Multi-centre Validation of a Deep Learning Model for Scoliosis Assessment

Scoliosis affects roughly 2–4% of adolescents, and treatment decisions depend on precise Cobb-angle measurement. Manual assessment is slow and subject to inter-observer variation. We conducted a retrospective, multi-centre evaluation of a fully automated deep-learning software (Carebot AI Bones, Spine Measurement functionality; Carebot s.r.o.) on 103 standing anteroposterior whole-spine radiographs collected from ten hospitals. Two musculoskeletal radiologists independently measured each study and served as reference readers. Agreement between the AI and each radiologist was assessed with Bland–Altman analysis, mean absolute error (MAE), root-mean-squared error (RMSE), Pearson correlation coefficient, and Cohen’s  \(\kappa \) for four-grade severity classification. Against Radiologist 1 the AI achieved an MAE of \(3.89^{\circ }\) (RMSE \(4.77^{\circ }\) ) with a bias of \(0.70^{\circ }\) (limits of agreement \(-8.59^{\circ }\) to \(9.99^{\circ }\) ); against Radiologist 2 it achieved an MAE of \(3.90^{\circ }\) (RMSE \(5.68^{\circ }\) ) with a bias of \(2.14^{\circ }\) (limits \(-8.23^{\circ }\) to \(12.50^{\circ }\) ). Pearson correlations were \(r = 0.906\) and \(r = 0.880\) (inter-reader \(r = 0.928\) ), while Cohen’s  \(\kappa \) for severity grading reached 0.51 and 0.64 (inter-reader \(\kappa = 0.59\) ). These results show that the proposed software reproduces expert-level Cobb-angle measurements and categorical grading across multiple centres, suggesting its utility for streamlining scoliosis reporting and triage in clinical workflows.