Gaussianmorph: deformable medical image registration with Gaussian noise constraints.

Abstract

Deep learning-based image registration methods offer advantages of time efficiency and registration outcomes by automatically extracting enough image features. Currently, more and more scholars choose to use cascaded networks to achieve coarse-to-fine registration. Although cascaded networks take a lot of time in the training and inference stages, they can improve registration performance. In this study, we utilize the advantage of high registration performance of cascaded networks. Two VoxelMorph convolutional neural networks are cascaded together. The first VoxelMorph network outputs the dense deformation field of registration. The second network outputs a noisy deformation field, which serves to boost the registration performance by minimizing the error in comparison with Gaussian noise. At the same time, the Enhancement Features-encoder (EF-encoder) block is introduced in the encoder and decoder part of the network to achieve enhancement features functions by attention mechanism. This paper conducted experiments on LPBA40 and HBN datasets. The experimental results show that the Dice similarity coefficient, Average Symmetric Surface Distance, Structural similarity and Pearson correlation coefficient of GaussianMorph are better than those of VM, VM × 2 and TST-Net. Experimental results show that GaussianMorph can improve the registration accuracy.