EMIDEC — Myocardial Infarction from Delayed-Enhancement Cardiac MRI

The MICCAI 2020 EMIDEC challenge: segment the left ventricle and characterise myocardial infarction in delayed-enhancement (late gadolinium) cardiac MRI. Acquired at the University Hospital of Dijon (ImViA Lab, Université de Bourgogne Franche-Comté).

Unlike most cardiac MRI benchmarks — which segment anatomy on cine images — EMIDEC targets tissue pathology on a static DE-MRI stack: normal myocardium vs. infarcted myocardium vs. the no-reflow (microvascular obstruction) core.

What this mirror contains — read first

The test set has no ground truth, and never did publicly. The official release ships 100 training cases with masks and 50 test cases with images only — the organizers withheld test GT for the leaderboard and scored submissions by email. This is confirmed structurally in the archive (zero Contours/ directories in the test archive) and in the test archive's own Readme.txt. The 50 test cases are mirrored here as test_unlabeled for completeness; they cannot be used to compute Dice.

Two different test sets exist upstream. The segmentation and classification contests did not use the same 50 patients ("the cases are not the same for the two contests"; cases were added and removed for the classification split). This mirror carries the segmentation test set.

Beware third-party mirrors. Seven Kaggle/HF re-uploads exist. All mislabel the license (MIT / Apache-2.0 instead of CC BY-NC-SA 4.0), and the HuggingFace one (viennh2012/cardiac_cine_emidec) is a corrupted derivative: images per-slice min–max normalised to uint8 (every slice max becomes exactly 254, destroying the cross-slice intensity relationship that is the DE-MRI signal), masks damaged by interpolated resampling, and the 100 training cases re-split into its own train/val/test — so its "test" set is official training data. It is also misnamed "cine"; EMIDEC is not cine. Use this mirror or emidec.com.

Dataset Details

Field	Value
Modality	Delayed-enhancement cardiac MRI (DE-MRI / LGE), T1-weighted PSIR, phase-sensitive image only
Acquisition	~10 min after Gd-DTPA 0.1–0.2 mmol/kg; TR 3.5 ms, TE 1.42 ms, TI 400 ms, flip 20°; ECG-gated, breath-hold
Scanners	Siemens Aera 1.5 T and Skyra 3 T
Body part	Heart — left ventricle only, short-axis base→apex (RV is not annotated)
Cases	150 patients — 100 train (33 normal + 67 pathological, with GT) + 50 test (images only)
Slices	708 across the 100 training cases; 4–10 per case
In-plane	1.367–1.875 mm; 92 distinct in-plane shapes across 100 cases
Slice gap	8.0 / 8.9 / 9.6 / 10.0 / 13.0 / 13.04 mm — 10 mm in only 66 of 100 cases
Format	.nii.gz; images float64, masks uint8 (86) / uint16 (14)
Intensity	Integral 12-bit values, global range 0–4095, per-case max 3684–4095 (no upstream normalisation)
License	CC BY-NC-SA 4.0 — stated on emidec.com and inside the archives
Paper	Lalande et al., Data 2020, 5(4):89 · doi:10.3390/data5040089

Label encoding

0 background · 1 cavity · 2 normal myocardium · 3 infarction · 4 no-reflow (PMO/MVO)

Papillary muscles are included in the cavity (label 1, not myocardium). The epicardial border excludes fat, and the contour is extrapolated across the junction of the two ventricles.

⚠️ The archives' own Readme.txt lists the classes in the wrong order. Its prose reads "background, myocardium, cavity, myocardial infarction and no-reflow", which implies 1=myocardium and 2=cavity. That is wrong — it contradicts the organizers' Evaluation-metrics code ({"background":0,"cavity":1,"normal_myocardium":2,"infarction":3,"NoReflow":4}). Verified here from the voxels across all 100 training masks: label 1 sits at mean in-plane radius 10.3 px from the heart centroid while label 2 sits at 18.8 px, and in 432 slices more than half of label 1 falls inside the filled holes of label 2 (zero slices vote the other way). Label 2 is the ring; label 1 is the blood pool. The eval code is right, the Readme prose is not.

⚠️ Evaluation targets are nested unions, not one-vs-rest

Straight from the organizers' main.py. Getting this wrong is the most common EMIDEC mistake:

Target	Label set	Cases with non-empty GT (of 100)
cavity	{1}	100
myocardium	{2, 3, 4}	100
infarction	{3, 4}	67
no_reflow	{4}	40

The no-reflow core is inside the infarct, and the infarct is inside the myocardial wall. Reported infarct volumes likewise include PMO. Dice is computed in 3D over the whole volume, not averaged per slice.

Verified label properties (measured on all 100 training masks)

• Only three label sets occur: {0,1,2} ×33, {0,1,2,3} ×27, {0,1,2,3,4} ×40. Always a contiguous prefix — there is no non-contiguous-label problem.
• The 33 cases with {0,1,2} are exactly the 33 normal (N) cases. Labels 3 and 4 are absent from every normal case.
• Label 3 present in 67/100 (all pathological), label 4 in 40/100.
• Label 4 is tiny: 13–2976 voxels per case (median 88). Expect near-zero no-reflow Dice from any naive model.
• Foreground is 2.92 % of voxels on average (range 1.30–5.54 %).
• All 708 slices contain foreground — only slices with visible myocardium were released, so the stack is not full-thorax coverage.

⚠️ Spacing lives ONLY in pixdim — the affine does not encode it

The single easiest thing to get wrong with this dataset.

Every one of the 100 training images has:

sform_code = 2   (ALIGNED_ANAT — i.e. declared valid)
qform_code = 0
affine     = diag(-1, -1, 1)      # LPS axcodes, zero translation, NO SCALE
pixdim     = (1.367-1.875, same, 8.0-13.04)   # the real spacing

The NIfTI spec says to use the sform when sform_code > 0, so any code that reads spacing straight off the affine gets 1 mm isotropic and silently mis-scales every physical-unit result: infarct volume, Hausdorff distance, and any resampling. With a true slice gap of 8–13 mm that is an order-of-magnitude error through-plane. The common form of this bug is np.diag(nib.load(p).affine)[:3], or anything deriving spacing from affine[:3,:3].

The two major toolkits both handle it correctly, so this is a hazard for hand-rolled readers rather than for MONAI/ITK pipelines — verified on this mirror:

Reader	Result
nibabel raw .affine	diag(-1, -1, 1) → wrong, 1 mm isotropic
nibabel header.get_zooms()	(1.5625, 1.5625, 10.0) → correct
MONAI LoadImaged / NibabelReader	rebuilds the affine from pixdim → diag(-1.5625, -1.5625, 10.0), correct
SimpleITK	warns has unexpected scales in sform, then recovers from pixdim → correct

These files are mirrored byte-identically and the header is deliberately NOT patched, so this mirror stays comparable to the official release and to published EMIDEC results. Instead, the true spacing is recorded per case in train.jsonl / test_unlabeled.jsonl (spacing_xyz, slice_gap_mm). Read spacing from header.get_zooms() or from the jsonl — never from the affine.

Geometry: the heart is centred but tiny — and not cropped

A widely repeated claim says EMIDEC images are cropped around the heart. They are not. The field of view is full-thoracic (e.g. 216 × 359 mm) and the heart occupies only ~2–5 % of the in-plane area. What upstream actually did is in-plane re-registration: "the slices are realigned according to the gravity centre of the area defined by the epicardial contour" — so the foreground centroid sits essentially on the image centre.

That is why a centre crop is the safe preprocessing choice here, and why every challenge entrant used one. It is a consequence of the re-centering, not of any upstream cropping.

Measured on this mirror: a 144 × 144 centre crop of the middle slice retains 100 % of the ground-truth foreground in all 100 training cases — not a single case loses a voxel. So a fixed centre crop is not merely conventional here, it is lossless at that size.

⚠️ Possible patient overlap with ACDC — unverifiable

EMIDEC and ACDC share a great deal: the same hospital (CHU Dijon), the same senior author, the same two annotators described identically ("two independent experts, 10 and 20 years of experience, reaching consensus"), the same n = 150, the same 100/50 split, and the same Siemens Aera 1.5 T. Decisively, EMIDEC was extracted from "a conventional cardiovascular exam [that] included cine-MRI and DE-MRI" — the very exam family ACDC's cine images come from. ACDC also contains a 30-patient MINF group (prior myocardial infarction).

Neither release carries a cross-reference ID, both are fully anonymised with DICOM headers stripped, and acquisition years are not published — so overlap cannot be checked or excluded by any downstream user, and no xref column could be preserved in this mirror.

Partial mitigation: EMIDEC is acute MI (imaged within ~1 month of angioplasty) with pathological LVEF 47.7 ± 13.2 %, mostly above ACDC's MINF threshold of < 40 %; and EMIDEC's 3 T cases used a Skyra where ACDC used a Trio Tim, suggesting a later acquisition window.

Do not place EMIDEC and ACDC on opposite sides of a train/test split in a combined cardiac benchmark, and caveat any claim of independence between them.

No overlap with MyoPS 2020 / MyoPS++ / MS-CMRSeg (Shanghai Renji cohort), LAScarQS (left atrium), CMRxMotion (Fudan volunteers), or M&Ms (multi-centre ES/DE/CA) — different institutions and cohorts; author overlap only.

Ground truth

A single gold-standard tier, which is also the only mask released:

Pass	Who	Role
1	Cardiologist, 10 yr experience	Drew all contours manually in QIR (CASIS, Quetigny)
2	Biophysicist, 20 yr cardiovascular MRI	"went through every outline and made some changes when necessary"

The expert-2-revised contours are the leaderboard reference. Manual (not semi-automatic) was deliberate: "questionable contours ... are only due to the choice of the experts and do not depend on algorithm settings."

Human ceiling (measured by the authors on 34 other cases, not released):

	Myocardium	Myocardial infarction
Intra-observer Dice	0.84	0.76
Inter-observer Dice	0.83	0.69

A model at infarct Dice ≈ 0.7 is already at inter-observer level. No multi-rater masks were released for the 150 cases.

Structure

train/images/Case_XXXX.nii.gz            # 100 DE-MRI volumes  (N###/P###)
train/masks/Case_XXXX.nii.gz             # 100 masks, same grid, values 0-4
test_unlabeled/images/Case_NNN.nii.gz    #  50 volumes, 101-150, NO masks exist
train.jsonl                              # per-case metadata
test_unlabeled.jsonl
clinical_metadata.csv                    # all 150 cases, 13 clinical fields
README.md
LICENSE.txt

Case IDs are the official ones. Train uses Case_N### / Case_P### where the number is a global sequence 001–100 and the letter is the class; test uses Case_101–Case_150.

The N/P prefix leaks the pathology label. It is retained for fidelity with the official release and the leaderboard, but a classifier that reads case filenames scores 100 % locally and 50 % on the real (numerically-named) test set. Use the pathology column deliberately, not the filename.

The parquet preview layer is display-only

data/*.parquet exists so the HF Dataset Viewer can render this dataset. Each row holds the middle slice of one volume as PNG: image (full-FOV, grayscale), mask (class-coloured), overlay, and overlay_zoom (the 144 × 144 centre crop, upscaled). Colours are 1 cavity blue, 2 normal myocardium green, 3 infarction yellow, 4 no-reflow red.

Do not train or evaluate on the preview. Its intensities are percentile-windowed to 8-bit for display and it holds one slice per case. Treating a rendered preview as the data is exactly the error that makes the third-party mirror unusable. The real data is the byte-identical .nii.gz at the repo root.

For test_unlabeled, mask / overlay / overlay_zoom / pathology and every label-derived column are null, because no test ground truth exists.

train.jsonl / test_unlabeled.jsonl columns:

Column	Meaning
case_id	"N006" … "P100" (train), "101" … "150" (test)
case_name	"Case_N006" — the official directory name
image, mask	repo-relative paths (mask is null for test_unlabeled)
split	"train" or "test_unlabeled"
pathology	"normal" or "pathological" (train only; null for test)
n_slices, shape_xyz	geometry — 92 distinct in-plane shapes
spacing_xyz, slice_gap_mm	true spacing, from pixdim (not the affine)
slice_gap_mm_declared	gap as written in the clinical txt — disagrees with pixdim in 6/100 cases
slice_gap_matches_pixdim	false for those 6
axcodes, sform_code, qform_code	header provenance for the spacing caveat
image_dtype, mask_dtype	float64; uint8 or uint16
intensity_min, intensity_max	per-case (never renormalised)
label_values	sorted labels present, e.g. [0,1,2,3]
label_voxels	{label: voxel_count} for 0–4
target_voxels	{cavity, myocardium, infarction, no_reflow} under the nested unions
has_infarction, has_no_reflow	booleans
foreground_fraction	fraction of voxels with label > 0
clinical	the 13 parsed clinical fields (see below)

Clinical metadata

Every case — all 100 train and all 50 test — ships a clinical text file, and all 13 fields are populated in 150/150 cases. Parsed into clinical_metadata.csv and the clinical object in the jsonl.

Field	Type	Domain over all 150
sex	categorical	M 89, F 61
age	int	27–89
tobacco	int	1 ×53, 2 ×32, 3 ×65 — see caveat
overweight (BMI > 25)	bool	Y 84, N 66
arterial_hypertension	bool	Y 61, N 89
diabetes	bool	Y 20, N 130
familial_history_cad	bool	Y 14, N 136
ecg_st_elevation	bool	Y 95, N 55
troponin	float (ng/mL)	0.1–420
killip_max	int	1 ×121, 2 ×23, 3 ×3, 4 ×3
lvef_percent	float (%)	20–70 — echocardiographic, field is labelled FEVG
ntprobnp	float (pg/mL)	3–22577
slice_gap_mm_declared	float (mm)	8 ×5, 10 ×120, 13 ×25

Parsing gotchas handled here, all of which bite a naive reader:

• The text files are ISO-8859-1 with CRLF, not UTF-8, and several keys carry a stray non-ASCII byte before the colon (Gap between slices :).
• Filenames use a space (Case N058.txt) while the image directories use an underscore (Case_N058/) — joining them naively fails.
• One Troponin value uses a French decimal comma (4,5) and must be normalised before float().
• Each archive also contains a Readme.txt at top level that is not a case.
• tobacco is provided as a bare 1/2/3. The descriptor paper describes this as yes / no / former, but the mapping is not stated in the data itself, so the raw integer is preserved here rather than a guessed decoding.
• slice_gap_mm_declared disagrees with the NIfTI pixdim in 6 of 100 training cases (N070 10→8.9, P008 10→9.6, P043 13→10.0, P059 10→8.0, P095 10→13.0, P097 10→13.0). The clinical text is a hand-entered note; pixdim is authoritative for geometry.

Source & Citation

• Official: https://emidec.com/dataset — CC BY-NC-SA 4.0. The site asks for a free self-service account; the archives themselves are served by an unauthenticated form POST.
• Organizers' evaluation code (the authority on label values and the nested eval groups): https://github.com/EMIDEC-Challenge/Evaluation-metrics
• Challenge design: https://doi.org/10.5281/zenodo.3755234 (PDF only — the data is not on Zenodo)

@article{lalande2020emidec,
  author  = {Lalande, Alain and Chen, Zhihao and Decourselle, Thomas and
             Qayyum, Abdulkadir and Pommier, Thibaut and Lorgis, Luc and
             de la Rosa, Ezequiel and Cochet, Alexandre and Cottin, Yves and
             Ginhac, Dominique and Salomon, Michel and Couturier, Raphael and
             Meriaudeau, Fabrice},
  title   = {Emidec: A Database Usable for the Automatic Evaluation of
             Myocardial Infarction from Delayed-Enhancement Cardiac MRI},
  journal = {Data},
  volume  = {5},
  number  = {4},
  pages   = {89},
  year    = {2020},
  doi     = {10.3390/data5040089}
}

@article{lalande2022deep,
  author  = {Lalande, Alain and Chen, Zhihao and Pommier, Thibaut and
             Decourselle, Thomas and Qayyum, Abdulkadir and Salomon, Michel and
             Ginhac, Dominique and Skandarani, Youssef and Boucher, Arnaud and
             Brahim, Khawla and de Bruijne, Marleen and others},
  title   = {Deep learning methods for automatic evaluation of delayed
             enhancement-MRI. The results of the EMIDEC challenge},
  journal = {Medical Image Analysis},
  volume  = {79},
  pages   = {102428},
  year    = {2022},
  doi     = {10.1016/j.media.2022.102428}
}