CTSpinoPelvic1K

A fused spine + pelvis 3D CT segmentation dataset built for one job the common tools get wrong: numbering the lumbar spine correctly in patients with a lumbosacral transitional vertebra (LSTV). 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. For most patients both annotations land on the same series (fused cases); for the rest, spine and pelvic labels target different prone/supine acquisitions (separate cases, exported as two records per patient — one spine_only, one pelvic_native).

Design principles (the decisions that make it LSTV-aware)

1. Vertebrae are radiologist ground truth, full stop — we never ship a pseudolabelled spine. The L1–L6 / L6 / sacrum calls and the transitional adjudications come from CTSpine1K's radiologists.

2. Two counting anchors — rostral and caudal. You can't tell L5 from L6 locally, but the lumbar column is bracketed by two fixed landmarks, so the count (hence L5-vs-L6) is deterministic:

   • Rostral — T12, the last rib-bearing vertebra (class 19 in the VerSe-native scheme), directly above ground-truth L1. Not a separate class — just the last thoracic vertebra in the native column. Free from GT (present on 783/784 abdominopelvic studies; earlier exports just dropped the thoracic column).
   • Caudal — S1, the first sacral body, segmented in v3 as a distinct class ((GT sacrum) ∩ (TS vertebrae_S1)) — the bottom bracket of the lumbar count and the sacral-endplate landmark for sacral slope / pelvic incidence.

   With a fixed cranial anchor above L1 and the sacrum/S1 below L5–L6, the model learns rib-bearing-ness from the thoracic-vs-lumbar distinction. See docs/RIB_ANCHOR_RATIONALE.md.

3. LSTV captured from both ends, and graded — vertebral count (CTSpine1K) and pelvic annotation (CTPelvic1K), expert Castellvi-graded (see below).

4. Honest about the pelvis — real where fused, pseudolabelled (leak-safe) for spine_only, with quality reported as held-out Dice on the pelvic_native scans.

Versions:

• v1 — the partial-annotation artifact (all configs, ignore protocol), i.e. the input used to train the pseudolabeller, now carrying the T12 anchor. pelvic_native keeps ignore on the spine (never a faked label).
• v2 — the clean, densely-labelled release: CTSpine1K spine ground truth + pelvis (real where fused, pseudolabelled on spine_only), fused + spine_only only. pelvic_native is dropped (held out for pelvis validation). This is the LSTV-segmenter training artifact, with no ignore voxels on the shipped cases.
• v3 — bone-augmented (TotalSegmentator): v2 plus both femurs, the GT thoracic column, and an S1 carve of the sacrum, in the VerSe-native label scheme (see Bone augmentation below). Ribs are deferred to a future v4 — their ids are reserved but unpopulated here.

Reviewing segmentations? If you were asked to help correct the AI-drafted labels, see docs/REVIEW.md — account/token setup, install, and how to connect to the distributed review system (and, for the project maintainer, how to stand up the HuggingFace Space backend).

Labels — VerSe-native scheme

The label scheme is VerSe-native: the spine keeps its VerSe ids verbatim, and every non-VerSe structure gets a fixed id above the VerSe range. Source of truth is scripts/label_scheme.py.

ID	Name	Source
0	background	—
1–7	C1–C7	CTSpine1K (VerSe 1–7), native
8–19	T1–T12	CTSpine1K (VerSe 8–19); T12 = 19 is the counting anchor
20	L1	CTSpine1K (VerSe 20)
21	L2	CTSpine1K (VerSe 21)
22	L3	CTSpine1K (VerSe 22)
23	L4	CTSpine1K (VerSe 23)
24	L5	CTSpine1K (VerSe 24)
25	L6 / LSTV	CTSpine1K (VerSe 25) — lumbarized S1
26	sacrum	CTPelvic1K (dataset2 1 → 26); CTSpine1K VerSe 26 fallback
27	coccyx	CTSpine1K (VerSe 27)
28	T13	CTSpine1K (VerSe 28) — supernumerary thoracic
29	S1 (carved from sacrum)	(GT sacrum) ∩ (TS vertebrae_S1)
30	left hip	CTPelvic1K (dataset2 2 → 30)
31	right hip	CTPelvic1K (dataset2 3 → 31)
32 / 33	femur_left / femur_right	TS
34–45	rib_left_1 … rib_left_12	RibSeg (numbered off GT thoracic)
46–57	rib_right_1 … rib_right_12	RibSeg (numbered off GT thoracic)
255	ignore	partial-annotation only — un-traced region, NOT bg

CTPelvic1K's sacrum takes priority over CTSpine1K's sacrum (VerSe label 26) to avoid the two labelling conventions colliding on lumbosacral transitional vertebrae. The vertebral column is VerSe-native (C1–C7 = 1–7, T1–T12 = 8–19, L1–L6 = 20–25, sacrum = 26, coccyx = 27, T13 = 28); the rostral counting anchor is the last thoracic vertebra (T12 = 19), not a stored class.

Bone augmentation (TotalSegmentator)

v3 augments the spinopelvic core with the GT thoracic column, an S1 carve of the sacrum, and the TotalSegmentator femurs, all in the VerSe-native scheme above:

ID	Name	Source	populated in v3?
1–7	C1–C7	radiologist GT	yes (FOV-visible)
8–19	T1 … T12	radiologist GT thoracic column (placed VerSe masks)	yes (FOV-visible)
20–25	L1–L6	radiologist GT	yes
26	sacrum	radiologist GT	yes
28	T13	radiologist GT	yes (FOV-visible)
29	S1 (carved from sacrum)	(GT sacrum) ∩ (TS vertebrae_S1)	yes
30 / 31	left_hip / right_hip	radiologist GT	yes
32 / 33	femur_left / femur_right	TS	yes
34–45	rib_left_1 … rib_left_12	reserved	no — deferred to v4
46–57	rib_right_1 … rib_right_12	reserved	no — deferred to v4
255	ignore	sentinel	yes

The thoracic column was always in the source GT (placed VerSe masks), dropped from v2; v3 ships it — but only the vertebrae inside each scan's field of view (on these spinopelvic acquisitions usually the lower thoracic, ~T8 down, not up to T1; T1–T13 is the id range, not the per-case extent). Femurs are added directly on background, so the underlying anatomy is unchanged.

S1 (id 29) is the part of the GT sacrum that TS identifies as the S1 body: the carve splits the sacrum along its principal axis (so the S1/S2 plane follows pelvic tilt) and relabels the cranial slab in place — the sacrum's outer boundary stays radiologist GT. It gives the S1 superior-endplate landmark for sacral slope / pelvic incidence. Ribs are deferred to v4: on these FOV-limited scans there is no full rib cage to count from, so no automatic segmenter (TotalSegmentator, or point-cloud labelers like RibSeg) can number ribs reliably; ids 34–57 are reserved for a future v4 built on manual / AI-assisted rib annotation.

Coverage is 802 of 1,153 volumes — the spine-anchored + orphan cases; the 351 separate-mode pelvic sides keep v2 labels only.

LSTV — captured from both ends, expert-graded

Transitional status is read independently from the vertebral count (lstv_vertebral, CTSpine1K) and the pelvic/sacral annotation (lstv_pelvic, CTPelvic1K filename qualifier), with an lstv_agreement flag. The vertebral calls match CTSpine1K's published cohort case-for-case (16 lumbarizations + 9 sacralizations on COLONOG); 8 further sacralization/semi-sacralization cases come from the CTPelvic1K pelvic annotations. All 33 transitional cases carry an expert Castellvi grade — 25 sourced from CTSpine1K (vertebral) + 8 from CTPelvic1K (pelvic). The full per-case source table is in the dataset card (docs/dataset_card.md); the grading sheet is _lstv_phenotypes.csv.

Orientation

All volumes are canonicalised to PIR (Posterior–Inferior–Right). The CT and its label map share exactly the same 4×4 affine; no resampling is needed before training. PHI fields (descrip, aux_file, db_name, intent_name) are stripped from every NIfTI header.

A post-write HU sanity check is run at export time on each saved CT/label pair: HU values sampled at hip-mask voxels must be ≥30% bone (HU > 200), otherwise the case is flagged in the manifest as postwrite_hip_bone_pct. This catches frame-mismatches between CT data and its affine that earlier passed the simple affine_equal check.

LSTV annotation

Each case carries two complementary LSTV (lumbosacral transitional vertebra) annotations:

• lstv_vertebral — derived from CTSpine1K by counting lumbar labels in the segmentation (4 → sacralization, 5 → normal, 6 → lumbarization).
• lstv_pelvic — derived from CTPelvic1K filename qualifiers (any substring containing "sacralization" → sacralization).
• lstv_agreement — True when both sources agree, False when they disagree, None when either side is uninformative.
• lstv_confusion_zone — True for cases at the sacralization ↔ lumbarization boundary where the two signals disagree; flagged for downstream audit.
• lstv_class — integer 0–3 summarising the dominant call (0=normal, 1=lumbarization, 2=semi-sacralization, 3=sacralization). Pelvic label takes priority when both sides disagree.

Configs

A single patient can contribute multiple records depending on how spine and pelvic masks align across their prone/supine acquisitions:

config	Meaning
fused	Both masks placed on the same CT series; spine + sacrum + hips present
spine_only	Record carries lumbar labels only (20–25); sacrum/hips absent
pelvic_native	Record carries sacrum + hip labels only (26, 30/31); lumbar absent

Use match_type to distinguish where each side's labels came from:

match_type	Meaning
fused	Spine + pelvic masks land on the same TCIA series
separate	Spine and pelvic masks target different series for this patient; patient appears twice (one spine_only + one pelvic_native record)
spine_only	Patient has only a spine mask available
pelvic_only	Patient has only a pelvic mask available

Splits

Stratified patient-level 5-fold cross-validation with a held-out test set. Stratification preserves the rare LSTV classes across every fold so that no fold is missing lumbarization or sacralization cases at validation time.

The current splits document (splits_5fold.json, schema v4) uses LSTV-first stratum ordering: each patient is binned by <lstv_subtype>|<match_type>, and rare buckets are coalesced by dropping the match_type qualifier first (preserving the LSTV signal). An invariant check at generation time enforces that every fold's validation split contains at least 3 lumbarization cases (configurable). This was tightened from schema v3, which sometimes dropped the LSTV tag during coalescing and produced folds with zero L6 vertebrae in their validation sets.

The export ships several views of the same splits for backward compatibility:

• splits_5fold.json — unified splits document (schema v4). Carries the test holdout and all five fold assignments in one file, with split invariants validated at generation time (patient-level disjointness, fold coverage, no overlap between test and trainval, ≥3 lumbarization per fold val).
• splits/test.json — flat list of unique test-set patient tokens (legacy path; still shipped for backwards compatibility).
• data_splits.json — earliest format, mapping ct_file entries to train / val / test. Consumed only as a last-resort fallback.
• splits_summary.json — aggregate per-stratum / per-fold counts.
• manifest_train.json, manifest_validation.json, manifest_test.json — per-record splits derived from the above.

File format

Each case is a pair of gzipped NIfTI files under ct/ and labels/. The filename schema (revised Apr 2026) is fully self-describing:

fused                      ct/<token:04d>_ct.nii.gz
                           labels/<token:04d>_label.nii.gz

spine-side (separate or
spine_only single-mask)    ct/<token:04d>_spine_ct.nii.gz
                           labels/<token:04d>_spine_label.nii.gz

pelvic-side (separate or
pelvic_only single-mask)   ct/<token:04d>_pelvic_ct.nii.gz
                           labels/<token:04d>_pelvic_label.nii.gz

A bare <token>_ct.nii.gz therefore unambiguously means a fused case (both regions present in one mask). The _spine / _pelvic suffix is applied uniformly to spine-side and pelvic-side files regardless of whether the source case is match_type="separate" (paired but non-coregistered) or a spine_only / pelvic_only single-mask case. Earlier the position (supine / prone) appeared in every filename; that was misleading because the prone/supine classifier rarely succeeded, and config is what every downstream consumer actually filters on. position still rides through to the manifest as a metadata column — it is no longer in the filename.

Per-case metadata lives in manifest.json (flat JSON list), with one record per NIfTI pair. Key fields:

token, config, match_type, position, ct_file, label_file, qc_file,
lstv_label, lstv_class, lstv_pelvic, lstv_vertebral,
lstv_agreement, lstv_confusion_zone, has_l6, n_lumbar_labels,
spine_series_uid, pelvic_series_uid,
spine_bone_pct, pelvic_bone_pct,
alignment_ok, ct_resampled_to_mask, postwrite_hip_bone_pct

ct_file and label_file are relative paths with the subdirectory prefix included (e.g. "ct/0017_ct.nii.gz"), so dataset_root / ct_file resolves directly. manifest.csv contains the same content for downstream tooling that prefers tabular ingest.

The two CT-vs-mask diagnostics are worth flagging:

• ct_resampled_to_mask is True when the source CT had to be resampled into the placed mask's grid because shape and/or affine differed. Most cases hit this; the few that don't are cases where the raw CT and placed mask happened to share both shape and affine.
• postwrite_hip_bone_pct is the percent of voxels under the saved label's hip mask that have HU > 200 in the saved CT, computed at export time. Values below ~30% indicate a CT/label frame mismatch and are logged as warnings during export.

Quickstart — NIfTI + nibabel

import json, nibabel as nib
from pathlib import Path

root = Path("anonymous-neurips-ED/CTSpinoPelvic1K")  # or local export dir
meta = json.loads((root / "manifest.json").read_text())

rec = meta[0]
ct  = nib.load(str(root / rec["ct_file"]))
lbl = nib.load(str(root / rec["label_file"]))

assert ct.shape == lbl.shape
assert (ct.affine == lbl.affine).all()

print(rec["token"], rec["config"], rec["match_type"], rec["lstv_class"])

Quickstart — dataset_interface (PyTorch)

The repo ships a dataset_interface.py with a CTSpinoPelvic1K class (directory-backed, no torch dependency — used by the benchmark and visualisation scripts) and a CTSpinoPelvicDataset PyTorch adapter:

from dataset_interface import CTSpinoPelvicDataset
from torch.utils.data import DataLoader

ds = CTSpinoPelvicDataset(
    root  = "anonymous-neurips-ED/CTSpinoPelvic1K",
    split = ("fold", 0, "train"),
)
dl = DataLoader(ds, batch_size=1, shuffle=True)

for batch in dl:
    ct, label = batch["ct"], batch["label"]   # (B,1,Z,Y,X) / (B,Z,Y,X)
    ...

Other supported split values: "trainval" (whole trainval pool), "test" (fixed test holdout), ("fold", i, "val") for fold i's validation side.

Quickstart — MONAI

from monai.transforms import (
    Compose, RandCropByPosNegLabeld, RandFlipd, NormalizeIntensityd,
)
from dataset_interface import CTSpinoPelvicDataset

transforms = Compose([
    NormalizeIntensityd(keys="ct", subtrahend=0, divisor=1000),
    RandCropByPosNegLabeld(keys=("ct","label"), label_key="label",
                           spatial_size=(96,96,96), pos=2, neg=1, num_samples=2),
    RandFlipd(keys=("ct","label"), prob=0.5, spatial_axis=(0,1,2)),
])

ds = CTSpinoPelvicDataset(root="anonymous-neurips-ED/CTSpinoPelvic1K",
                          split=("fold", 0, "train"), transform=transforms)

Citation

Please cite all three source datasets alongside this derivative release:

@misc{smith2015ctcolonography,
  author       = {Smith, K. and Clark, K. and Bennett, W. and Nolan, T. and
                  Kirby, J. and Wolfsberger, M. and Moulton, J. and
                  Vendt, B. and Freymann, J.},
  title        = {Data From CT COLONOGRAPHY},
  year         = {2015},
  publisher    = {The Cancer Imaging Archive},
  doi          = {10.7937/K9/TCIA.2015.NWTESAY1},
  howpublished = {\url{https://doi.org/10.7937/K9/TCIA.2015.NWTESAY1}},
}

@article{deng2021ctspine1k,
  title   = {{CTSpine1K}: A Large-Scale Dataset for Spinal Vertebrae
             Segmentation in Computed Tomography},
  author  = {Deng, Yang and Wang, Ce and Hui, Yuan and Li, Qian and
             Li, Jun and Luo, Shiwei and Sun, Mengke and Quan, Quan and
             Yang, Shuxin and Hao, You and Liu, Pengbo and Xiao, Honghu
             and Zhao, Chunpeng and Wu, Xinbao and Zhou, S. Kevin},
  journal = {Machine Learning for Biomedical Imaging},
  volume  = {3},
  number  = {MICCAI Open Data 2024-2025},
  pages   = {824--832},
  month   = {5},
  year    = {2021},
  doi     = {10.59275/j.melba.2025-gf84},
  url     = {https://arxiv.org/pdf/2105.14711},
}

@article{liu2021ctpelvic1k,
  title   = {Deep Learning to Segment Pelvic Bones: Large-Scale {CT}
             Datasets and Baseline Models},
  author  = {Liu, Pengbo and Han, Hu and Du, Yuanqi and Zhu, Heqin and
             Li, Yinhao and Gu, Feng and Xiao, Honghu and Li, Jun and
             Zhao, Chunpeng and Xiao, Li and Wu, Xinbao and Zhou, S. Kevin},
  journal = {International Journal of Computer Assisted Radiology
             and Surgery},
  volume  = {16},
  pages   = {749--756},
  year    = {2021},
}

BibTeX entries are also provided in CITATION.cff.

License

• Source datasets — CT COLONOGRAPHY (TCIA), CTSpine1K, CTPelvic1K — retain their respective licenses.
• Derivative fused labels, splits, and code: CC BY-NC 4.0 (non-commercial).