fidopt Interactive view
Abdomen CT, segmentation

3D-IRCADb-01

Abdominal CT of an adult, three in four with liver tumours, the liver and its tumours segmented (3D-IRCADb-01, IRCAD Strasbourg).

10 test casesabdomenimage CTresearch only
Look at the images
Read before comparing (9)
  • Pool. 20 patients, 10 women and 10 men born 1929 to 1987, with liver tumours in 15. The test set holds tumours on 8 of its 10 patients; patient_14 (a resection site, not drawn as tumour) and patient_20 have none.
  • Images. The DICOM series as delivered, 512 x 512 pixels of 0.56 to 0.87 mm, slices 1 to 4 mm apart, 16-bit, in Hounsfield units. On 15 of the 20 the values stop at 1023, saturating bone and contrast; three pad outside the field with -2048. Most are contrast-enhanced, but the liver reads 27 HU on patient_16 and 50 HU on patient_18. The headers are anonymised: their sex and dates are not the page's. No mask is given.
  • Truth. IRCAD's page lists the segmented structures without saying who drew them. LiTS's reference for liver_28 to liver_47 is these masks, voxel for voxel once two axes are flipped (the liver at a Dice of 1.000 on all 20, the tumours on the 15 that have them; the other five have none on either side), and LiTS says each centre's scans were annotated by one rater of that centre. Graded expert with that limit: one rater, no agreement published.
  • Labels. Label 1 is the liver without its tumours, as MSD_Liver's; cysts stay in it. Label 2 holds 5.7 to 161 mL on the 8 test patients with tumours (1 to 46 tumours a patient on the page). On patient_04 (5 voxels) and patient_10 (218 voxels) a tumour mask reaches beyond the liver mask; it stays tumour. The vessels, the other organs, the bone and the skin, drawn on some patients only, are not read.
  • Room. The test livers (1,240 to 2,112 mL) moved one voxel keep a Dice of 0.95 to 0.99; the tumours 0.75 to 0.98.
  • Start. Labelling the whole body (above -500 HU) as liver scores a mean Dice of 0.23 on the test cases and 0.008 for the tumours; Otsu's bright voxels 0.24 and 0.009.
  • TotalSegmentator. Version 2.18 (task total) overlaps the whole liver, labels 1 and 2 together, at a Dice of 0.976 on patient_01 and 0.906 on patient_12, whose 161 mL of tumour it mostly leaves out of its liver; label 1 alone at 0.939 and 0.958. It finds the liver where the truth has it: the masks and the headers agree.
  • Overlap. These 20 CTs are LiTS's liver_28 to liver_47, with their geometry lost (1 mm in every direction); MSD_Liver, CTORG and CTSpine1K, built from LiTS's copies, leave those 20 out by name, and the label match above confirms the crosswalk. No built image of the corpus has the grid of any of the 20 (12,304 stores compared by size and spacing, a check that finds AbdomenDIRQA's case20 in MSD_Pancreas); those with 1 mm voxels and the same slice count are LiverPhasesCT's 8-bit CTs from Chongqing.
  • Outside numbers. None: LiTS scored on its own test scans, and papers on 3D-IRCADb-01 use their own splits.

No method has a result on every case of this dataset yet.

Case by case

Click a score to look at that result: the image, the truth in green and the method's map in red. Underlined scores have their image; for the others you see the case without the result.

About the data

Test cases
10 test patients drawn with seed 2026 among the 10 outside development, and 10 development patients, the last of the same seeded order: the database's 20 patients, one CT each. The provider publishes no split.
Truth
contours by experts the database's masks, as one label map: 1 the liver mask less its tumours, 2 the union of its liver tumour masks; the page does not say who drew them, and LiTS, which took these 20 scans in as liver_28 to liver_47, uses them voxel for voxel as its reference
Licence
CC BY-NC-ND 4.0
10 cases for tuning, never tested

patient_07, patient_03, patient_19, patient_06, patient_02, patient_15, patient_17, patient_05, patient_11, patient_13