CTSpinoPelvic1K

This repository holds an earlier export. The current release is v6, archived at 10.5281/zenodo.22139643.

v6 is a different dataset in every respect that matters: 802 records rather than 1153 files, a scheme of 84 identifiers rather than 10, per-level ribs, femurs, an S1 class carved from the sacrum, surgical instrumentation, and a crosswalk giving the TCIA SeriesInstanceUID each annotation was drawn on. It also carries corrections that this export does not — twenty-two records had left and right hip labels swapped or interdigitated here, and the transitional-anatomy flags in this repository's manifest are unreliable (see below).

Use the Zenodo record for anything new. What follows documents what is actually in this repository, because it is still downloadable and it should not mislead anyone who does.

A fused spine + pelvis 3D CT segmentation dataset built by patient-level crosswalk between three public sources:

1. TCIA CT COLONOGRAPHY — DICOM CT volumes (prone + supine per patient)
2. CTSpine1K (COLONOG subset) — VerSe-convention vertebral label masks
3. CTPelvic1K dataset2 — sacrum + bilateral hip label masks

Annotations are placed onto the TCIA CT volume with the highest bone coverage (HU > 200), separately per anatomy. Where both land on the same series the record is fused; otherwise spine and pelvic labels target different prone/supine acquisitions (separate).

Labels

ID	Name	Source
0	background	fused files only — see the warning below
1	L1	CTSpine1K (VerSe 20 → 1)
2	L2	CTSpine1K (VerSe 21 → 2)
3	L3	CTSpine1K (VerSe 22 → 3)
4	L4	CTSpine1K (VerSe 23 → 4)
5	L5	CTSpine1K (VerSe 24 → 5)
6	L6	CTSpine1K (VerSe 25 → 6) — six-lumbar spines
7	sacrum	CTPelvic1K (dataset2 1 → 7)
8	left hip	CTPelvic1K (dataset2 2 → 8)
9	right hip	CTPelvic1K (dataset2 3 → 9)
10	not annotated (ignore)	the region this file does not cover

Identifier 10 is not anatomy and earlier versions of this card omitted it. In a *_spine_label.nii.gz the pelvis is unannotated, and in a *_pelvic_label.nii.gz the spine is; identifier 10 marks that region. It is not a rare edge case — measured on these files it covers 99.7% of a spine-only volume and 99.0% of a pelvic-only one, and in those files identifier 0 does not appear at all, because 10 has taken the place of background.

Treat 10 as ignore, not as a class. A loss computed over identifiers 0–10 without masking it trains the network to predict "not annotated" across almost the entire volume. Fused files (NNNN_label.nii.gz, no _spine_/_pelvic_ qualifier) use 0 for background and contain no 10.

CTPelvic1K's sacrum takes priority over CTSpine1K's sacrum (VerSe 26) so the two conventions do not collide on a transitional vertebra.

Orientation

All volumes are ('P','I','R'). Each CT and its label share exactly the same 4×4 affine, so no resampling or reorientation is needed to overlay them — verify with np.allclose(ct.affine, label.affine) rather than assuming it.

File format

NIfTI pairs, not archives:

import nibabel as nib
import numpy as np

lab = nib.load("labels/0002_label.nii.gz")
ct = nib.load("ct/0002_ct.nii.gz")
assert np.allclose(lab.affine, ct.affine)

L = np.asanyarray(lab.dataobj)
print(sorted(int(v) for v in np.unique(L)))   # includes 10 on separate-mode files

manifest.json carries one record per file, and splits_5fold.json carries frozen patient-grouped folds. Earlier revisions of this card described a 70/15/15 train/val/test split and a .npz layout; neither is what this repository contains.

Transitional-anatomy flags: do not filter on them

has_l6, has_lumbar_rib and n_lumbar_labels in the manifest are unreliable in this export and were corrected only for v6. Counted in the released label volumes, has_l6 was true for one record that contains no L6 and false for all eighteen that do, and has_lumbar_rib was false for every record including the sixteen that carry one. Anyone selecting the six-lumbar cases on those fields receives one record that is not one, and anyone selecting lumbar ribs receives an empty set.

Derive the flags from the label volumes, or use the v6 record, where the fields are computed from the voxels.

Citation

Please cite the source datasets — CTSpine1K, CTPelvic1K, and TCIA CT COLONOGRAPHY — alongside this derivative release, and cite the dataset itself at 10.5281/zenodo.22139643.

Licence

CC BY-NC-SA 4.0. The ShareAlike term is inherited rather than chosen: CTSpine1K, from which the vertebral annotations derive, is released under CC BY-NC-SA, and a derivative of a ShareAlike work must carry the same licence. Earlier versions of this card said CC BY-NC 4.0 in the body while the metadata said CC BY-NC-SA 4.0; CC BY-NC-SA 4.0 is correct.

The source datasets retain their own licences. Research use only — these labels are not a medical device and are not validated for clinical decision-making.