Background. Bone age (BA) assessment is a fundamental tool in pediatrics, pediatric endocrinology, and orthopedics for diagnosing growth and developmental disorders. Traditional manual assessment methods (the Greulich-Pyle atlas, Tanner-Whitehouse method) are characterized by high subjectivity, low reproducibility, and significant time consumption for the radiologist. In the Russian Federation, there are no domestic validated software solutions for automating this process, making the development of artificial intelligence (AI)-based systems highly relevant.
Objective. To develop and validate a neural network model for automated BA assessment from hand radiographs, adapted to the Russian population, and to create applied software for integration into clinical practice.
Materials and Methods. A combined dataset was used for training and testing the model, including 12,611 radiographs from the open RSNA Bone Age Challenge dataset and 200 verified studies from the Unified Radiological Information Service (URIS) of the EMIAS of Moscow. The data processing pipeline included hand detection (YOLOv11), segmentation (U-Net), and regression analysis (ResNet50).
Results. The Mean Absolute Error (MAE) of the developed model on the test set was 7.7 months for boys and 8.2 months for girls. The analysis time for a single image does not exceed 2 seconds. The model's accuracy surpasses that of traditional assessment using the Greulich-Pyle atlas (error up to 18 months) and is comparable to expert assessment (12 months).
«MosMedSoft» software has been developed (Certificate of State Registration No. 202566158).
Conclusion. The first domestic automated BA assessment system has been created, demonstrating high accuracy and processing speed. Implementation of the system into clinical practice will allow for the standardization of the diagnostic process, reduce the workload on radiologists, and improve the quality of diagnostics for endocrine and orthopedic pathologies in children.
