Evaluation of Skeletal Gender and Maturity for Hand Radiographs using Deep Convolutional Neural Networks

Abstract

Assessment of skeletal maturity is typical strategy applied in clinical pediatrics today. The main goal of a Bone Age Assessment (BAA) is to determine endocrinology and growth disorders by comparing the bone and chronological age of the patient. Several methods are developed to determine skeletal maturity, but Greulich-Pyle and Tanner-Whitehouse represent the two most common methods that involve left hand and wrist radiographs. However, these methods are extremely time-dependent and rely on an experienced radiologist, who further evaluates bone age using hand atlas as a reference. In this paper, VGG-16 and ResNet50 are two Deep Convolutional Neural Network (DCNN) models applied with ImageNet pre-trained weights in order to estimate correct bone age and achieve high accuracy of gender prediction using public RSNA dataset that includes 12611 radiographs. The experimental results show month discrepancy of approximately eight months and 82% accuracy during the process of gender classification.