digiBONE: An automated tool for segmental Greulich-Pyle bone age assessment of Indian children and adolescents

Abstract

Accurate bone age assessment (BAA) is essential for diagnosing and managing pediatric endocrine and growth disorders. Currently available BAA automation methods often overlook that skeletal maturation progresses at different rates in the anatomical regions of the hand, leading to potential diagnostic variability. This study presents digiBONE, a deep learning–based BAA framework that explicitly incorporates segmental skeletal maturity information to enhance prediction accuracy and interpretability. Using convolutional neural networks (CNN), we trained separate models for full-hand radiographs and models of anatomically defined segments: short bones, carpals, and wrist bones, before integrating their outputs into a composite estimate of bone age. Although segmental models exhibited a higher mean absolute difference (MAD) when used independently due to reduced contextual information, combined segmental-full-hand integration achieved the lowest MAD, 4.75 months for boys and 4.93 months for girls, outperforming predictions made by the full-hand-only model. This biologically oriented approach increases prediction accuracy and improves clinical interpretability by identifying instances of asynchronous skeletal maturation. digiBONE predictions can be used in high-volume clinical workflows, as it takes less than five seconds to compute each prediction. Our results demonstrate that integrating biologically relevant segmental information into deep learning pipelines offers a scalable, automated bone age assessment solution applicable to Indian populations.