A radiation-free screening system for adolescent idiopathic scoliosis using deep learning on 3D back surface point clouds

Authors

Abstract

Widespread screening for Adolescent Idiopathic Scoliosis (AIS) is critical for timely intervention but is currently constrained by the radiation risks of X-rays and the subjectivity of physical examinations. Here, we present PointScol, a radiation-free triage system leveraging 3D back surface point clouds. To reconcile the conflicting clinical demands for "zero-miss" screening and "fine-grained" severity assessment, we developed a two-stage deep learning framework. First, an automated segmentation module extracts the dorsal region of interest (ROI) to standardize input geometry. Second, the system employs a dual-branch diagnostic strategy: a binary classification network designed for maximal sensitivity to rule out health, and a 5-class grading network designed to stratify severity (0-10{degrees}, 11-20{degrees}, 21-30{degrees}, 31-40{degrees}, >40{degrees}). Validation on a multi-center dataset (n=128) confirmed the distinct utility of this hierarchical approach. For the scoliosis screening task using a 10{degrees} Cobb angle threshold, the binary classification model achieved a sensitivity of 100.00% in the external cohort, ensuring that no cases requiring further clinical attention were missed. While the 5-class grading task inherently faces greater complexity, it successfully achieved an overall accuracy of 84.48% and, crucially, demonstrated a high specificity of 98.42% for severe surgical cases (>40{degrees}). This performance profile establishes PointScol as a safe clinical filter: the binary module reliably excludes healthy individuals, while the 5-class module flags high-risk patients for prioritized intervention, collectively offering a non-invasive, resource-efficient paradigm for scoliosis management.

Tags

Topics