Learning from the unseen: synthetic data for scoliosis AI segmentation.
TL;DR
Pathology-informed synthetic data improve robustness to severe scoliosis without new annotations or architecture changes in nnU-Net CNN and PRIMUS models under matched compute.
Abstract
Purpose
Severe scoliosis remains challenging for 3D vertebral segmentation because annotated pathological computed tomography (CT) data are scarce.
Methods
We synthesize scoliosis-like CT volumes by jointly warping public healthy spine scans and labels and then train nnU-Net CNN and PRIMUS models under matched compute. We evaluate on in-house out-of-distribution (OOD) adolescent scoliosis CTs and the public Spine1K dataset.
Results
On 18 in-house adolescent scoliosis CTs, the best CNN setting increased mean Dice from 0.765 to 0.842 and vertebra success rate from 73.2 to 81.6%. In 55 surgeon-defined apical labels, mean Dice increased from 0.720 to 0.854. Spine1K performance remained high when synthetic data supplemented real scans; synthetic-only training was less stable.
Conclusion
Pathology-informed synthetic data improve robustness to severe scoliosis without new annotations or architecture changes. Code and interactive visualizations are available at: https://chepchikvaleri.github.io/synth-data-s/.