A comparative study of the interpretation results of different artificial intelligence bone age assessment software

Objectives To compare the interpretation results of different AI bone age assessment software using the TW3 (Tanner–Whitehouse 3) method. Methods The comprehensive analysis included bone age readings for various age groups (3–17 years) using five methods: Reference-TW3, YIZHUN-TW3, LIANYING-TW3, AIBAA-TW3, and manual-TW3. Researchers calculated mean values and standard deviations, employed linear regression analysis for predictive accuracy comparison, and generated Bland–Altman plots to assess agreement with the reference standard. Results The comparative analysis of different TW3 methods revealed distinct variations in predictive accuracy. Notably, the YIZHUN-TW3 and LIANYING-TW3 models exhibited exceptional predictive accuracy, with multiple R-squared values indicating strong correlation with the reference standard (0.9861 for both methods). This was markedly higher than the AIBAA-TW3 model, which had a multiple R-squared value of 0.9591, and the manual-TW3 method with 0.9732. Statistical significance testing further confirmed the superiority of YIZHUN-TW3 and LIANYING-TW3, with p-values < 0.001 in linear regression analysis, while AIBAA-TW3 and manual-TW3 showed non-significant results in some comparisons. The AIBAA-TW3 model consistently predicted lower bone ages across various age groups, particularly in teenagers, with a notable underestimation trend. This pattern was supported by Bland–Altman analysis, which showed a mean difference of −0.6928 years for AIBAA-TW3, indicating a substantial underestimation. Conclusions The study highlights the advanced predictive accuracy and reliability of the YIZHUN-TW3 and LIANYING-TW3 methods in bone age assessment. While the AIBAA-TW3 method showed strong predictive power, its consistent underestimation trend suggests a need for recalibration.