fidopt Interactive view
Lesions and tumours, segmentation

TotalSegmentator liver_lesions, CT

Contrast-enhanced abdominal CT of a University Hospital Basel patient whose radiology report mentions a liver lesion, every liver lesion segmented under one label (TotalSegmentator's liver_lesions training set).

30 test casesabdomenimage CT
Look at the images
Read before comparing (9)
  • Source. TotalSegmentator's training set for liver_lesions: 842 CTs whose maps its page calls manual segmentations. Nicoli et al. 2025 describe 429 of them, late arterial or portal venous CTs whose report mentions a liver lesion of any kind; a junior physician drew 414 voxel by voxel under a radiologist with 3 years of experience, who corrected 12.
  • Doubt. Where the other 413 come from is not said. The paper also has 172 CTs around a chemoembolisation whose tumours its own model segmented. None of 11 CTs sampled shows iodised oil, but if such CTs are here, part of the truth is a network's and one patient's two CTs could straddle the split. The deposit gives no patient, date or set: graded expert, with that doubt.
  • Agreement. On 40 studies segmented by both readers apart, reader 1 against reader 2 found 64.7% of the lesions with 0.78 false positives a case, a Dice of 0.46 (0.34 to 0.58) and a Cohen's kappa of 0.24. One label holds lesions of every kind, and what counts as one is the reader's call.
  • Cases. The 30 test CTs hold 0.11 to 1,125 mL of lesions, median 16 mL. Three hold under 0.3 mL (Auqzq3k4LbDBh04B, MEdqPXvGFS37edbe, u7qCgzi9j3Ji9fJH): small lesions a method finds or misses, a Dice near 0 or not.
  • Images. The deposit's CTs as delivered, cropped by the provider around the body, in Hounsfield units through the header; slices lie 1 mm apart on 649 of the 842, 0.4 to 5 mm on the rest. xAVNlwaAdzLNYNLx holds a padding of -8192 over 12% of its voxels. No mask is given.
  • Checks. On the cases built first each map holds 0 and 1 on its CT's grid; on the ten test cases drawn first the lesions are darker than the liver, and on nine 88 to 100% of their voxels lie in TotalSegmentator's liver. xAVNlwaAdzLNYNLx's truth is one exophytic mass of 57 mL against the spleen, kept. The 40 added hold 0 and 1 on their CT's grid.
  • TotalSegmentator. Its liver_lesions model, version 2.18, was trained on this deposit and has seen every case: Dice 0.974 on 3pnTnEfD3wRhu9TO and 0.935 on Vo5MyVU69MbeWaNV. That is the model remembering its training set; the paper's own model reached 0.66 on its 46 test CTs.
  • Overlap. No image is in another dataset of the corpus. The 20 CTs built first were compared with every corpus image of the same voxel size, HCCTACE's, MSD_HepaticVessel's, CRLMCT's and LiverMetsCT's among them, and with TotalSegmentator's CTs, TotalSegCT's source; the 40 added, with every image of the same size and voxel size in the corpus as built on 2026-10-19. Nor is the deposit LiTS, the decathlon's hepatic vessel CTs or Colorectal-Liver-Metastases.
  • Outside numbers. None on these cases: the paper's test CTs are not named. The provider sets no surface tolerance.

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

The first test case of TotalSegmentator liver_lesions, CT
The first test caseClick to look at every case.
Test cases
30 test CTs drawn with seed 2026 among the 766 outside development, and 30 development CTs, the last of the same seeded order. The pool is the deposit's 796 CTs, of 842, whose lesions hold 0.06 mL or more, the smallest lesion the paper counts; 27 maps are empty and 19 hold less. The provider publishes no test set and no patient identifier.
Truth
contours by experts every lesion of the liver, benign or malignant, under one label, which the deposit calls manual segmentations; for the 414 CTs its paper describes, drawn voxel by voxel, slice by slice, by a junior physician under a radiologist with 3 years of experience, who corrected 12 of them (Nicoli et al. 2025)
Licence
CC BY 4.0
30 cases for tuning, never tested

icWKYwJBtsQO5sgA, Z3qpH62JejdRciIf, pAF0BZmcGpQI3F1A, GWWKkiru4yYEUxeP, qrj7E96P2HZAev31, kqogixwNc9fbDSfK, TB78AMKH4380b8H3, DdmZIydGuTM3Fett, wXPokKusbhpziLKz, hZ9rIZ3lJOUjeSpb, gfQOwFWge782svgh, LYNeYgFtpeUqeZj8, xxCR0q1jedPX1buv, lteswBsZxJDpyoPR, drCiM8jdRkx85TtG, mOLpUg8RqI723eDG, IVWpOzlJSAoBWHrk, V7zlLsDGYRKgi6zF, I6Lsn0coRYv29AqM, xT6wxc6Ivuk49fEn, rUspdMGq0FWRtgzf, exrnZu1mNbQpuwKZ, crjUNA7USUvstTPX, Rt6pSQZmiaRZRMRG, euR7OEDhkn5scJJ0, gk9ux0Hgk8SG1fXV, wdcNFvHw5V66gLUz, dKAwbDoIpgD6Nq3z, WvkA47s9UJh8qpQo, wGBlgPBr3ofLjoxX