Severe scoliosis CT segmentation without new labels: bending healthy spines into pathological shapes lifts vertebra success from 73.2% to 81.6%

Vertebral segmentation in severe scoliosis can be shored up with training data synthesized from pathological shape changes, instead of collecting and annotating more real scoliosis CT. High-quality segmentation of severe deformities has depended on labor-intensive manual annotation of CT or MR scans, and public scoliosis datasets are scarce; standard augmentation models local variability such as intensity perturbations, noise, cropping and mild elastic deformation, and misses the global curvature change. On 18 in-house adolescent scoliosis CTs, the CNN with VerSe+2x under default augmentation raised mean Dice from 0.765 to 0.842, vertebra success rate from 73.2% to 81.6%, and cut HD95 p95 from 31.082 mm to 24.900 mm; generic elastic deformation alone reached only 0.794 mean Dice.

The synthetic data bends healthy spine CT and its labels together into scoliosis-like shapes while keeping vertebral order intact. The pipeline extracts vertebral centroids from the label mask, fits a spinal centerline, deforms it into a randomized pathological curve with a fourth-degree Bezier curve, and builds local frames and circular polar control grids along the curve. Control points are ordered by an arc-length monotonicity constraint to avoid fold-overs, then tessellated by Delaunay triangulation into a tetrahedral mesh for a piecewise-linear volumetric backward map, resampling CT with linear interpolation and masks with nearest-neighbor, followed by fully automated tube masking and morphological opening and closing.

The pipeline targets vertebral segmentation in severe scoliosis, trains on public VerSe 2019/2020 data, and is evaluated on 18 in-house adolescent scoliosis CTs and the 1005-scan Spine1K; the in-house data are not public because of privacy and institutional restrictions. Synthetic generation and post-processing are fully automated with no manual inspection or curation, and the authors state that not every generated volume corresponds to clinically realistic anatomy.

The in-house test cohort has only 18 cases, 17 with Cobb angles; the authors note that warping and resampling can introduce interpolation effects or local distortions, and report that on the in-house cohort the gain from geometric diversity outweighed those artifacts in aggregate. The severity analysis is exploratory and the authors call for validation on broader multi-center cohorts with pre-defined severity bins and evaluation across scanners and protocols; the elastic baseline was tuned once on a single CNN VerSe fold and then frozen, and model scale and training budget were fixed without extensive model-specific tuning.