BY 4.0 — creativecommons.org/licenses/by/4.0, as declared on all three figshare items. Commercial use and derivatives are permitted with attribution.

The Boston Children's Hospital IRB approved this retrospective evaluation and waived written informed consent (IRB-P00011748), and determined that sharing the 60 de-identified scans under an open license is not inconsistent with that approval. De-identification includes cropping at the chin, so no facial features are present.

Citation

Pace DF, Contreras HTM, Romanowicz J, Ghelani S, Rahaman I, Zhang Y, Gao P, Jubair MI, Yeh T, Golland P, Geva T, Ghelani S, Powell AJ, Moghari MH. HVSMR-2.0: A 3D cardiovascular MR dataset for whole-heart segmentation in congenital heart disease. Scientific Data 11, 721 (2024). doi:10.1038/s41597-024-03469-9

If you use the pace_media2022 split, also cite:

Pace DF, Dalca AV, Brosch T, Geva T, Powell AJ, Weese J, Moghari MH, Golland P. Learned iterative segmentation of highly variable anatomy from limited data: Applications to whole heart segmentation for congenital heart disease. Medical Image Analysis 80, 102469 (2022).