DisCLE-BAR: A Dynamic Region-of-Interest Guided Framework for Hand Bone Age Assessment

Accurate bone age assessment (BAA) is crucial for evaluating pediatric health, predicting growth, and supporting legal and athletic tasks. While deep learning-based BAA methods have improved precision and efficiency, many produce features misaligned with bone growth stages, leading to suboptimal performance and limited interpretability. To address this, we propose DisCLE-BAR, a distillation and contrastive learning-enhanced framework for bone age regression. DisCLE-BAR dynamically identifies key regions of interest (ROIs) in hand X-ray images, adapting to specific bone growth stages. The training of DisCLE-BAR involves two phases: a full ROIs capturing phase, where knowledge distillation identifies regions critical for BAA, and an adaptive ROI weight adjustment phase, where weighted supervised contrastive learning refines attention maps. This design focuses the model on the most significant ROIs while minimizing the influence of less relevant areas. Experiments on the RSNA dataset show that DisCLE-BAR achieves a mean absolute error (MAE) of 3.73 months, outperforming other state-of-the-art methods by effectively capturing dynamic bone development characteristics. The results demonstrate that DisCLE-BAR offers a reliable, interpretable solution for BAA with strong clinical potential.