3D medical image registration based on simplified Transformer block and multi-scale iterative structure

The registration of large deformation medical images is a challenging task, due to some conditions during image acquisition, the image will produce large deformation, such as myocardial dilatation. Most of the existing registration models adopt multi-layer pyramid structure, which leads to the loss of feature information when combined at different scales. We propose a multi-scale iterative structural model of Transformer-CNN, which has the advantages of Transformer to obtain remote information while simplifying modules and reducing the number of model parameters. The multi-scale iterative structure of CNN can obtain local information and connect feature information at different scales. Experimental results on medical image datasets show that the proposed method has excellent performance in terms of registration accuracy, which is superior to the existing registration methods, while reducing the number of model parameters to a certain extent.