Hierarchical transfer learning and region of interest standardization for bone age assessment in Thai children

Abstract \normalfont\unboldmath Bone age assessment (BAA) from hand radiographs is routinely used to evaluate child growth and endocrine disorders, but manual atlas-based reading is slow and reader-dependent, while deep learning models trained on large international datasets may lose accuracy in other populations. We developed a population-specific AI framework for BAA in Thai children. It combines ROI standardization using contrast enhancement, automatic hand detection, and segmentation with an EfficientNetB7 regression model trained through hierarchical transfer learning. The model first learned skeletal-maturity representations from the public RSNA and DHA datasets and was then fine-tuned on Thai hospital data. On a held-out Thai test set excluded from BAA training and validation, hierarchical adaptation achieved a mean absolute error (MAE) of 4.64 months (RMSE, 5.85 months; \((R^2=0.98)\)), compared with 5.73 months for pooled training and 7.60 months for the public-only baseline. The full preprocessing pipeline also reduced downstream MAE relative to raw radiographs. It met the predefined detection-and-segmentation processing criterion for all evaluated RSNA and DHA images; however, this finding represents processing completeness rather than independent quantitative validation of segmentation accuracy. Among GP, TW, and a hybrid strategy using TW when applicable and GP outside the validated TW3-RUS range, the hybrid produced the lowest point-estimate MAE, although its advantage over TW was small. Grad-CAM provided exploratory evidence of attention to anatomically relevant regions but also identified localized or non-skeletal attention. Improved RSNA hand ROI annotations are released to support future research.