Pediatric bone age assessment with AI models based on modified Tanner-Whitehouse

Abstract

Assessment of bone age, which represents the development and maturity of bones. It helps treat various pediatric conditions and address legal issues. Conventional bone age assessment is a complex and laborious procedure that is susceptible to inconsistencies between different reviewers and within the same reviewer. Artificial intelligence is a new automatic, accurate, and fast method used to evaluate bone age from X-ray images. In this work, a new network design based on AI methods is proposed. This network is based on the dataset obtained from the Paediatric Bone Age Challenge organized by the Radiological Society of North America. The collection comprises 12,600 radiological pictures of left hands, each labeled with the patient’s bone age and sex. The design involves two steps: first, using a Faster R-CNN mask by training the ResNet50 model to select regions of interest and then entering the selected regions into three models (Inception v3, GoogleNet, and ResNet50) for regression-based bone age estimation. The results from these models vary based on their internal structure. ResNet50 yields a mean absolute error (MAE) of 10.6 for males and 9.5 for females. Inception v3 has an MAE of 7.5 and 8.3 for Males and females, and GoogleNet has an MAE of 8.4 for males and 9.2 for females. These models can enhance the precision and effectiveness of bone age prediction.