fidopt Interactive view
Body CT, many structures, segmentation

ENHANCE.PET 1.6k

The CT of a whole-body FDG-PET/CT study of a patient with lung cancer at Leipzig or Careggi, with 127 structures (organs, heart and vessels, muscles, bones, ribs, vertebrae) drafted by a network and corrected by people (ENHANCE.PET 1.6k).

30 test caseswhole bodyimage CT
Look at the images
Read before comparing (8)
  • Pool. The 384 Leipzig and 199 Careggi lung cancer FDG-PET/CT studies of the release (LuCaPET consortium, CC BY 4.0), arms up; its 1,014 autoPET studies (CC BY-NC 4.0, already in AutoPET_Seg, AutoPETCT and AutoPET_CTPET_Syn) are left out; the UC Davis studies the paper describes are not in the bucket.
  • Patients. The studies carry no patient key. Three Leipzig exams are delivered twice under two numbers (1028 and 1240, 1200 and 1208, 1201 and 1209: same date, series time, dose and slope; their PETs correlate at 0.99999, or match slice for slice where one is cut short), and the higher number is kept. As a precaution, studies of a site whose sex, height (2 cm), weight (10 kg) and age against date gap agree are also kept to one (prepare.py), 514 studies in all; such matches occur no more often than chance, so this rule finds no repeat patient and most likely joins different people.
  • Truth. MOOSE's network drafted 130 structures on each CT; one medical student corrected them in 3D Slicer, a second reviewed and corrected, then a radiology resident and a nuclear medicine resident reviewed. Tracer-avid lesions seen on the PET were cut out of the organs, so an organ holding a tumour is drawn without it. Graded semi-automatic: most of each outline is the network's.
  • Labels. Six of the seven groups in one map, 127 structures under TotalSegmentator's names where it has them; Body-Composition (muscle and fat on a slab at L3, which the muscles overlap) is left out. Where two groups both claim a voxel, mostly where the hip bones meet the femurs and where ribs meet vertebrae, the voxel is in the ignore region, out of the truth and the prediction before scoring: a median of 5,849 voxels a test case (74 to 23,154). A test case holds 89 to 117 structures (median 100), those in its field of view; left structures lie on the patient's left on all 85 pairs checked.
  • Images. The provider's CTs as delivered, 32-bit floats in Hounsfield units, -1024 outside; pixels 0.98 or 1.17 mm, slices 1 to 5 mm, whole body or head to thighs, defaced by the provider.
  • Overlap. No other dataset of the corpus holds Leipzig or Careggi studies (QuadraHC is Vienna's). ENHANCEPET_CTPET_Syn draws other Leipzig patients.
  • Start. A mask of the whole image (ignore region out) scores a Dice of at most 0.012 for any structure at the median of the test cases, 0.03 on any case.
  • Outside numbers. The provider's nnU-Net reports Dice on its own 80/20 split, not these cases; MOOSE was trained on this release and may have seen the test cases.

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
30 test patients drawn with seed 2026 by site (22 Leipzig, 8 Careggi) among the 484 outside development, then 30 development patients, the last of the same seeded order, the sites in proportion. One study a patient: of the 583 studies of the two sites, three Leipzig exams delivered twice are taken once and prepare.py keeps one study of each group it links, 514 in all. The release has no test set; ENHANCEPET_CTPET_Syn takes other Leipzig patients.
Truth
contours by a program, verified by people the provider's six groups (organs, cardiac, muscles, peripheral bones, ribs, vertebrae) in one map: MOOSE's network drafted them on the CT, one medical student corrected them in 3D Slicer, a second reviewed and corrected, then a radiology resident and a nuclear medicine resident reviewed; tracer-avid lesions on the PET were taken out of the organs; voxels two groups both claim are ignored
Licence
CC BY 4.0 (Leipzig and Careggi studies)
30 cases for tuning, never tested

1223, 1480, 1146, 1242, 1274, 1590, 1209, 1383, 1034, 1456, 1119, 1181, 1476, 1061, 1308, 1212, 1406, 1192, 1311, 1064, 1545, 1143, 1358, 1435, 1366, 1179, 1046, 1520, 1171, 1215