Research On Medical Image Registration Based On Attention Mechanism

In medical image analysis,image registration is a very important and fundamental step.Deformable image registration is the most complex and widely used type of image registration.The purpose of deformable image registration is to transform images from different coordinate systems to a common coordinate system,and to match their contents by finding dense nonlinear spatial correspondences between image pairs.It is the basis of many clinical tasks and has been widely used in medical image preprocessing and diagnosis,such as tumor growth detection,surgical navigation,and atlas creation.However,there are still many issues to be explored in deformable image registration.For example,in feature-based registration methods,how to introduce advanced deep learning general visual models into registration tasks,how to effectively combine the classic matching algorithm with deep learning models to establish more accurate and automated correspondences between regions of two images,and how to further improve the accuracy and speed of registration models are all worth exploring by researchers.Against the above backdrop,this article presents research work on deformable image registration algorithms.Based on the Transformer model and self-attention mechanism,two endto-end monomodal deformable image registration algorithms are proposed.The main contributions are as follows:(1)A Transformer-based unsupervised single-modality deformable image registration network is proposed.This network introduces the visual Transformer feature extraction module into the image registration network based on convolutional neural networks.The Transformer feature extraction module is used to replace the convolutional module to extract deep features of the image,leveraging the advantage of the Transformer model in extracting features with longrange dependencies,and compensating for the poor long-range feature dependency processing ability of the convolutional module.In this paper,the structure of the progressive image pyramid is also introduced into the proposed network to enhance the model’s ability to capture features at different scales and improve the registration results.The network is evaluated and analyzed on one brain MRI dataset and one lung CT dataset.The experiments show that the proposed network exhibits good registration results in subjective qualitative evaluation,and the DSC,HD and other indicators also have different degrees of improvement in objective quantitative evaluation.(2)A deformable image registration network based on self-attention mechanism and feature explicit matching is proposed.This network combines the feature matching idea in traditional iterative registration algorithms with deep learning implementation.It independently extracts features from individual images of the registration image pairs,and further establishes explicit multi-level feature matching between image pairs under the implementation of deep learning through the "Query-Key" matching mechanism of self-attention mechanism,improving the interpretability of the network.The network is evaluated and analyzed on three brain MRI datasets and one lung CT dataset,and the experimental results prove that the proposed network can achieve high-precision image registration.