Developing new methods for the accurate evaluation of children's bone age based on X-ray images: a study on the...

Bone age assessment is crucial to monitoring children's growth and development. Artificial intelligence (AI)-based bone age assessment systems are widely used in clinical practice, significantly enhancing the efficiency of evaluations and enabling the objective quantification of results. However, obtaining precise assessments can be challenging, and the available evaluation methods for bone age involve certain limitations. This study aimed to propose a novel evaluation method based on a comparative study between an AI system and specialized pediatric radiologists to achieve the goal of precise bone age assessment.

A retrospective analysis was conducted involving the continuous collection of 1,083 images from children with comprehensive clinical data. Bone age was evaluated with both an AI system and three experienced pediatric radiologists applying the Greulich and Pyle (G-P), radius-ulna-short bones Chinese (RUS-CHN), and Tanner-Whitehouse 3 (TW3)-Carpal methods. The children's chronological age served as the reference standard. For intergroup comparisons, AI results were compared with chronological age, manual results with chronological age, and AI results with manual results. This design was applied to identify children with nonstandard bone morphological development. For these comparisons, intraclass correlation coefficients (ICCs) and Bland-Altman plots were used to assess the consistency, the accuracy of the results were each calculated, and the mean absolute error (MAE) was determined.

The detection rate of nonstandard bone morphological development was significantly higher (29.19%) than the rate of children clinically diagnosed with growth and developmental abnormalities (7.83%). The AI evaluation results demonstrated high agreement with manual assessment results (ICC: 0.962-0.982; MAE: 0.38-0.45 years). In contrast, both AI and manual evaluation results showed variable and poor agreement with chronological age across different age groups. The MAE of chronological age ranged from 0.75 to 1.45 years for AI and from 0.68 to 0.95 years for manual assessments. Performance analysis revealed that the different methods had higher consistency and better accuracy for certain age groups. Specifically, the G-P method performed better in the lower age group (approximately 3-5 years old), TW3-Carpal method in the middle-to-lower age group (approximately 7-10 years old), and the RUS-CHN method for the middle-to-higher age group (approximately 11-13 years old).

Although the capability of AI evaluation is comparable to that of specialized pediatric radiologists, both manual and AI evaluations are lacking in precision compared to chronological age. However, the development of a novel AI evaluation method that involves automatically allocating weights and values can address this deficiency to provide greater accuracy in future bone age assessment.