Bone Age Assessment Based on Deep Convolution Neural Network

Abstract

Bone age assessment has become a common clinical practice in the diagnosis of endocrine disorders and bone maturity. An end-to-end bone age assessment method based on deep convolutional neural networks is proposed, which uses deep learning to segment the hand region and predict key points for correcting hand posture. By optimizing the dice loss function in DenseUNet for segmentation, the model is easier to converge and has better segmentation quality. Squeeze-and-excitation (SE) module and Position/Channel Attention (PA/CA) module can help network to focus on useful feature channels and position. By merging InceptionV3 with SE-module and PA/CA module, the assesssment model can extract the features of hand bone more efficiently. The experimental results show that the proposed network is superior to traditional and other automated bone age assessment methods both in accuracy and efficiency on the data of the 2017 children's bone age challenge organized by the Radiological Society of North America. Finally, the visualization of the region activation for bone age assessment is implemented, which is helpful to understand the connection between the image appearance and the bone age.