3D-IRCADb — IRCAD Liver / Multi-Organ CT Segmentation (re-host)

Re-host of the official 3D-IRCADb database from IRCAD (subsets 01 + 02), converted from the original nested-DICOM release into a single dataset/case_XXXXX/ + train.jsonl layout (NIfTI volumes) matching the EasyMedSeg KiTS23 / SLIVER07 convention so one Base3DDataset subclass can load it.

Composition

Subset	Cases	Notes
01 (ircadb01)	20	abdominal CT, hepatic tumors in ~75% of cases
02 (ircadb02)	2	one patient, inspiration + expiration scans (respiratory motion)
total	22	all cases have masks; no official train/val/test split

case_00000..case_00019 are subset 01; case_00020..case_00021 are subset 02.

File layout

dataset/case_00000/
  imaging.nii.gz        # CT volume, int16 Hounsfield units
  segmentation.nii.gz   # uint8 multi-label mask (see label table)
...
train.jsonl
README.md
LICENSE.txt

train.jsonl lists one entry per case with image, mask, label (full canonical class list), labels_present (classes actually annotated in that case), modality, dataset, official_split, subset, patient_id. Image/mask paths are prefixed with data/nii/3D-IRCADb/ for the EasyMedSeg Base3DDataset.HF_JSONL_PREFIX convention.

Mask labels (canonical multi-label map)

The original release stores each anatomical structure as a separate binary DICOM stack (MASKS_DICOM/<structure>/), with inconsistent folder naming across patients (livertumor01/livertumor1/livertumors, liverkyst/liverkyste/livercyst, etc.). These were normalized to the canonical classes below and folded into a single integer-labelled volume. Structures were painted in label-ID order, so vessels and lesions (higher IDs) remain visible inside the organs that contain them. The imaging artifact class metal was dropped. Not every case contains every class — see labels_present per entry.

Value	Class
0	background
1	skin
2	bone
3	lung
4	heart
5	kidney
6	spleen
7	liver
8	gallbladder
9	pancreas
10	stomach
11	colon
12	smallbowel
13	bladder
14	uterus
15	adrenalgland
16	biliarysystem
17	lymphnodes
18	venoussystem
19	venacava
20	portalvein
21	artery
22	aorta
23	livercyst
24	livertumor
25	adrenaltumor
26	othertumor

CT voxel intensity

Hounsfield units are preserved (RescaleSlope/RescaleIntercept applied; stored int16). Per-volume spacing varies (0.56–0.96 mm in-plane; 1.0–4.0 mm slice); read it from each NIfTI affine rather than assuming a fixed value.

License

Re-host inherits the IRCAD terms: CC BY-NC-ND 4.0 (research / non-commercial, no derivatives). The upstream LICENSE.txt is reproduced at the repo root. See https://www.ircad.fr/research/data-sets/liver-segmentation-3d-ircadb-01/ .

@techreport{soler20103dircadb,
  title       = {3D image reconstruction for comparison of algorithm database:
                 A patient-specific anatomical and medical image database},
  author      = {Soler, Luc and Hostettler, Alexandre and Agnus, Vincent and Charnoz, Arnaud
                 and Fasquel, Jean-Baptiste and Moreau, Johan and Osswald, Anne-Blandine
                 and Bouhadjar, Mourad and Marescaux, Jacques},
  institution = {IRCAD, Strasbourg, France},
  year        = {2010}
}