Non-rigid medical image registration using multi-scale residual deep fully convolutional networks

Abstract

Medical image registration technology is important in medical image processing, which can establish correspondence between complex medical images and ensure the comparability of image data. Since the deep convolutional neural network model is prone to degradation and the convolution kernel has limited detection range during training, a Multi-scale residual full convolutional network (MS-ResFCN) model is proposed for unsupervised non-rigid medical image registration tasks. The model introduces a residual structure in the Fully Convolutional Network (FCN) model to ensure effective and stable training. At the same time, a hierarchical multi-scale convolution kernel is constructed within a single convolutional layer of the residual structure, which enhances the nonlinear mapping ability of the network. The local texture information and the contextual information of the image are synchronously extracted and fused to enrich the diversity of features. The experiment results on LPBA40 data sets show that the MS-ResFCN model can effectively eliminate the degradation phenomenon of the deep network during the training process and extract multi-scale features, which achieves better feature representation ability and registration accuracy.