Using Convolutional Neural Networks and Transfer Learning for Bone Age Classification

The bone age of a child indicates the skeletal and biological maturity of an individual. The most commonly applied clinical methods for Bone Age Assessment (BAA) are based on the visual examination of ossification of individual bones in radiographs of the left hand and the wrist by comparing with standard hand atlas. This kind of method is highly subjective and the performance extremely depends on practitioners' experiences. This paper investigates the use of Deep Convolutional Neural Networks (DCNNs) for the automatic bone age assessment. As there exists no large-scale annotated medical image dataset comparable to ImageNet for medical image analysis, this paper uses transfer learning within DCNNs to perform bone age classifications making full use of advantages of DCNNs. We define various Regions of Interest (ROIs) based on domain knowledge, for each of which a local bone age classification model is achieved by fine-tuning the pre-trained VGGNet with corresponding ROI patches. A final bone age classification is obtained by fusing multiple regional models. The results show that the proposed approach outperforms the current state-of-the-art classification methods in BAA with small dataset.