P4.014: Advancing Pediatric Growth Analysis: A Systematic Review of Bone Age Assessment Using Artificial Intelligence

Background: Artificial Intelligence (AI) is increasingly utilized in pediatric bone age assessment, aiming to enhance accuracy and reduce interobserver variability. However, clinical implementation remains limited due to insufficient validation across diverse populations and the lack of FDA-approved models.

Methods: We conducted a systematic review of PubMed and IEEE Xplore publications evaluating AI-based bone age assessment models. Studies were included if they performed external or prospective validation, adhered to gold-standard methods, and compared outcomes to expert consensus or chronological age. Study quality was assessed using the Oxford Level of Evidence, and a diagnostic protocol was synthesized from the data.

Results: We included 17 studies that validated 7 AI models for bone age assessment. The two most extensively validated models were BoneXpert and 16-bit Bone Age-Physis. BoneXpert was validated in 7 studies with 4,833 samples across diverse populations, demonstrating a mean absolute error (MAE) ranging from 0.03 to 0.58 years. However, it rejected 50% of cases with achondroplasia and showed variability in accuracy depending on age and sex. 16-bit Bone Age – Physis was validated in 4 studies with 6,067 samples, reporting a MAE between 0.35 and 1.1 years, with limitations in distinguishing inappropriate images and maintaining accuracy after image resolution adjustments. The remaining 5 models (Vuno-BoneAge, IB LAB-Panda, Deeplasia/AIMI-Bonn, Deep Wise Intelligence Lab, and YITU-ZheJiang) had limited validation, with small sample sizes or restricted to specific populations, making their generalizability uncertain.

Conclusion: BoneXpert and 16bit Bone Age-Physis demonstrated the most robust validation among AI models for bone age assessment. However, further revalidation is required before clinical implementation, particularly in diverse populations and pathological conditions. The remaining models showed limited external validation, highlighting the need for further studies. No AI models for bone age assessment have received FDA approval.