Dual-input deep learning bone age assessment in late adolescence using hand and knee radiographs

To develop and internally evaluate an age-assisted dual-input model for bone age assessment in late adolescence using paired hand-wrist and knee radiographs.

This single-center retrospective study included 547 adolescents aged 13-19 years with paired left hand-wrist and left knee radiographs (1094 images) acquired during routine clinical care. Data were split at the subject level into training, validation, and internal held-out test sets (70%/10%/20%; 382/55/110 subjects), keeping paired images from each subject in the same subset. A ShuffleNet v2-based model used hand and knee radiographs as visual inputs and incorporated chronological age and sex as auxiliary variables. The validation set was used for model selection, and final performance was evaluated on the internal held-out test set using mean absolute error, root mean square error, and intraclass correlation coefficient.

The final model achieved a mean absolute error of 3.77 months, root mean square error of 4.58 months, and intraclass correlation coefficient of 0.984, with 98.18% of predictions within 12 months of the reference standard. Internal ablation comparisons showed lower errors with paired hand-knee inputs and demographic variables than with single-site inputs.

In this single-center retrospective cohort, the age-assisted dual-input framework showed promising internal performance for late-adolescent bone age assessment. Independent multicenter external validation is required before clinical deployment, forensic/disputed-age use, or legal decision-making. Knee radiographs should be used only when already available or medically justified, not as routine additional imaging.