Scoliosis CT labels without hand annotation: bending healthy spines into pathological shapes lifts vertebra Dice from 0.765 to 0.842
Synopsis
Vertebra segmentation in severe scoliosis has been limited by scarce annotated pathological CT, since public data contain almost no scoliotic cases. An anatomy-aware synthesis pipeline jointly deforms public healthy spine CT scans and their segmentation labels into scoliosis-like volumes, then trains nnU-Net CNN and PRIMUS models under matched compute. On 18 in-house adolescent scoliosis CTs, the best CNN setting raised mean Dice from 0.765 to 0.842 and vertebra success rate from 73.2% to 81.6%; on 55 surgeon-defined apical labels, mean Dice rose from 0.720 to 0.854.
Interpretation
Adding synthesized scoliosis-like CT volumes to the training set clearly improves vertebra segmentation on real severe cases, with no change to network architecture or inference. Prior practice relied on labor-intensive manual annotation of pathological CT or MR scans, while generic elastic deformation models only local variability and misses the global shape change of scoliosis. On 18 in-house adolescent scoliosis CTs, VerSe+2x with default augmentation raised the CNN's mean Dice from 0.765 to 0.842, vertebra success rate from 73.2% to 81.6%, and reduced HD95 p95 from 31.082 mm to 24.900 mm; adding only generic elastic deformation reached just 0.794 mean Dice.
The deformation is not random perturbation: a spinal centerline is reconstructed and then bent into a scoliosis-like curve, with image and label deformed together. Conventional augmentation perturbs intensities, mirrors, or applies local elastic warps on voxels, without guaranteeing that vertebral ordering and topology stay consistent after bending. The pipeline extracts vertebral centroids from the mask, fits a cubic spline centerline, generates C- or S-shaped curvature with a fourth-degree Bézier curve, builds local frames by parallel transport, realizes a volumetric backward map through a tetrahedral mesh, and enforces arc-length monotonicity to prevent control-plane crossing; training sets were generated from 408 public CT volumes after merging and deduplicating VerSe 2019/2020.
Synthetic data supplement real scans rather than replace them. Training on synthetic data alone was less stable, particularly for the transformer model. On the 1005-scan Spine1K public benchmark, mixed real-synthetic training gave the CNN a best mean Dice of 0.908 versus 0.904 for the baseline; the transformer rose from 0.886 to 0.903 and its vertebra success rate from 91.0% to 92.5%; 4x-only training was weaker, with a large HD95 tail outlier for the transformer under the default setting.
Perspective
The pipeline targets vertebra segmentation in severe scoliosis, with training data drawn from public VerSe 2019/2020 healthy spine CT. Evaluation covers 18 in-house adolescent scoliosis CTs, 55 surgeon-verified apical labels, and the 1005-scan Spine1K public benchmark. Synthetic volumes are generated offline, and generation and post-processing are fully automated without manual curation. The in-house pediatric dataset is not publicly available due to privacy and institutional restrictions, while public training data are available through the VerSe challenge resources.
The authors note that warping and resampling can introduce interpolation effects or local distortions that may reduce fidelity in some regions; the net effect on segmentation was positive in aggregate, but systematic characterization of artifact types and their downstream impact remains future work. The elastic baseline was tuned once on a single CNN VerSe fold and then frozen, so broader architecture-tailored tuning could strengthen that comparison. The severity analysis was exploratory, with a small sample and uneven angle distribution, and should be validated on broader multi-center cohorts using predefined severity bins and evaluation across scanners and protocols. Because deformation is computationally expensive during training, synthetic volumes were generated offline; more efficient generation and stratified analyses by etiology remain open.
