External Validation of Deeplasia for Automated Bone Age Assessment Compared with Four Commercial AI Systems

Background/Objectives: Artificial intelligence (AI)-based systems enable automated bone age (BA) assessment according to the Greulich and Pyle (G&P) method with expert-level performance. Deeplasia is a recently introduced deep learning-based approach that demonstrated promising results in previous studies. This study aimed to externally validate Deeplasia for G&P-based BA and chronological age (CA) estimation. Methods: This retrospective single-center study included two independent cohorts. For BA assessment, 306 children and adolescents aged 1-18 years were analyzed using the mean rating of three expert readers as the reference standard. For CA assessment, 1653 children and adolescents undergoing hand radiography after trauma were included after exclusion of pathological findings. Deeplasia was compared with four CE-certified AI systems. Performance was evaluated using mean error, mean absolute error (MAE), root mean squared error (RMSE), and Bland-Altman limits of agreement. Results: Deeplasia achieved a very good overall agreement with the human reference standard, with the lowest RMSE (0.59 years in boys, 0.55 years in girls) and MAE (0.45 years in boys, 0.43 years in girls). However, no significant differences between the AI systems were observed within the age range representing 90% of the clinically relevant cohort. Estimation of the CA was substantially less accurate than G&P-based BA assessment across all systems. All programs showed systematic overestimation of CA, particularly in adolescent girls. Conclusions: Compared to commercial AI systems, Deeplasia demonstrated excellent external validity for automated G&P-based BA assessment. However, the findings again highlight the intrinsic limitations of G&P-based models for precise CA estimation in contemporary pediatric populations.