AI model for automatic spinopelvic and spinal alignment parameters measurement from EOS

PURPOSE: To evaluate the reliability of spinopelvic and spinal alignment parameters measurements generated by Gleamer BoneMetrics® artificial intelligence (AI) tool on full-spine EOS® exams. MATERIALS AND METHODS: Consecutive full-spine EOS® radiographs were retrospectively evaluated to extract a total of 12 spinopelvic and spinal alignment parameters, including Cobb angle, thoracic kyphosis, lumbar lordosis, pelvic incidence, sacral slope, pelvic tilt, sagittal vertical axis (SVA), and spino-sacral angle automatically obtained by Gleamer BoneMetrics®. Each parameter was manually measured by a final-year radiology resident, blinded to AI results. The first 100 cases were reviewed jointly with a senior musculoskeletal radiologist; remaining doubtful cases were resolved by consensus. Agreement was assessed using Bland–Altman analysis, intraclass correlation coefficients (ICC[3,1] and ICC[A,1]), Pearson’s correlation, root-mean-square error (RMSE), and predefined clinical thresholds (± 2°/±3 mm and ± 5°/±5 mm). RESULTS: Two-hundred exams were analyzed (mean age 27.3 ± 21.8 years; 69.5% females). For most parameters (n = 200; Cobb angle n = 174), Gleamer and radiologist measurements showed excellent agreement: Pearson’s r = 0.92–0.99 (all p < 0.001), ICCs = 0.912–0.997, with small biases (median 0.65, maximum 1.4). RMSE ranged 0.7–5.0. Bland–Altman plots showed centered biases and symmetric limits of agreement. Significant heteroscedasticity was detected for some pelvic and coronal parameters. Clinical threshold concordance was high: 39.5–99.5% within ± 2° and 75–100% within ± 5° for angular measures, and 81–98% within ± 3 mm and 93.5–99.5% within ± 5 mm for linear measures. CONCLUSION: Gleamer BoneMetrics can generate reliable and accurate spinopelvic and spinal alignment parameters from EOS®.