MedOtter/CETUS2014

CETUS 2014 — Endocardial Three-dimensional Ultrasound Segmentation

The MICCAI 2014 CETUS challenge: segment the left-ventricular endocardial surface in 3-D transthoracic echocardiography, at end-diastole (ED) and end-systole (ES).

This is 3-D echo, not 2-D. Where CAMUS gives you a 2-D apical plane, CETUS gives a full pyramidal volume per cardiac phase — so the LV cavity is segmented as a solid, and the ED/ES pair yields stroke volume and ejection fraction directly.

What this mirror contains — read first

⚠️ ED and ES only. The full cardiac sequences are not in this release. The 2014 challenge distributed complete cine loops as MHD/RAW through the now-decommissioned MIDAS server. The 2022 CREATIS Girder re-release mirrored here ships exactly two frames per patient — ED and ES — as NIfTI. There are no intermediate frames and no ED_ES_time metadata anywhere in the archive. Verified: all 45 patient folders contain exactly 4 files, zero exceptions. Anything requiring the cardiac cycle cannot be reproduced from this release.

⚠️ …but ground truth is now public for all 45 patients. The challenge withheld the 30 test masks for its leaderboard. This release includes a _gt.nii.gz for every patient and both phases — 90 annotated volumes. So relative to 2014 this is fewer frames but more labels.

⚠️ There is no train/test split — and the original 15/30 assignment is not recoverable. The archive is a flat list of 45 patient folders with no split file, no Training//Testing/ directories, and empty Girder metadata on every item. This mirror therefore ships one train split containing all 45 patients and declares a split fallback, rather than inventing a boundary. Any published "CETUS test set" number refers to a partition this release does not identify.

⚠️ Beware third-party mirrors. zeahub/cetus-miccai-2014 reformats to HDF5 and carves out a test/ split (patients 39–45) that appears to drop ground truth — but all 45 patients have GT here, so that split is invented and its "test" set discards usable labels. Prefer this mirror or the CREATIS source.

Dataset Details

Field	Value
Modality	3-D transthoracic echocardiography (B-mode), volumetric
Body part	Heart — left ventricle, endocardial surface
Target	LV endocardium (blood pool) — single binary structure
Cases	45 patients × 2 phases (ED, ES) = 90 annotated volumes
Centres	Rennes University Hospital · University Hospitals Leuven · Erasmus MC Rotterdam
Vendors	GE Vivid E9 · Philips iE33 · Siemens SC2000 (per-patient vendor not published)
Volume shapes	33 distinct; 200–380 × 191–424 × 176–348
Slices (z)	176–348 per volume, median 231
Spacing	0.5763 mm isotropic — see the units warning below
Format	.nii.gz; images and masks both float32
Mask values	{0.0, 255.0} — not {0,1}, not uint8
Split	none upstream — all 45 patients in train
License	CC BY-NC-SA 4.0 — shipped inside the archive itself
Paper	Bernard et al., IEEE TMI 35(4):967–977, 2016 · doi:10.1109/tmi.2015.2503890

⚠️ The header declares millimetres but stores metres

The single easiest thing to get wrong with this dataset.

Every one of the 90 volumes has:

xyzt_units = 2          # NIfTI code 2 == MILLIMETRES
pixdim     = (5.763e-4, 5.763e-4, 5.763e-4)
affine     = diag(-5.763e-4, -5.763e-4, +5.763e-4), zero translation
sform_code = 1, qform_code = 1

Taken literally that is 0.00058 mm — 0.58 micrometres per voxel, i.e. a heart the size of a bacterium. The values are metres; the true spacing is 0.5763 mm isotropic, uniform across every patient and every axis.

Confirmed physiologically rather than by assertion — computing LV volume from the ground-truth voxel counts under the metres reading gives:

	min	median	max
EDV (mL)	79.9	139.5	407.9
ESV (mL)	36.0	87.8	352.7
EF (%)	13.5	38.2	56.7

Textbook values for a cardiac cohort (14/45 patients have EF < 30 %, 11/45 have EDV > 200 mL — CETUS deliberately included impaired and dilated ventricles). Under the literal-millimetres reading the same voxels give EDV ≈ 2.9 × 10⁻⁷ mL.

Dice and other overlap metrics are unaffected. Anything in physical units — volume, ejection fraction, Hausdorff distance, mm-based resampling — is wrong by 10³ per axis unless you override the spacing. The headers are deliberately NOT patched here so this mirror stays byte-identical to the official release; the corrected spacing is recorded per case in train.jsonl as spacing_mm.

⚠️ Masks are float32 {0.0, 255.0}

Both the image and the mask are stored float32, and the mask's header is byte-identical to its image's header in all 90 pairs. Foreground is 255.0, not 1.

Binarize on the raw values (arr > 0). Measured over all 90 masks: exactly two unique values, no intermediate/anti-aliased voxels, no empty masks, and the image/mask grids always agree. A {0,255} binary mask survives a min–max→>0.5 recipe, but reading the raw values is the honest operation and is what the official notebook does.

All 90 images are likewise integral and within 0–255 despite the float32 container, so they are losslessly uint8-representable — the float32 storage costs 4× for nothing. Files are mirrored as-is regardless.

Ground truth

One reference mask per volume — there is no tier or rater to choose. The challenge's evaluation protocol involved three expert observers to establish inter-observer variability, but the distributed _gt.nii.gz is a single consensus reference, and it is the only annotation released.

Measured over all 90 masks:

	min	median	max
Foreground fraction of volume	1.20 %	3.35 %	6.84 %

• Zero empty masks — every one of the 90 volumes has annotated foreground.
• ED cavity > ES cavity in 45/45 patients, as physiology requires. This is a free correctness check on the phase labelling, and it passes without exception.

Choosing a slicing axis — axis 2 (z)

For 2-D slice-wise use, axis 2 is both the anatomically correct stack and the empirically best choice. It is the beam/depth axis, running apex → base, so slicing it yields the conventional short-axis cross-sections; axes 0 and 1 yield long-axis (apical) views.

Measured over all 90 masks:

Axis	View	Slices containing foreground (median)	Leading background slices (median)
0	long-axis	31.6 %	102
1	long-axis	32.7 %	88
2 (z)	short-axis	64.4 %	41

Axis 2 roughly doubles the fraction of useful slices and more than halves the leading empty run — which matters for any pipeline that samples slices at random and gives up after a bounded number of attempts.

Note the organizers' own script_cetus.ipynb (carried over here) visualises along a long-axis plane. That is a display choice, not a processing convention.

Structure

train/images/patient01_ED.nii.gz    #  90 B-mode volumes (45 patients x ED/ES)
train/masks/patient01_ED.nii.gz     #  90 masks, same grid, values {0., 255.}
train.jsonl                         # per-case metadata (90 rows)
manifest.csv                        # sha256 + bytes + shape for all 180 originals
script_cetus.ipynb                  # organizers' official reader, verbatim
LICENSE_TERMS.md                    # from the archive, verbatim
MANDATORY_CITATION.md               # from the archive, verbatim
README.md
LICENSE.txt

Case IDs are patient01_ED … patient45_ES; patient folders upstream are lowercase, zero-padded 2-digit (patient01, not Patient1 — that was the 2014 MHD release).

Group on patient_id, not case_id. Each patient contributes two rows (ED and ES) of the same heart on the same grid. Splitting them across a train/test boundary leaks.

train.jsonl columns:

Column	Meaning
case_id	"patient01_ED" — unique per row
patient_id	"patient01" — the grouping key; 2 rows share it
phase	"ED" or "ES"
image, mask	repo-relative paths
split	always "train" (no upstream split exists)
shape_xyz, n_slices	geometry; n_slices is the axis-2 extent
spacing_mm	[0.5763, 0.5763, 0.5763] — corrected, use this
pixdim_raw, xyzt_units_code	what the header literally says (the defect)
axcodes, sform_code, qform_code	header provenance
image_dtype, mask_dtype	both float32
intensity_min, intensity_max	per-volume; 88/90 span 0–255
image_fits_uint8	true for all 90
mask_values	[0.0, 255.0] for all 90
fg_voxels, n_voxels, foreground_fraction	cavity size
lv_volume_ml	fg_voxels x 0.5763^3 / 1000 — corrected spacing
ef_percent	patient-level EF from this patient's ED/ES pair
fg_slice_fraction, leading_bg_slices	per-axis ("0","1","2") coverage
image_sha256, mask_sha256, image_bytes, mask_bytes	fidelity to source

Overlap and contamination

• CAMUS — believed disjoint, but not provably so. CAMUS is single-centre (Saint-Étienne) 2-D echo; CETUS is three-centre (Rennes / Leuven / Rotterdam) 3-D echo. Different modality, cohorts and scanners. No author asserts disjointness, and both releases use positional anonymous IDs (patient01…), so overlap is unverifiable by ID even in principle. What the two genuinely share is authors (Bernard, D'hooge, Pedrosa), not patients. Note the ID schemes collide — namespace by dataset key.
• ⚠️ CETUS2014 appears in the IMed-361M / IMIS-Net training corpus (arXiv 2411.12814, Table 4, cited to the CREATIS challenge URL, 90/10 split). Benchmarking IMIS-Net on CETUS is contaminated.
• Clean with respect to MedSAM, SAMUS/US30K, UltraSam/US-43d, MedSAM2, SAM-Med2D, SAM-Med3D and BiomedParse — those corpora absorbed CAMUS, not CETUS. "CETUS" appears zero times in MedSAM's supplement.
• No overlap with EchoNet-Dynamic (Stanford), MITEA, or the Medical Segmentation Decathlon (which contains no echocardiography at all).

Known per-case notes

• patient29 is the only intensity outlier: max 168 (ED) / 190 (ES) rather than 255 — a dimmer acquisition, not a truncated file.
• Volume shape is unique per patient (33 distinct shapes over 45 patients); ED and ES always share their patient's shape.

Source & Citation

• Official: CREATIS Human Heart Project Girder, collection 62eb991b73e9f0048c3a6c45 — served anonymously, no account required. https://humanheart-project.creatis.insa-lyon.fr/database/
• The same server hosts the official ACDC, CAMUS and TED releases.

@article{bernard2016cetus,
  author  = {Bernard, Olivier and Bosch, Johan G. and Heyde, Brecht and
             Alessandrini, Martino and Barbosa, Daniel and Camarasu-Pop, Sorina
             and Cervenansky, Frederic and Valette, Sebastien and Mirea, Oana
             and Bernier, Michel and Jodoin, Pierre-Marc and Domingos, Joao S.
             and Stebbing, Richard V. and Keraudren, Kevin and Oktay, Ozan and
             Caballero, Jose and Shi, Wenzhe and Rueckert, Daniel and
             Milletari, Fausto and Ahmadi, Seyed-Ahmad and Smistad, Erik and
             Lindseth, Frank and van Stralen, Marijn and Wang, Chen and
             Smedby, Orjan and Donal, Erwan and Monaghan, Mark and
             Papachristidis, Alexandros and Geleijnse, Marcel L. and
             Galli, Elena and D'hooge, Jan},
  title   = {Standardized Evaluation System for Left Ventricular Segmentation
             Algorithms in {3D} Echocardiography},
  journal = {IEEE Transactions on Medical Imaging},
  volume  = {35},
  number  = {4},
  pages   = {967--977},
  year    = {2016},
  doi     = {10.1109/TMI.2015.2503890}
}