fidopt Interactive view
Spine, bones and joints, segmentation

CTSpine1K

CT of one patient from one of four public collections (CT colonography, liver tumour CT, COVID-19 chest CT, head and neck CT), with each vertebra in the field segmented, C1 to L6 (CTSpine1K).

368 test casesspineimage CTresearch only
Look at the images
Read before comparing (9)
  • Pool. 1,005 CTs, one a patient, converted to NIfTI by the provider: 784 of TCIA's CT COLONOGRAPHY (one of the two positions at random), 150 of the decathlon's liver task, 40 of TCIA's CT Images in COVID-19 and 31 of HNSCC-3DCT-RT. The provider left out scans with metal artefacts and split the rest into trainset 610, test_public 197 and test_private 198.
  • Source. The files come from the Hugging Face copy the provider's README links (alexanderdann/CTSpine1K, commit 9b454add), the only route without a sign-in; two cases compared with the provider's XNAT project are the same bytes. The provider keeps CC BY-NC-SA 4.0 for its four sources, which TCIA and the decathlon give as CC BY 3.0 and CC BY-SA 4.0: the stricter reading applies.
  • Truth. An nnU-Net trained on 41 VerSe scans proposed the masks; junior annotators refined them, two senior annotators checked and refined every one, a spine surgeon settled the hard cases and coordinators double-checked a random sample. The model was retrained every 100 cases. No agreement between readers is given. Graded semi-automatic: a method close to that nnU-Net may be favoured.
  • Labels. Every vertebra in the scan is labelled, slivers at the field's edges included; the sacrum and ribs never are. L6 (label 25, a lumbarised S1) is in none of the 20 cases built first, and the provider's nnU-Net scores a Dice of 0 on it. prepare.py removes editing leftovers, pieces under 50 voxels off the first and last slices.
  • Checks. On every case the labels away from the edges follow each other from the head down. Their vertebrae read a median of 192 to 359 HU on the 20 cases built first and 100 to 700 HU on the 378 others, checked by script, but C1 of HN_P028 (798 HU), whose labels lie on their vertebrae.
  • Images. The copy's files as delivered, CT in Hounsfield units, each mask on its CT's grid. The colonography scans are low-dose and noisy, 500 to 700 slices of 512 x 512; the COVID-19 chest CTs have slices 4.5 mm apart on average.
  • TotalSegmentator. Version 2.18 scores a Dice over the whole spine of 0.942 on 1.3.6.1.4.1.9328.50.4.0640, 0.936 on liver_142 and 0.919 on volume-covid19-A-0380_ct, each level on the label of the same name. The provider's nnU-Net reaches a mean Dice of 0.869 on test_public and 0.840 on test_private, lowest from T5 to T8 and C4 to C6.
  • Seen. The provider's released model and any model trained on its training split have not seen the test cases. All 1,005 masks are public, so a model trained on all of CTSpine1K has.
  • Overlap. Left out by name: the 41 CT COLONOGRAPHY patients of CTColonography and HQColon, and the 20 decathlon liver scans of MSD_Liver and the 20 of CTORG. CTORG_Reg's 28 CT-ORG scans, matched by size and voxel size or by resampling, are left out too. liver_142 matches no image built for the corpus. No other descriptor names HNSCC-3DCT-RT or TCIA's CT Images in COVID-19.

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
The provider's two test splits taken whole, test_public (197) and test_private (198), less the scans another dataset of the corpus holds: 368 test cases, 287 COLONOG, 50 MSD-T10, 20 COVID-19 and 11 HNSCC-3DCT-RT. 30 development cases are drawn in the seeded order from the provider's training split, 530 left of its 610. liver_28 to liver_47 (3D-IRCADb-01, broken headers) are left out too.
Truth
contours by a program, verified by people the provider's masks, one label a vertebra in its numbering (1 to 25, C1 to L6): an nnU-Net trained on VerSe proposed each mask, junior annotators refined it in ITK-SNAP, two senior annotators checked and refined every mask, a spine surgeon (12 years on average) settled the hard ones, and coordinators double-checked a random share; the network was retrained every 100 cases (Deng et al., MELBA 2024). The sacrum is not labelled. Pieces of a label under 50 voxels off the scan's first and last slices, editing leftovers, are set to 0 (prepare.py).
Licence
CC BY-NC-SA 4.0 (the provider's own deposit, its XNAT project africai_miccai2024_ctspine1k, which names the version; its README says CC-BY-NC-SA, for the images and the masks it redistributes)
30 cases for tuning, never tested

1.3.6.1.4.1.9328.50.4.0431, HN_P016, 1.3.6.1.4.1.9328.50.4.0229, 1.3.6.1.4.1.9328.50.4.0041, 1.3.6.1.4.1.9328.50.4.0224, 1.3.6.1.4.1.9328.50.4.0185, 1.3.6.1.4.1.9328.50.4.0485, 1.3.6.1.4.1.9328.50.4.0372, 1.3.6.1.4.1.9328.50.4.0018, 1.3.6.1.4.1.9328.50.4.0308, 1.3.6.1.4.1.9328.50.4.0009, 1.3.6.1.4.1.9328.50.4.0133, 1.3.6.1.4.1.9328.50.4.0257, 1.3.6.1.4.1.9328.50.4.0503, 1.3.6.1.4.1.9328.50.4.0247, 1.3.6.1.4.1.9328.50.4.0370, 1.3.6.1.4.1.9328.50.4.0376, 1.3.6.1.4.1.9328.50.4.0400, 1.3.6.1.4.1.9328.50.4.0368, 1.3.6.1.4.1.9328.50.4.0091, 1.3.6.1.4.1.9328.50.4.0189, 1.3.6.1.4.1.9328.50.4.0326, 1.3.6.1.4.1.9328.50.4.0321, HN_P017, 1.3.6.1.4.1.9328.50.4.0379, 1.3.6.1.4.1.9328.50.4.0375, 1.3.6.1.4.1.9328.50.4.0094, 1.3.6.1.4.1.9328.50.4.0256, 1.3.6.1.4.1.9328.50.4.0200, 1.3.6.1.4.1.9328.50.4.0352