Spine labels no longer assume 12 thoracic and 5 lumbar: anomaly-aware labeling lifts all-vertebra accuracy to 99% on CT

Vertebra labeling now returns correct level labels when the spine carries an enumeration anomaly, rather than numbering straight through the usual count. Earlier automatic methods predicted at most twelve thoracic and five lumbar vertebrae, and the few that theoretically allowed anomalies were either unreliable at arbitrary fields of view or never validated on a large dataset. On 1000 T2w sagittal MRI images, 98.30% of subjects had all vertebrae correctly labeled versus 94.24% for the comparison method; on 365 CT images the figures were 99.18% versus 77.26%, with Wilcoxon signed-rank tests giving p < 0.01.

The method judges each vertebra on its own, then solves the whole sequence as one constrained optimization problem. The usual design lets a single classification head emit labels such as C3 or T5 and takes the per-vertebra maximum. A 3D DenseNet169 trained on individual vertebra crops carries four output heads for vertebra label, region, transition position and visibility; a dynamic-programming sequence predictor searches for the lowest-cost path under constraints on consecutive labels, spatial order and permitted cervical, thoracic and lumbar counts, then rewrites two consecutive T12 or L5 labels as T13 or L6.

The same model returns a nearly unchanged sequence when images are cropped to different numbers of visible vertebrae. Under a limited field of view, conventional counting from the cranial or sacral end cannot fix the thoracolumbar level. Cropping test images to every combination of visible vertebra counts left predictions almost identical to those on the whole spine, with thoracic anomaly recall dropping only for a few samples at very small fields of view.

The method targets vertebra labeling in CT and T2w sagittal MRI at arbitrary fields of view, and the test sets were deliberately enriched with enumeration anomaly cases for evaluation. It performs no localization or detection and requires the centroid of each vertebra beforehand; the authors note that public segmentation models already supply this step. The MRI model was trained only on the healthy NAKO cohort, so the authors state that the same performance cannot be guaranteed on other routine clinical data; the CT data are diverse, and the authors expect more robust behavior in that modality.

The authors state that because the MRI model was trained only on a healthy cohort, the same performance cannot be guaranteed on other routine clinical datasets. At very small fields of view, thoracic anomaly recall drops, which the authors attribute to individual vertebrae being hard to judge rather than to particular fields of view. The method contains no localization step, so an additional vertebra localization is needed before use.