Bone age estimation from lateral cephalograms using cervical vertebrae–centered ROIs: a multicenter deep learning study

This multicenter retrospective study compared cervical vertebrae (CV)-centered region-of-interest input strategies for bone-age estimation from lateral cephalograms and evaluated whether incorporating upper-airway regions provides complementary information, with emphasis on cross-site generalizability. In total, 1,100 paired lateral cephalograms and hand–wrist radiographs (ages 4–18 years) from two institutions were analyzed. The second to fourth cervical vertebrae and upper airway were manually delineated to generate three input strategies: CV masking, combined CV and airway masking, and CV-centered cropping without masking. Reference bone ages were AI-derived surrogate labels obtained from commercial hand–wrist bone-age software. A deep-learning regression model was evaluated using five-fold cross-validation, an internal test set, and an external test set. Performance was comparable across strategies under internal evaluation but differed under external validation. Although overall errors increased substantially in the external test set, CV-only masking showed the largest degradation (mean absolute error, 1.91 years); airway-inclusive masking (1.72 years) and unmasked CV-centered cropping (1.67 years) yielded significantly lower errors. Interpretability analyses highlighted mainly cervical vertebral structures. However, Grad-RAM showed prominent airway activation for airway-inclusive inputs. Together, these findings suggest that CV-centered inputs preserving surrounding anatomical context are more robust for cross-site bone-age estimation than strictly vertebra-only masking.