Age-based morphological prediction for automated bone age assessment: a deep learning approach

This study presents a novel approach to skeletal bone age assessment, which is a critical clinical practice for evaluating various disorders, including endocrinology, genetics, and growth issues in adolescents. The regions of interest within hand bones change with age, making these variations essential for clinicians who rely on the Greulich and Pyle (G&P) atlas for accurate diagnoses. Our proposed framework diverges from traditional deep learning models by introducing an age-based morphological strategy (AMPS) that automatically analyzes hand bone characteristics across different age groups. The assessment process involves two main stages: first, we utilize U-Net for segmenting the entire hand bone and then crop the relevant regions after standardizing the images' tone and contrast. In the second stage, AMPS employs Xception blocks to facilitate automatic bone age assessment. To tackle the issue of data imbalance among various age groups, we implement a data-balance (DB) loss function. Our model's effectiveness was validated using the RSNA dataset and a local hospital dataset (L-HOSP), achieving mean absolute errors (MAE) of 5.78 months (0.48 years) and 6.95 months (0.57 years), respectively. These results demonstrate the model's robust generalization capabilities. The findings emphasize the potential of AMPS in enhancing automated bone age assessment protocols, with future research aimed at extending this methodology to other regression challenges in the field of medical imaging.