AI-Based Longitudinal Scoliosis Monitoring on EOS Whole-Spine Radiographs in a Pediatric to Young Adult Cohort

Manual Cobb angle measurement can exhibit high interrater variability, potentially affecting scoliosis management decisions. This study aimed to examine the accuracy of an artificial intelligence (AI) software for automated Cobb angle measurement compared to manual assessments on coronal EOS (EOS Imaging, Paris, France) radiographs of the spine in a pediatric to young adult cohort.

In this retrospective single-center study we selected patients from 5-25 years with confirmed scoliosis and multiple follow-up exams. Cobb angles and coronal balance were measured on EOS radiographs: manually by two expert readers, representing the reference standard, and an AI-driven commercial software (IB Lab SQUIRREL, IB Lab GmbH, Vienna, Austria). Performance among readers and the AI was compared using mean differences, intraclass correlation coefficients (ICC), Pearson's correlation coefficients, Skillings-Mack and Bland-Altman analyses.

A total of 144 anteroposterior whole-spine EOS radiographs from 29 patients were included (median age 13 years, 8 male). The AI demonstrated excellent agreement with human readers for all measurements (ICCs 0.969-0.993). Mean differences were minimal: (0.2°-0.7°) for Cobb angles and 0.7 mm for coronal balance. Skillings-Mack analysis confirmed no significant temporal drift in AI measurement errors across sequential examinations (all p > 0.05). The AI identified scoliosis progression (increase of or over 5° between measurements) with 90.9% accuracy.

AI measurements demonstrated excellent absolute and longitudinal agreement with expert human readers for longitudinal Cobb angle measurement across multiple follow-ups, offering potential for improved diagnostic consistency and efficiency in scoliosis assessment.