Proof-of-concept comparison of an artificial intelligence-based bone age assessment tool with Greulich-Pyle and Tanner-Whitehouse version 2 methods in a pediatric cohort

Comprehensive Summary

This study compared an artificial intelligence-based software, BoneView BoneAge, against two widely-used clinical methods for estimating bone age: the Greulich–Pyle (GP) atlas and the Tanner–Whitehouse (TW2). This AI application was trained on data from the GP method and incorporated sex and chronological age as inputs to produce a decimal estimate of skeletal age from left hand and wrist radiographs of individuals under 18. Data for this study was procured based on patient age (2–17 years), availability of a clear left hand radiograph, and complete demographic information. Two radiologists independently assessed bone age using the GP atlas for 157 patients, with the mean of their estimates used for statistical analysis, while an endocrinologist evaluated a subset of 35 patients using the TW2 method. Demographic information was available to both clinicians and the AI model during evaluation. All 3 methods were assessed based on the patients’ chronological age. Bias, reflecting trends in under or overestimation, was minimal for the AI model (−0.05 years), mild for human GP assessments (−0.43 years), and pronounced for TW2 evaluations (−2.63 years). Mean absolute error (MAE) was similar for the AI model and GP method (1.38 and 1.30 years, respectively), while the TW2 method showed a substantially higher MAE of 2.86 years. Root mean square error (RMSE) followed a similar pattern, measuring 1.75 years for the AI model, 1.80 years for human GP, and 3.88 years for TW2. Correlation between chronological and estimated age also reflected this trend, with coefficients of 0.857 for the AI model, 0.894 for human GP, and 0.490 for TW2, indicating strong correlations for the AI and GP methods but only moderate correlation for TW2. While the reliability of AI-based bone age estimation has been supported by previous studies, the poorer performance of TW2 is likely attributable to limitations in benchmarks derived from a monocultural reference population. Interobserver agreement between the two radiologists using the GP atlas yielded a Cohen’s coefficient of 0.72, indicating high consistency, and substantial agreement was also observed between the AI model and radiologists. However, the AI model occasionally misidentified normal anatomical features as fractures, underscoring the need for further pediatric-specific training and continued human oversight. Study limitations include the small size of the TW2 subgroup, the absence of stratification by ethnicity or pubertal stage, and the retrospective and monocentric nature of the dataset.

Outcomes and Implications

Bone age estimation is a critical tool in pediatric endocrinology and radiology, enabling the prediction of skeletal growth and the diagnosis of conditions affecting or resulting from the bone. Following a 2017 AI bone age assessment competition, numerous software applications for automated bone age estimation have been introduced, including BoneView BoneAge. While these tools are valuable in bone age estimation by reducing interobserver variability and streamlining the assessment process, their reliability is limited by variations in skeletal development across different populations. Independent evaluations of AI-based systems have highlighted insufficient training on diverse patient populations. Given that bone age assessment methods directly inform diagnosis and guide future patient care, such as in cases of growth delay or precocious puberty, high accuracy through diversity is essential.