fidopt Interactive view
Spine, bones and joints, segmentation

CSI 2014 vertebrae

Spine CT of a young adult or of a patient with vertebral compression fractures, each thoracic and lumbar vertebra segmented (CSI 2014 vertebra segmentation challenge, University of California, Irvine).

10 test casesspineimage CT
Look at the images
Read before comparing (9)
  • Cohort. Spine CTs of the Department of Radiological Sciences, University of California, Irvine: the 10 training scans and test scans 11 to 15 of young adults aged 16 to 34, test scans 16 to 20 of patients aged 59 to 82 with osteoporotic compression fractures. Philips iCT 256, two test scans on a Siemens Sensation 64, with contrast, a small field centred on the spine (Yao et al. 2016).
  • Images. The record's normalised copy: Hounsfield units clipped to -1000 to 3095, origin at 0 and axes reordered to the identity direction by the uploader, nothing resampled. 512 x 512 pixels of 0.31 to 0.38 mm, slices 0.7 to 2 mm thick; the fracture scans are the coarser (2 mm). As 16-bit integers. No mask is given.
  • Truth. An automatic segmentation corrected and refined by a medical fellow and a research fellow, the test references by consensus of two operators; then cleaned by the uploader (islands removed, 2 x 2 x 2 median filter) and cut to T1 to L5. Graded semi-automatic: a method close to the starting algorithm may be favoured. The uploader warns that the fracture scans' references are of lower quality.
  • Labels. 17 vertebrae a scan, T1 to L5 (8 to 24, as in the record); test scans 13 and 17 have four lumbar vertebrae, L5 left out (label 24 absent). The cervical vertebrae, the sacrum and the ribs are not labelled even where the field holds them. On every test scan the labels follow each other from the head down, and the vertebrae read a median of 149 to 457 HU.
  • TotalSegmentator. Version 2.18 (task total) scores a mean Dice of 0.913 over the 17 vertebrae of case11 (0.88 to 0.93) and 0.909 on case16 (0.86 to 0.93), each level on the label of its own name: the numbering, the sides and the headers agree.
  • Room. Each test vertebra moved one voxel keeps a Dice of 0.81 to 0.98.
  • Start. Labelling the whole body (above -500 HU) as spine scores a mean Dice of 0.09 against all vertebrae together on the test scans, the bone above 200 HU 0.44; neither names the levels.
  • Overlap. No other dataset of the corpus names the CSI 2014 challenge, SpineWeb or UC Irvine; VerSe's scans come from TUM and Glocker's set, CTSpine1K's from four other collections. No built image has the grid of any of the 20 scans (12,304 stores compared by size and spacing, a check that finds AbdomenDIRQA's case20 in MSD_Pancreas).
  • Outside numbers. The challenge's five entries, scored on the original references, reached a mean Dice of 0.93 to 0.96 per region on the healthy test scans and 0.88 to 0.92 on the fracture scans (best per region, Yao et al. 2016): close to, but not the same as, the cleaned masks scored 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
The challenge's test set taken whole: its 10 scans, 11 to 15 of young adults and 16 to 20 of older patients with osteoporotic compression fractures. Development: its 10 training scans, 1 to 10, all of young adults. One scan a person.
Truth
contours by a program, verified by people one label a vertebra, T1 to L5: an automatic segmentation (adaptive thresholding, watershed, graph search) corrected and refined by hand by a medical fellow and a research fellow, the test references by consensus of two operators (Yao et al. 2016); in this copy islands removed, the masks smoothed by a 2 x 2 x 2 median filter and the cervical vertebrae dropped by the uploader
Licence
ODC-By 1.0
10 cases for tuning, never tested

case01, case02, case03, case04, case05, case06, case07, case08, case09, case10