Research On Medical Image Registration Algorithm Based On Deep Learning

Medical image registration can establish the spatial consistency of corresponding anatomical structures between different medical images,which is of great significance in medical image analysis and medical diagnosis.Excellent registration algorithms can achieve full registration and information fusion of images,enrich the information in images,and help doctors diagnose and treat diseases in time.The traditional medical image registration algorithm is to establish an appropriate similarity measure function and obtain the best registered image through multiple iterations.However,the traditional medical image registration algorithm has the problems of long time and slow speed,which can meet the needs of clinical application.With the rapid development of deep learning methods,especially the application of convolutional neural network(CNN)in the field of image processing,researchers have begun to apply deep learning methods to the task of medical image registration.Deep learning methods can build a more efficient and accurate registration model to complete the image registration task through a large amount of data.However,there are still some problems in the registration process of deep learning methods,such as the accuracy is not high enough and the topology structure of the image cannot be guaranteed.In this paper,aiming at the problems of medical image algorithms based on deep learning,the medical image registration algorithm is studied more deeply,and a new registration model is proposed.The main contents are as follows: Medical image registration can establish the spatial consistency of corresponding anatomical structures between different medical images,which is of great significance in medical image analysis and medical diagnosis.Excellent registration algorithms can achieve full registration and information fusion of images,enrich the information in images,and help doctors diagnose and treat diseases in time.The traditional medical image registration algorithm is to establish an appropriate similarity measure function and obtain the best registered image through multiple iterations.However,the traditional medical image registration algorithm has the problems of long time and slow speed,which can meet the needs of clinical application.With the rapid development of deep learning methods,especially the application of convolutional neural network(CNN)in the field of image processing,researchers have begun to apply deep learning methods to the task of medical image registration.Deep learning methods can build a more efficient and accurate registration model to complete the image registration task through a large amount of data.However,there are still some problems in the registration process of deep learning methods,such as the accuracy is not high enough and the topology structure of the image cannot be guaranteed.In this paper,aiming at the problems of medical image algorithms based on deep learning,the medical image registration algorithm is studied more deeply,and a new registration model is proposed.The main contents are as follows:1.This paper proposes a registration network TD-Net that combines the global modeling ability of Transformer and the local modeling ability of convolutional neural network(CNN).The CNN-based method can provide rich local registration information,but its global modeling ability is weak,and it is difficult to conduct longdistance information interaction,which limits the accuracy of registration.Compared with CNN,Transformer can provide rich global information,while the local modeling ability of Transformer is not as good as CNN.In this paper,TD-Net,a hybrid registration network similar to the U-Net network,is proposed to combine Transformer and CNN to extract global and local information.The experimental results on the brain dataset show that the proposed method can effectively improve the registration accuracy.2.In this paper,a new two-stage coarse-to-fine registration network TS-Net is proposed,and a new smooth constraint function is proposed.Previous medical image registration methods based on deep learning usually fail to deal with large and complex deformations in the image well,and destroy the topological properties of the image in the process of deformation.The proposed model TS-Net learns deformation from coarse to fine in two stages,and can gradually learn large and complex deformations in the image.In the second stage,the local receptive domain can be expanded and more local information can be obtained by performing the feature maps that are downsampled by jumping concatenation in the first stage.The smooth constraint function used by previous algorithms imposes the same constraint on the whole world and is not targeted.In this paper,a new smoothness constraint function is proposed for each voxel deformation,which can better ensure the smoothness of the transformation and preserve the topological properties of the image.Experiments on complex deformed brain datasets and large deformed heart datasets show the effectiveness of the algorithm.