CTSpinoPelvic1K: spine, pelvis, ribs and femora in one coordinate frame, annotated for lumbosacral transitional anatomy
Gregory Schwing, Ashley Schehr, Annika Tekumulla, Margret Khoushi, Ryan Christian, Dane Hubers, Faris Mahjoub, Hassan Saad, et al.
cs.AI · cs.LG
Quick verdict
Joins two public label sets into one 802-CT spine-pelvis-rib dataset with rare anomaly classes; unusually transparent, though key labels stay reader-derived.
Summary
This dataset paper releases CTSpinoPelvic1K: 802 abdominopelvic CT records from the ACRIN 6664 colonography trial that join CTSpine1K's vertebra labels and CTPelvic1K's pelvic masks on one series per patient, adding ribs, femora and surgical hardware, plus classes for a sixth lumbar vertebra, T13 and lumbar ribs. Because the two sources had often annotated different series of the same patients (only 49.4% doubly annotated fell on the same series), the authors built a bone-overlap crosswalk, corrected 112 spine and 13 pelvic labels by hand, and completed missing pelvises with an out-of-fold five-fold nnU-Net ensemble (sacrum Dice 0.980). Ribs unite TotalSegmentator and Möller's network; 152 triaged records were human-corrected. Validation reports geometric invariants passing 802/802, rib-vertebra concordance in 5,751 of 5,753 evaluable ribs (two one-level offsets, 0.035%), and spinopelvic means matching published supine CT (pelvic incidence 52.1° vs 52.1°). A logistic regression on five morphometric measures separates T12 from L1 at AUC 0.990, but no classifier is attempted at the lumbosacral junction, the motivating question. Main limitations, largely self-declared: no held-out test set; ribs are triaged-review pseudolabels; Castellvi grades cover only 33 records; thoracic truth is field-of-view limited; border-vertebra names remain reader-derived; and the abstract's 'separate S1' class conflicts with the body, which retires it.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- CTSpinoPelvic1K is the first public CT collection placing spine, pelvis, ribs and femora in one coordinate frame while expressing enumeration anomalies (L6, T13, lumbar ribs) as explicit classes across 802 records from a single prospective protocol.
- A bone-overlap crosswalk correctly recovers which series each source annotation was drawn on, agreeing with 558/565 CTSpine1K and 702/713 CTPelvic1K self-declared mappings.
- Local morphology separates T12 from L1 with AUC 0.990 and 97.3% accuracy over 1,485 vertebrae using five released per-level measures plus six ratios, demonstrating decidability at the thoracolumbar border.
- Released spinopelvic means (pelvic incidence 52.1±11.1°, n=756) and anomaly prevalences (stump ribs 12.4%, lumbar ribs 2.0%) reproduce published whole-spine CT values.
- Whether local morphology resolves lumbosacral naming remains open — the release supplies the classes and measurements but not yet enough transitional records.
Strengths
The crosswalk is a real, documented curation contribution
Sec. II.A shows why a naive join fails: CTSpine1K's Path.csv names no position for 217/782 records, CTPelvic1K's filename-encoded series field matches nothing, one mask names a patient with no archived series. Bone-overlap resolution agrees with 558/565 and 702/713 self-declared mappings, exposing that the two sources annotated the same series for only 342/693 (49.4%) doubly annotated patients.
Anomaly-aware class vocabulary is genuinely novel for public CT
Table I shows no prior collection carries L6 plus pelvis plus ribs plus a lumbar-rib class; this release adds ids 60/61 for lumbar ribs (the one class with no published counterpart) and 62-68 for hardware, so phenotypes (b), (c), (d) of Fig. 2 become writable as distinct data rather than three readings of one count.
Validation discipline with stated denominators and gate-capability testing
External anchoring of measurements and prevalences
Complete release and honest LLM disclosure
Weaknesses
Ground truth at the motivating junction is reader-derived, creating circularity risk
The central lumbosacral question is left unattempted
Label strength is heterogeneous and validated only at Dice level where weakest
No held-out test set
Abstract-body inconsistencies suggest version drift
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
CTSpinoPelvic1K is a dataset release, not a method paper: 802 abdominopelvic CTs from the single-protocol ACRIN 6664 colonography trial (patients aged 50+, supine and prone, 15 centers, 5 scanner vendors), pairing CTSpine1K vertebra masks with CTPelvic1K pelvis masks on one series per patient and adding ribs, femora and surgical hardware under a VerSe-native identifier scheme extended to 68. The motivating claim is that existing public collections cannot express enumeration anomalies at the lumbar borders — TotalSegmentator has no L6 or lumbar-rib class, VerSe stops at the sacrum — while this release can (L6=25, T13=28, lumbar ribs 60/61, hardware 62–68). Secondary claims: the bone-overlap crosswalk recovers the true source series (558/565 and 702/713 agreement with self-declared mappings); local morphology decides T12-vs-L1 at AUC 0.990; and spinopelvic measures reproduce published supine CT values.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the CTSpinoPelvic1K dataset?
A public dataset of 802 abdominopelvic CT scans (from the ACRIN 6664 colonography screening trial) with segmentation labels for vertebrae, sacrum, pelvis, ribs, femora and surgical hardware in one coordinate frame. It uniquely includes explicit classes for anatomical numbering anomalies — a sixth lumbar vertebra, a thirteenth thoracic vertebra, and lumbar ribs — built by joining the CTSpine1K and CTPelvic1K label sets on the same image series.
Why does lumbosacral transitional vertebra (LSTV) numbering matter?
When a person has four or six lumbar-type vertebrae instead of five, the same count can mean different anatomies (four may be a lumbar rib on L1 or an L5 fused to the sacrum; six may be a genuine L6, a T12 with rudimentary ribs, or a lumbarized S1). Miscounting levels contributes to wrong-level spine surgery, reported at roughly 1 in 3,110 procedures, and lumbar surgery is planned on scans that lack the C2 reference the gold-standard counting method requires.
Can this dataset train a model to name vertebrae from local shape alone?
It demonstrates feasibility at the thoracolumbar border: a logistic regression on five released shape measures separates T12 from L1 with AUC 0.990 over 1,485 vertebrae. The harder lumbosacral question is left open because only about 34 transitional records and 33 Castellvi grades exist in the release; the paper ships the per-vertebra measurements needed to attempt it.
Where can I download CTSpinoPelvic1K and its code?
The data are archived on Zenodo (concept DOI 10.5281/zenodo.22139642, v11 at 10.5281/zenodo.22797014) as NIfTI image/label pairs with frozen patient-grouped, LSTV-stratified five-fold splits. Code, the QC scripts behind every figure and table, the reference loader, and nnU-Net checkpoints are on GitHub (OpenSpineConsortium/CTSpinoPelvic1K, OpenSpineToolkit) and Hugging Face.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (Fig. 2 phenotypes, Fig. 3 morphometry, S1-S3) are described textually but not visually verifiable from the extracted text.
- Numbers could not be checked against the live Zenodo/GitHub artifacts; the abstract-body discrepancies (S1 class, 5,749 vs 5,753 ribs) may reflect version drift between the arXiv listing and v11.
- External anchors (Veilleux, Hasegawa, Nagata, Panjabi comparators) are taken as reported; the cited source papers were not consulted.
- Whether the released archive matches the described contents (splits_5fold.json, per-fold Dice files, three unfinalized rib reviews) is unverified.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.