Not All Spines Are Created Equal: How CT Segmenters Fail on Lumbosacral Transitional Vertebrae

Wrong-level spine surgery remains a persistent never-event, with the Lumbosacral Transitional Vertebra (LSTV, 5–35% prevalence) as its largest anatomic driver. We audit TotalSegmentator — the most widely deployed CT segmentation system and default back-end of 3D Slicer — and show that on lumbarization cases its output is anatomically indistinguishable from a wrong-level surgical plan: lacking an L6 class, it shifts lumbar labels caudally by one level, with junction-DSC dropping 29 points on Any-LSTV. No existing CT benchmark surfaces this — not TotalSegmentator's own evaluation, not Li et al. (2025) (the per-class DSC frontier on the matched COLONOG split), and not VERIDAH (vertebra-labeling SOTA on a private CT cohort, operating on pre-localized crops with no pelvis). We contribute: (i) an LSTV-stratified evaluation protocol (per-class DSC, junction-DSC over a 40 mm L5/S1 window, per-class voxel confusion at the junction) translating the level-shift mechanism into quantifiable wrong-level surgical risk; (ii) CTSpinoPelvic1K — 1,153 CT volumes across 802 patients with unified spinopelvic masks, 33 LSTV-positive cases across a 6-way phenotype taxonomy, and radiologist Castellvi typing (I–IV) on all 33; (iii) the first publicly-released merge-based LSTV-handling dual-spinopelvic CT segmenter, exceeding TotalSegmentator on sacrum (TS=0.817 zero-shot) and uniquely delivering both spine and pelvis in a single forward pass with released weights; (iv) an empirical isolation showing VERIDAH's training-side L5/L6 merge eliminates the collision driving TotalSegmentator's level shift but exposes a residual L4/last_lumbar collision on count-style sacralization, motivating VERIDAH's sequence predictor as the necessary disambiguation component. CTSpinoPelvic1K, the protocol, and 5-fold checkpoints are released publicly.