Research On Brain Image Registration Of Alzheimer’s Disease Based On Generative Adversarial Network

Medical image registration is a key step in image processing.In recent years,with the development of deep learning,medical image registration methods based on deep learning have also developed rapidly.Due to the limitation of image segmentation and marking,unsupervised image based the registration method has become the mainstream of image registration.In view of the superior performance of the generative adversarial network combined with other fields,the image registration method of the generative adversarial network has also received extensive attention,and it is also a branch of unsupervised image registration.Therefore,many applications of generative adversarial networks in the field of medical image registration have also been derived.At present,there are several challenges in the image registration method based on the generative adversarial network.One is that the feature loss will occur during the upsampling process of the U-Net network model,which will reduce the accuracy of image registration;the other is to use it in the training and testing phase.The cascaded network will increase the network parameters and increase the time of network training and testing;the third is that the discriminative network in the generative adversarial network is not strong in discrimination,which limits the accuracy of image registration.Aiming at the challenges in the above image registration,this paper studies the image registration method based on generative adversarial network,and pays attention to the application of image registration in Alzheimer’s disease.The main research contents are divided into the following points:(1)Aiming at the problem of feature loss during the upsampling process of the U-Net network model,this paper proposes a dual attention mechanism network structure.In the process of sampling on the U-Net network,the channel attention mechanism and the position attention mechanism are added,so that more important relevant information can be retained in the process of sampling on the U-Net network,and the network can better learn the image information between.Compared with the most classic Voxel Morph,the experimental results of this network structure on the OASIS-3 Alzheimer’s disease dataset show that the Dice values of the three brain regions of cerebrospinal fluid,cerebral white matter and gray matter have increased by 0.101,0.076 and 0.059 respectively.(2)In view of the problem that the cascaded network will increase the training and testing time,this paper proposes the enhanced network as an image registration method to enhance the regularization term,and in the enhanced network,the mean square error loss function is introduced to constrain the deformation field to tend to zero deformation field,so that the registration network learns a better deformation field,thereby improving the accuracy of image registration.This network framework is only used during training,and only the registration network is used during testing.This method uses the HBN dataset and the ABIDE dataset to verify the experimental results,and achieves comparable accuracy to Voxel Morph.(3)Aiming at the problem that the discrimination network is not strong in the generative adversarial network,this paper proposes the MATGAN method,which incorporates a multiscale attention mechanism into the discrimination network.The multi-scale attention mechanism improves the feature recognition ability by combining hole convolution.To reduce the influence of redundant features,and in order to reduce the interference of outliers in image registration,the Huber loss function is added to the training to improve the discrimination ability of the discrimination network and improve the accuracy of image registration.The MATGAN method has been tested on the ADHD200 and OASIS-1 datasets to verify the effectiveness of the method,and the average Dice index of the method has increased by 0.015.