Scoliosis Detection System using Zero-Shot segmentation

Scoliosis is a progressive spinal disorder involving abnormal lateral deviation and vertebral rotation, which may lead to functional and structural complications if not detected early. The Cobb angle is widely accepted as the primary clinical indicator for quantifying spinal curvature and guiding treatment decisions. However, conventional Cobb angle measurement from radiographic images requires manual identification of vertebral landmarks, making the process time-intensive and susceptible to observer variability. While recent artificial intelligence-based approaches have improved automation in spinal analysis, most existing methods depend on extensive annotated datasets and supervised training, which restricts generalizability across institutions and imaging protocols.

This study introduces a scoliosis detection framework built upon zero-shot segmentation using a foundation vision model. The proposed system employs the Segment Anything Model (SAM), a large-scale pre-trained Vision Transformer, to generate segmentation masks directly from spinal radiographs without additional task-specific training. The most anatomically relevant mask is selected and refined using morphological operations to enhance structural continuity. Subsequently, skeletonization and regression-based geometric fitting are applied to estimate spinal orientation and compute the Cobb angle automatically. Severity categorization is performed according to established clinical thresholds. By removing the dependency on labelled datasets and retraining, the proposed method provides a scalable, adaptable, and interpretable solution for automated scoliosis screening and curvature estimation in real clinical environments.