An attention-based residual neural network for deformable image registration

Abstract

Deformable image registration is a basic image processing task, especially widely used in medical image processing and analysis. Different from rigid registration, its purpose is to find the optimal nonlinear transformation between two images and establish corresponding relationship, so as to achieve image consistency. In recent years, deformable registration methods based on deep learning have been studied a lot. Compared with traditional methods, they show great advantages in registration performance. This paper proposes an attention-based residual neural network for deformable image registration, which utilizes the U-Net encoder-decoder structure to design a convolutional neural network to predict the deformation field, and uses the residual module and attention mechanism enhances the ability of the model to extract features, and finally uses the spatial transformation function to obtain the registered image, and the entire network is trained in an unsupervised manner. We conducted experiments on the MNIST dataset and 2D brain magnetic resonance images (MRI) respectively. The experimental results show that the deformable registration network proposed in this paper has good performance and shows good results in registration accuracy.