fidopt Interactive view
One lesion in a CT volume of interest, segmentation

ULS23, lesions in CT volumes of interest

A CT volume of interest of 256 x 256 x 128 voxels centred on one lesion of the chest, abdomen or bones, the lesion segmented in 3D (ULS23 challenge, novel training data: DeepLesion and Radboudumc bone lesions).

30 test caseswhole bodyimage CTresearch only
Look at the images
Read before comparing (8)
  • Pool. The release's novel data: 750 DeepLesion volumes (743 with a mask, 549 patients) chosen by the organisers where a 3D nnU-Net disagreed most with DeepLesion's RECIST diameters, across the chest and abdomen; and 744 bone lesions, sclerotic and lytic, of Radboudumc studies whose reports mention bone disease (697 with a mask, 151 ids). One volume a patient, the first by name; the release's DeepLesion names number patients first, and the bone id is the only one it gives, listed as a patient id in the paper's held-out split though its Table 1 counts 151 series.
  • Task. Each volume is cut around one lesion, centred on one of its voxels drawn at random by the provider; other lesions in the volume carry no label. It asks for the lesion at the centre, a task segmenters of whole scans do not take: ranked apart.
  • Truth. DeepLesion volumes: trained (bio)medical students drew each lesion in 3D from its RECIST measurement, three times, the majority kept; bone volumes: an experienced radiologist found and drew each lesion. No agreement figure is released for these parts.
  • Volumes. 256 x 256 x 128 voxels on the scan's spacing, 0.47 to 0.98 mm in plane. DeepLesion released a few slices around each key slice, so most DeepLesion volumes are 13 to 128 real slices (1 or 5 mm apart) padded to 128 at -1024 HU; bone volumes have 0.8 or 1.5 mm slices. The test lesions hold 0.16 to 202 mL (median 0.93 mL).
  • Values. Eight of the twelve bone volumes drawn (those on 1.5 mm slices) are delivered in Hounsfield units plus 1024: prepare.py takes 1024 off any volume whose minimum is 0 or more, so that all are in Hounsfield units. Padding stays as delivered (-1024, -2048 or the scan's minimum less one).
  • Start. Labelling the whole volume scores a Dice at the median of the test volumes of 0.0001 (at most 0.020), Otsu's bright voxels 0.0005 (at most 0.046). TotalSegmentator has no lesion task; the challenge's own nnU-Net baseline is not run here.
  • Overlap. Not used: the pancreas part, whose Radboudumc patients may be PANORAMA's, and the release's other parts, which recut KiTS21, LIDC-IDRI, LiTS, the decathlon, NIH lymph nodes, CCC18 and DeepLesion's 2D-measured lesions, held elsewhere in the corpus. No other dataset holds DeepLesion; its patients are the NIH Clinical Center's, as are CTLymphNodes_Abdomen's, CTLymphNodes_Mediastinum's and PancreasCT8's, and DeepLesion's anonymised ids cannot be linked to theirs: a patient in two of them cannot be ruled out. The bone ids are not linked to other Radboudumc sets (PANORAMA, PICAI, LUNA25): a patient in two of them cannot be ruled out.
  • Outside numbers. None on these volumes: the challenge ranks on its hidden test set of 775 lesions.

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 volumes drawn with seed 2026 among the 700 patients of the release's DeepLesion and Radboudumc bone parts, the two in proportion (24 and 6), then 30 development volumes; one volume a patient, the first by name that has a mask. The pancreas part is not used, and the challenge's test volumes are hidden.
Truth
contours by experts the lesion the volume is centred on, segmented in 3D: on the DeepLesion volumes by trained (bio)medical students from DeepLesion's RECIST measurement, three a lesion, their majority kept; on the bone volumes by an experienced radiologist; other lesions in the volume are not labelled; no radiologist review of this training data is stated (the DeepLesion masks are the majority of three trained students, drawn from the radiologists' RECIST marks)
Licence
CC BY-NC-SA 4.0
30 cases for tuning, never tested

000711_02_01_537_lesion_01, 003746_01_01_270_lesion_01, bone_00504_lesion_01, 000258_01_01_047_lesion_01, 002642_01_02_022_lesion_01, 001129_02_01_257_lesion_01, bone_00229_lesion_01, 001220_01_01_118_lesion_01, 001994_01_01_046_lesion_01, 000859_01_01_117_lesion_01, 001767_02_01_073_lesion_01, bone_01048_lesion_01, 002358_02_03_017_lesion_01, 003096_01_01_079_lesion_01, 000392_02_01_089_lesion_01, 002016_01_01_007_lesion_01, bone_00254_lesion_01, 003398_01_01_030_lesion_01, 003327_01_01_586_lesion_01, 001969_01_02_328_lesion_01, bone_01064_lesion_02, 000603_05_01_004_lesion_01, 001229_02_01_105_lesion_01, 001079_03_03_140_lesion_01, 003953_01_02_246_lesion_01, bone_00519_lesion_01, 000548_03_01_036_lesion_01, 001152_02_01_066_lesion_01, 000517_04_02_463_lesion_01, 002455_01_01_149_lesion_01