VERSE20 — Vertebrae Segmentation Challenge 2020

3D CT vertebrae segmentation benchmark from the MICCAI 2020 challenge. Multi-detector CT scans of the cervical, thoracic, and lumbar spine with manually corrected per-vertebra segmentation masks.

Dataset Summary

Field	Details
Modality	CT (multi-detector, multi-vendor)
Body Part	Spine (cervical C1–C7, thoracic T1–T12, lumbar L1–L6)
Subjects	~286 unique subjects
Series (CT volumes)	303
Splits	01_training (104) / 02_validation (99) / 03_test (100)
Annotated vertebrae	4,142 (paper figure across all splits)
License	CC BY-SA 4.0 (data) / MIT (code)
Source	OSF b2wxj (MICCAI structure) — https://osf.io/b2wxj/

Mask Convention

Voxel values 1–25 encode individual vertebrae:

• 1–7 cervical (C1–C7)
• 8–19 thoracic (T1–T12)
• 20–25 lumbar (L1–L5; L6 is rare transitional)
• 28 T13 (rare transitional thoracic)

Data Layout

VERSE20/
  01_training/{subject}/{subject}{_CT-iso|_CT-sag|_CT_ax|}.nii.gz       # CT volume
  01_training/{subject}/{subject}{_CT-iso|_CT-sag|_CT_ax|}_seg.nii.gz   # segmentation
  01_training/{subject}/{subject}{_CT-iso|_CT-sag|_CT_ax|}_iso-ctd.json # centroids
  02_validation/...
  03_test/...

data/
  train-*.parquet, validation-*.parquet, test-*.parquet  # Dataset Viewer preview
                                                          # (middle slice + mask + overlay)

Some subjects have multiple acquisitions (_CT-iso, _CT-sag, _CT_ax) — each is a separate (CT, seg, centroids) triplet.

Manifests (train.jsonl, validation.jsonl, test.jsonl)

One JSON object per CT series:

{
  "image": "VERSE20/01_training/verse008/verse008.nii.gz",
  "mask":  "VERSE20/01_training/verse008/verse008_seg.nii.gz",
  "centroids": "VERSE20/01_training/verse008/verse008_iso-ctd.json",
  "subject_id": "verse008",
  "sample_id":  "verse008",
  "label":      [<26 vertebra-name strings>],
  "vertebrae_present": [19, 20, 21, 22, 23, 24],
  "modality":   "CT"
}

The 26-name label list mirrors the existing VerSe convention so the EasyMedSeg data_preprocess.VerSe function can be reused without changes.

Citation

@article{sekuboyina2021verse,
  title   = {VerSe: A Vertebrae labelling and segmentation benchmark for multi-detector CT images},
  author  = {Sekuboyina, Anjany and Husseini, Malek E. and Bayat, Amirhossein and Loeffler, Maximilian and Liebl, Hans and Li, Hongwei and Tetteh, Giles and others},
  journal = {Medical Image Analysis},
  volume  = {73},
  pages   = {102166},
  year    = {2021},
  doi     = {10.1016/j.media.2021.102166}
}

@article{liebl2021ctvertebral,
  title   = {A computed tomography vertebral segmentation dataset with anatomical variations and multi-vendor scanner data},
  author  = {Liebl, Hans and Schinz, David and Sekuboyina, Anjany and others},
  journal = {Scientific Data},
  volume  = {8},
  number  = {284},
  year    = {2021},
  doi     = {10.1038/s41597-021-01060-0}
}