AI-driven Assessment of Vertebral Rotation in Adolescent Idiopathic Scoliosis Using Weight-bearing Cone-beam CT.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, West China Hospital of Sichuan University, No. 37 Guoxue Alley, Chengdu 610041, China (J.H., Z.L., Y.Z., Y.Z., M.Z., Z.L., J.T.).
- Shenzhen Angell Technology Co., Ltd., Shenzhen, China (F.R., Z.C., H.Z.).
- Department of Procurement and Supply, West China Hospital of Sichuan University, Chengdu, China (H.L.).
- Department of Orthopedics, Chengdu Seventh People's Hospital, Chengdu, China (J.L.).
- Department of Radiology, West China Hospital of Sichuan University, No. 37 Guoxue Alley, Chengdu 610041, China (J.H., Z.L., Y.Z., Y.Z., M.Z., Z.L., J.T.). Electronic address: [email protected].
Abstract
RATIONALE AND OBJECTIVES: To evaluate an artificial intelligence (AI)-based algorithm for quantifying vertebral rotation angles on weight-bearing cone-beam CT images in adolescent idiopathic scoliosis (AIS) patients, and to investigate how well the 2D Nash & Moe grading system represents 3D vertebral rotation under weight-bearing conditions. MATERIALS AND METHODS: This prospective study included 72 adolescents with AIS. Six predefined vertebrae per patient were manually measured by two radiologists, while all vertebrae were also processed using a U-Net-based algorithm to obtain automated rotation angles. Nash & Moe grades were assessed from standing radiographs. Agreement was evaluated using intra-class correlation coefficients (ICCs), Bland-Altman analysis, and correlation tests. The explanatory power of Nash & Moe grades for 3D vertebral rotation was analyzed by linear regression. RESULTS: Automated vertebral rotation angles demonstrated good agreement and strong correlation with manual measurements (ICC = 0.881, 95% CI 0.852-0.905; Spearman correlation coefficient = 0.900, 95% CI 0.871-0.920). Linear regression showed strong relationships between automated and manual measurements for thoracic (R<sup>2</sup> = 0.776) and lumbar vertebrae (R<sup>2</sup> = 0.792). Nash & Moe grades exhibited moderate explanatory power for 3D rotation measured by the algorithm (R<sup>2</sup> = 0.549 thoracic and R<sup>2</sup> = 0.628 lumbar), with similar results for manual measurements (R<sup>2</sup> = 0.532 thoracic; R<sup>2</sup> = 0.621 lumbar). CONCLUSION: The AI-driven algorithm showed good agreement with manual measurements and enabled reliable automated quantification of vertebral rotation under weight-bearing conditions, whereas the 2D Nash & Moe method provides only approximate estimates and lacks precision for detailed assessment. This automated 3D approach may improve scoliosis evaluation and support more informed clinical decision-making in AIS management.