fidopt Interactive view
Spine, bones and joints, segmentation

whole-spine T2 MR

T2-weighted MR of the whole spine of one healthy volunteer, with each vertebra and disc from C1 to the sacrum, the spinal canal and the cord labelled by a network and corrected by hand (whole-spine, TotalSpineSeg's training data).

10 test casesspineimage MRresearch only
Look at the images
Read before comparing (6)
  • Cases. One sagittal 3D T2-weighted SPACE a volunteer, 1 mm isotropic, from C1 down to L2 at least and to the sacrum on one test case, as delivered, of healthy adults aged 21 to 56 scanned on a Siemens Verio (AMU, Marseille) or Trio (UNF, Montreal). The dataset's whole-body T1 MPRAGE and its B0 simulations are not used. No mask.
  • Labels. 50 structures: the cord, the canal around it, C1 to L5, the sacrum and the 23 discs from C2-C3 to L5-S1, by TotalSpineSeg's names; the canal map holds the cord and is written under it. Levels below the field are absent from a case and not scored there: L3 is on 7 test cases, L4 on 3, L5 and the sacrum on 1.
  • Truth. Released by TotalSpineSeg's authors as training labels: initial labels from a registration to the PAM50 template, a network trained on them, its output corrected by hand in 3D Slicer (sidecar: TotalSpineSeg r20240601, then 'Manual', Y. Warszawer). Graded semi-automatic; no second reader.
  • Grid. The labels lie on the image's voxels with the first axis reversed, at times with one more empty slice: prepare.py puts them on the image's grid without moving a voxel.
  • Start. A mask of the whole image scores a Dice of 0.0015 at the median over the test cases and structures, an Otsu mask 0.0005; on the canal 0.012 and 0.089. TotalSegmentator 2.18.0's vertebrae_mr scores a median Dice of 0.86 over the vertebrae of sub-amuMT and 0.87 on sub-unfErssm003, where it falls to 0.57 to 0.69 from T6 to T9 and near 0 on L1 and L2.
  • Overlap. SpineGeneric's built pair holds AMU volunteers sub-amu02 and sub-amu03 of spine-generic, scanned at AMU in 2019, with no identifier shared with this dataset. Matching sex, height within 2 cm and weight within 6 kg against spine-generic's five AMU volunteers flags ALT, HB, LJ, MD, TR and VC; FR, GB and MLL have no demographics; VP (a man, 180 cm, 66 kg) is 3 cm from sub-amu02 (183 cm, 67 kg), a SpineGeneric test case, on the same Verio; all ten are held out. Several UNF volunteers match spine-generic's UNF volunteers as loosely, but none of those is built. SPIDER and spine-generic, TotalSpineSeg's other sources, are not used here.

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 volunteers drawn with seed 2026 by site (AMU Marseille, UNF Montreal) and 10 development volunteers, one scan each. The pool is the 56 volunteers whose T2w image the snapshot holds and the authors labelled; ten AMU volunteers who could be SpineGeneric's are held out (Overlap).
Truth
contours by a program, verified by people TotalSpineSeg's released labels of the T2w image: initial labels from a registration to the PAM50 template, then a network's output corrected by hand in 3D Slicer (Warszawer et al. 2025); the canal and cord masks merged into the vertebra and disc map
Licence
images: CC0 (OpenNeuro ds005616); labels: released in the TotalSpineSeg repository, LGPL-3.0
10 cases for tuning, never tested

sub-unfErssm009, sub-unfPain003, sub-unfErssm021, sub-unfErssm016, sub-unfErssm001, sub-unfErssm012, sub-amuFL, sub-unfErssm018, sub-unfPain004, sub-amuAM