VerSeFusion

A re-fused, PIR-canonical version of the VerSe 2019 and VerSe 2020 vertebra segmentation challenges, with VERIDAH (Möller 2026) label corrections applied for thoracolumbar transitional vertebrae.

Dataset stats

• Total scans: 374
• Total patients: 355
• Splits: training=141, validation=120, test=113
• Source: VerSe 2019 + VerSe 2020 (combined) with VERIDAH corrections
• Canonical orientation: PIR (axis 0 = P, axis 1 = I, axis 2 = R)
• VERIDAH-corrected subjects: 14

Orientation

Every scan in this dataset has been reoriented to a single canonical frame:

• axis 0 increases toward P (posterior — i.e., anterior → posterior)
• axis 1 increases toward I (inferior — i.e., superior → inferior; this is the spine axis)
• axis 2 increases toward R (right — i.e., left → right)

This is verified end-to-end: see orientation_audit.json for the per-subject report. Rendering conventions in previews/:

• Coronal: head at top, patient's right at viewer's right
• Axial: anterior at top, patient's right at viewer's right
• Sagittal: head at top, anterior at left

Structure

gregoryschwingmdphd/VerseFusion/
├── README.md
├── LICENSE
├── splits.csv                  # series_id → split (training/validation/test)
├── orientation_audit.json      # per-subject orientation verification
├── scans/
│   └── <series_id>/
│       ├── ct.nii.gz           # CT volume, HU values, PIR-oriented
│       ├── mask.nii.gz         # vertebra labels (uint8), PIR-oriented
│       └── meta.json           # per-scan provenance
├── corrections/
│   └── veridah_manifest.json   # which subjects had labels corrected
└── previews/                   # optional QC renders
    └── <series_id>.png

Label schema

Label	Anatomy		Label	Anatomy
1–7	C1–C7		20	L1
8	T1		21	L2
9	T2		22	L3
10	T3		23	L4
11	T4		24	L5
12	T5		25	L6 (supernumerary lumbar)
13	T6		26	sacrum (variably annotated)
14	T7		27	coccyx
15	T8		28	T13 (supernumerary thoracic)
16	T9
17	T10
18	T11
19	T12

Loading example

import nibabel as nib

ct  = nib.load("scans/verse001/ct.nii.gz")
msk = nib.load("scans/verse001/mask.nii.gz")

# Both are guaranteed to be PIR-oriented:
assert nib.aff2axcodes(ct.affine)  == ('P', 'I', 'R')
assert nib.aff2axcodes(msk.affine) == ('P', 'I', 'R')

Citation

If you use this dataset, please cite the original VerSe challenges and the VERIDAH corrections paper:

@article{sekuboyina2021verse,
  title={VerSe: A vertebrae labelling and segmentation benchmark for multi-detector CT images},
  author={Sekuboyina, A. and others},
  journal={Medical Image Analysis},
  year={2021}
}

@article{loffler2020verse2020,
  title={A vertebral segmentation dataset with fracture grading},
  author={Löffler, M.T. and others},
  journal={Radiology: Artificial Intelligence},
  year={2020}
}

@article{moller2026veridah,
  title={VERIDAH: Vertebral identification and transitional anomaly detection},
  author={Möller, H. and others},
  year={2026}
}

Acknowledgments

VerSe challenge data: Technical University Munich. VERIDAH corrections: H. Möller et al. (2026).