Age Prediction from Low Resolution, Dual-Energy X-ray Images Using Convolutional Neural Networks

Abstract

Age prediction from X-rays is an interesting research topic important for clinical applications such as biological maturity assessment. It is also useful in many other practical applications, including sports or forensic investigations for age verification purposes. Research on these issues is usually carried out using high-resolution X-ray scans of parts of the body, such as images of the hands or images of the chest. In this study, we used low-resolution, dual-energy, full-body X-ray absorptiometry images to train deep learning models to predict age. In particular, we proposed a preprocessing framework and adapted many partially pretrained convolutional neural network (CNN) models to predict the age of children and young adults. We used a new dataset of 910 multispectral images that were weakly annotated by specialists. The experimental results showed that the proposed preprocessing techniques and the adapted approach to the CNN model achieved a discrepancy between chronological age and predicted age of around 15.56 months for low-resolution whole-body X-rays. Furthermore, we found that the main factor that influenced age prediction scores was spatial features, not multispectral features.