An Unsupervised Dual Attention Method for 3D Medical Image Registration

Abstract

Medical image registration has always been a research hotspot in medical image processing. Compared with other fields (such as image segmentation), the training parameter space is very large and often lacks expert supervised correspondence annotations. These two challenges lead to slow image registration progress. We present a novel unsupervised medical image registration model, which incorporates channel and spatial weight module and spatial structure attention module. Firstly, we weigh the channels and spaces in the convolutional network process, which is an improvement to the Convolution structure, it can output the attention map in turn from the input feature map along the two independent dimensions of the channel and space, and then multiply the attention map by the input feature map for adaptive feature refinement. Then, we focus on the 3D characteristics of medical images, the proposed spatial structure attention can aggregate the spatial information of medical images to avoid the loss of important information during the convolution process, which is more conducive to retaining the content information of the input images. In addition, we use a learnable activation function in the network to improve the experimental results without adding more parameters. Combine the above parts with a convolution neural network to form our network. Finally, on LPBA40 and OASIS datasets, the method proposed by this work outperformed most work.