Research On Human Brain Multimodal Medical Image Registration Technology Based On Deep Learning

Medical image analysis plays an important role in disease diagnosis and research,and in which,image registration is the most important preprocessing method.At present,traditional medical image registration technology has already well developed,it is timeconsuming,and its accuracy is not better enough to meet the requirements in many clinical applications.Considering the superiority of deep learning in the field of Computer vision,this work intends to design deep learning models for multimodal image registration.The main contents are as follows:(1)For the registration between structural images,we proposed an ensemble attention-based and dual similarity guidance registration network named as EADSGReg Net,which includes three parts: feature extraction,deformation field estimation,and a resampler.A cascaded encoder and a decoder are designed to realize multi-scale feature extraction and deformation field estimation.The ensemble attention module(EAM)is introduced into the cascaded encoder to learn the importance of the extracted features through training,so as to select features that are more useful for the registration task,making the decoder estimate the deformation field more accurately.In order to be able to estimate the global and local deformations accurately,the global gray-scale similarity NMI(normalized mutual information,NMI)and the local feature similarity based on the SSC(self-similarity context)descriptor are used guiding the training of the network.The results show that the proposed registration model can effectively improve the speed and accuracy of image registration.(2)For the registration of functional and structural images.Since there is nonlinear relationship between the gray levels,and difference in textures and contents,how to learn the correct correspondence from the features with larger differences,extracted by deep convolutional neural network,is difficult.For these reasons,we proposed residual fusion registration network named as RF-Reg Net,which is like to Laplace pyramid,to learn the residual information,and designed an independent residual encoder to learn the deformation parameters of the local details in the image pair.The results show that the proposed method can effectively retain local detailed information and achieve a better registration effect.