Lightweight deep learning system for automated bone age assessment in Chinese children: enhancing clinical efficiency and diagnostic accuracy

Bone age assessment (BAA) is a critical diagnostic tool for evaluating skeletal maturity and monitoring growth disorders. Traditional clinical methods, however, are highly subjective, time-consuming, and reliant on clinician expertise, leading to inefficiencies and variability in accuracy. To address these limitations, this study introduces a novel lightweight two-stage deep learning framework based on the Chinese 05 BAA standard. In the first stage, the YOLOv8 algorithm precisely localizes 13 key epiphyses in hand radiographs, achieving a mean Average Precision (mAP) of 99.5% at Intersection over Union (IoU) = 0.5 and 94.0% within IoU 0.5–0.95, demonstrating robust detection performance. The second stage employs a modified EfficientNetB3 architecture for fine-grained epiphyseal grade classification, enhanced by the Rectified Adam (RAdam) optimizer and a composite loss function combining center loss and weighted cross-entropy to mitigate class imbalance. The model attains an average accuracy of 80.3% on the training set and 81.5% on the test set, with a total parameter count of 15.8 million—56–86% fewer than comparable models (e.g., ResNet50, InceptionV3). This lightweight design reduces computational complexity, enabling faster inference while maintaining diagnostic precision. This framework holds transformative potential for pediatric endocrinology and orthopedics by standardizing BAA, improving diagnostic equity, and optimizing resource use. Success hinges on addressing technical, ethical, and adoption challenges through collaborative efforts among developers, clinicians, and regulators. Future directions might include multimodal AI integrating clinical data (e.g., height, genetics) for holistic growth assessments.