Trustworthy Vertebral Segmentation in Spine CT via External Validation and Per-Vertebra Reliability Assessment

Abstract Purpose. Clinical adoption of automated vertebral segmentation from spine CT depends on two forms of trust: performance under genuine external validation rather than same-source testing, and a signal for when an individual prediction is unreliable. We evaluate nnU-Net, decompose the cross-domain gain into training-data expansion versus ensembling, and test whether per-vertebra ensemble disagreement can flag vertebrae for review. Methods. An nnU-Net v2 residual-encoder (ResEnc-M) model was trained on VerSe alone (112 cases) and on an expanded VerSe-plus-CTSpine1K set (387 cases), then evaluated on the held-out CTSpine1K validation split (191 cases), external to the VerSe-only model. Single-model and five-fold ensemble inference were compared. Per-vertebra Dice (DSC) was computed against reference labels; ensemble disagreement was tested against DSC by Spearman correlation and AUROC with bootstrap CIs. Results. The expanded ensemble reached a mean per-vertebra Dice of 0.9206 (lumbar 0.957, cervical 0.888). Five-fold ensembling raised single-source Dice from 0.8315 to 0.8779 (+ 4.6), exceeding the gain from adding a second source (0.8578, + 2.6). Ensemble disagreement predicted per-vertebra error (Spearman 0.862; AUROC 0.920); flagging vertebrae below 0.8 fold agreement captured 76.1% of poorly segmented vertebrae while referring 9.4% for review. Conclusion. Ensemble disagreement provides a validated, cost-free per-vertebra reliability flag that predicts segmentation error and localizes review, catching about three-quarters of poor vertebrae while referring under one-tenth. Ensembling drove more cross-domain gain than added data here. The emphasis is trustworthy deployment of an existing model, not a new one.