Abstract
BACKGROUND: Bone-age assessment is a crucial tool in pediatric healthcare for monitoring children's growth and diagnosing endocrine disorders. Traditional manual assessment methods such as Greulich-Pyle (GP) and Tanner-Whitehouse (TW) standards have significant limitations including high subjectivity, complex operation, and a time-consuming process. The currently available automated methods struggle to effectively highlight clinically relevant growth regions and capture anatomical associations between skeletal structures across different developmental stages. This study developed a novel framework integrating anatomical knowledge with developmental stage awareness to improve automated bone-age assessment accuracy and interpretability. METHODS: We evaluated our proposed framework using the publicly available Radiological Society of North America (RSNA) pediatric bone age dataset. The model was trained and tested using the standard data split, with performance primarily assessed by the mean absolute error (MAE) in months. We designed a developmental stage-aware graph transformer framework (DSGTF) for automated bone-age assessment. Our framework integrates image preprocessing with automatic key region detection, along with subsequent feature extraction from both local skeletal regions and the whole-hand area. The core innovation lies in our graph transformer architecture that models anatomical relationships with a biologically informed skeletal graph structure. This is enhanced by our developmental stage-aware visual transformer fusion module, which adaptively identifies each sample's developmental stage and dynamically adjusts processing strategies to accommodate variations in bone characteristics across maturity levels. The system is optimized through multitask learning, which balances bone age prediction with anatomically meaningful feature representation. RESULTS: The DSGTF model achieved an MAE of 4.82 months in bone-age assessment. The model exhibited consistent performance across different age groups and sexes, with the MAE ranging from 3.39 to 6.03 months. Robustness evaluation showed that the model maintained stable performance under various image transformations, with 11 out of 16 transformations showing no statistically significant performance degradation (P>0.05). Geometric transformations resulted in minimal increases in MAE (<0.43 months), while moderate photometric changes produced MAE differences within 0.21 months. Visualization analysis revealed that the developmental stage-aware module successfully captures differentiated attention patterns across skeletal regions, with early stage processing focusing primarily on wrist regions, the middle stage showing balanced attention across structures, and the late stage emphasizing the proximal phalanges while maintaining high attention to radius-ulna regions-a pattern that demonstrates the model's ability to capture clinically relevant developmental priorities across different maturation stages. CONCLUSIONS: By leveraging a skeletal graph structure based on anatomical relationships and incorporating a developmental stage-aware processing mechanism, the DSGTF framework provides an accurate and efficient automated solution for clinical bone-age assessment. The proposed approach exhibits strong stability and consistency across different age groups and sexes, making it a reliable tool for real-world medical applications. This model effectively simulates the clinical assessment process by dynamically adjusting attention to different skeletal regions based on developmental stages, mirroring the decision-making process of radiologists.