Datasets:

MedOtter
/

Aortic-Dissection-Graz

license: cc-by-4.0
task_categories:
  - image-segmentation
tags:
  - medical
  - cta
  - ct
  - aorta
  - aortic-dissection
  - vascular
  - 3d
pretty_name: Aortic Dissection (Graz) - true/false lumen CTA
size_categories:
  - n<1K

Aortic Dissection (Graz) - CTA with true / false lumen expert annotations

Re-host of "Aortic Dissection Dataset and Segmentations" (figshare 10.6084/m9.figshare.22269091, CC BY 4.0) - 40 type-B aortic dissection CTA volumes collected in clinical routine 2005-2021 across the Medical University of Graz regional hospital network, with expert true-lumen and false-lumen annotations.

Mayer C, Pepe A, Hossain S, Karner B, Arnreiter M, Kleesiek J, Schmid J, Janisch M, Fuchsjaeger M, Deutschmann H, Zimpfer D, Egger J, Maechler H. Type B Aortic Dissection CTA Collection with True and False Lumen Expert Annotations for the Development of AI-based Algorithms. Scientific Data 11, 596 (2024). doi:10.1038/s41597-024-03284-2

Ethics: Medical University of Graz EK-34-161 ex 21/22; consent waived (retrospective), head and face cropped for anonymisation.


Modality	CTA, arterial phase (not ECG-gated)
Cases	40, one volume each
Classes	`0` background, `1` true lumen, `2` false lumen
Split	none - single cohort of 40 (`split="train"` for all rows)
Shape	mostly 512x512xZ; in-plane spacing 0.525-0.965 mm, through-plane 1.25-3.0 mm
Foreground	TL 125,516-659,055 voxels; FL 25,242-496,428 voxels; no empty masks

There is no thrombus class. Per the paper, "clearly thrombosed sections have not been segmented"; calcifications and atherosclerotic changes are likewise excluded. Ambiguous dark regions judged to be late-filling artifact were segmented as lumen. If you need a thrombus label, see ImageTBAD instead.

Layout

dataset/ctaNNs/ctaNNs.nrrd           image, int16 HU - byte-identical to figshare
               truelumenNN.seg.nrrd  true lumen,  uint8 {0,1}
               falselumenNN.seg.nrrd false lumen, uint8 {0,1}
               maskNN.nrrd           merged {0=bg, 1=TL, 2=FL}
               meshNN.stl            surface mesh - byte-identical to figshare
original_masks/ctaNNs/*.seg.nrrd     untouched source masks (~13 MB total)
train.jsonl                          40 rows (schema below)

train.jsonl per row: case_id, case_num, image, mask, mask_true_lumen, mask_false_lumen, mesh, original_mask_*, shape, spacing_mm, coarse_axis, tl_voxels, fl_voxels, tl_volume_ml, fl_volume_ml, tl_fl_overlap_voxels, fl_label_value_raw, resampled_to_image_grid, annotator_tier, split.

What was normalised here (and why)

Everything below is flagged per-case in train.jsonl, and the untouched source masks ship under original_masks/.

False-lumen masks binarised to {0,1}. 17 of 40 cases store the false lumen with label value 2, the other 23 with 1 - cases 4, 5, 11, 12, 13, 14, 15, 16, 17, 18, 22, 25, 26, 30, 35, 36, 38. Code doing mask == 1 against the raw files silently drops those 17 false lumens. Original value kept in fl_label_value_raw.
Case 40 masks resampled onto the image grid. As shipped, case 40's masks sit on a grid rotated ~6.85 deg about Z (512x1103x187 @0.545 mm) while its image is axis-aligned (512x512x401 @0.602 mm) - index-wise pairing raises IndexError. Nearest-neighbour world-space resampling retains 100.01% / 100.04% of physical volume, and mean HU inside the resampled masks is 409 HU (TL) / 339 HU (FL), i.e. contrast-filled arterial lumen and in line with the other 39 cases. Flagged by resampled_to_image_grid. Every other case's masks already match their image exactly.
Merged maskNN.nrrd added, painted FL first then TL, so the true lumen wins ties - matching the authors' own converter apepe91/AD_NRRD_TO_STL (cited in the paper's Code Availability), which subtracts a dilated true lumen from the false lumen. Ties are negligible either way: 16 of 40 cases overlap by 1-106 voxels, at most 0.028% of the TL-or-FL union - flap-boundary jitter, not a nested structure. Because ties are resolved, the stored map is strictly disjoint, so mask == 1 / mask == 2 are exact. The two source binaries remain available if you prefer to fan out.

Deliberately not changed: no reorientation, no resampling of the other 39 cases, no intensity windowing, original filenames kept.

Gotchas

Key masks off the FILENAME, never the segment name. Internal Slicer names are inconsistent: case 2's true lumen is called artery, case 19's false lumen is also called artery, and elsewhere you find true lumen 5, true_lumen, truelumen, True Lumen neu, flase lumen, False_Lumen.
Cases 19 and 20 are sagittally reformatted: the coarse 1.5 mm axis is NRRD axis 0, not axis 2 (all other 38 cases have it at axis 2). A hard-coded z-slicer emits ~950 thin reformats instead of ~240 native slices. Use coarse_axis from train.jsonl.
No official split. Group any split you make on case_id; one patient owns exactly one volume, so a plain case-level split is safe.
Anisotropic, up to 3.0 mm through-plane.
The images are the anonymised, head/face-cropped reconstructions. The separate raw CTs.zip on figshare is not mirrored here on purpose: its case 24 is a different reconstruction (776x776x101 oblique @0.625/0.625/2.4 mm) that does not align with the masks.

Annotation provenance - one tier, all of it gold

All 40 masks were seeded semi-automatically (local thresholding / Grow Cut, fill-between-slices interpolation, or region growing - the paper's Table 2 gives the per-case method), then given 2-3 hours of manual slice-by-slice paint/erase, because the paper reports the semi-automated results "were not precise enough for AI applications". A radiologist then checked every case.

annotator_tier records who did the manual work: 15 cases (1-10, 26, 27, 28, 29, 39, the paper's Table-1 nabla marker) by a medical student, the other 25 by cardiac-surgery residents. The paper marks these deliberately to show there is no difference between software users - so all 40 are gold and none should be filtered; the column is there for stratification only.

Measured vs published volumes. The tl_volume_ml / fl_volume_ml values here are measured directly from the mask files. They reproduce the paper's Table 2 to within +/-1% for 35 of 40 cases (most within 0.1%), which independently confirms the image-mask pairing and the binarisation. Five cases (5, 6, 7, 32, 36) disagree in the true-lumen column only; in three of them the paper's TL cell exactly equals its FL cell, which looks like a transcription slip. Each mask file contains exactly one correctly-named segment, so this is a table artifact, not extra structures in the data. Trust the values here.

Related datasets / leakage

No patient overlap with MedOtter/SegA (the AVT collection: KiTS 20 + RIDER 18 + Dongyang 18 - US and Chinese cohorts). The two share authors (Pepe, Egger), not patients, and AVT's few incidental dissection cases carry only a single binary aortic-vessel-tree mask, not lumen labels. No overlap with AortaSeg24 (Univ. of Florida), ImageTBAD (Guangdong), TotalSegmentator (Basel), or any TCIA collection. No cross-reference ID to other datasets exists; case_id (cta01-cta40) joins only to Tables 1 and 2 of the paper.

License

CC BY 4.0 - redistribution and commercial use permitted with attribution. Cite the Scientific Data paper above and the figshare record.