Deep Learning-Based Bone Age Assessment Using Multi-Site MRI Epiphyseal Imaging: A Retrospective Study

This study presents BA-Net, a deep learning framework designed to automate juvenile bone age assessment by analyzing epiphyseal union patterns across wrist and knee MRI scans. Leveraging 474 axial wrist, 1,149 coronal knee, and 105 combined wrist-knee T1-weighted MRI scans collected between 2015 and 2021, the model integrates convolutional neural networks to extract multidimensional skeletal features and sex-specific parameters for multi-label classification. Data were partitioned into training, validation, and test sets to evaluate performance. Comparative analysis involved 1,728 subjects (mean age 14.81±1.62 years; 1,013 males), with results benchmarked against manual assessments by three specialists and chronological age. Single-site evaluations demonstrated over 90% accuracy for 12-, 14-, and 16-year thresholds (wrist marginally superior to knee), while multi-site fusion improved accuracy beyond 95%, statistically outperforming isolated anatomical assessments. Although manual evaluations achieved perfect accuracy for 16-year knee classifications—aligning with BA-Net—they lagged behind automated methods in younger age groups and alternative anatomical regions. The proposed system eliminates radiation exposure while maintaining clinical-grade precision, demonstrating that coordinated analysis of wrist and knee MRI data significantly enhances bone age prediction robustness. These findings highlight the potential of multi-site deep learning integration to standardize and streamline pediatric skeletal maturity evaluations in clinical practice.