Research On Brain MRI Image Registration Method Based On Joint Learning

In the clinical field of medicine,doctors need to compare anatomical images and functional images,and compare medical images at different time periods to judge the condition.The image matching criteria play a key role in the doctor’s judgment of the condition.Accompanying it is the development of traditional methods and deep learning methods in medical image registration.At this stage,the main research is focused on the field of deep learning.The unsupervised registration accuracy based on Convolutional Neural Network(CNN)There is still room for improvement.In addition,although the accuracy of supervised network learning is higher,its labeled data(deformation field)is more difficult to obtain,so it cannot be realized in practical applications.At present,the neural network model based on weak supervision has gradually become the focus of scholars’ research.In view of the above situation,a method based on CNN and joint learning is proposed,which combines the two tasks of image segmentation and image registration to apply to the work of medical image registration,which significantly improves the registration effect.Based on the semantic information of the segmentation results in the medical image,the network model will learn the hypothesis of the edge information of the image,and the structure of the twin network is introduced,so that the parameter optimization of multi-task learning is concentrated in the same shared layer,enriching the semantic information of the network model Feature learning.In addition,because the nature of the diffeomorphism is easily destroyed during the registration process,in order to maintain the nature of the diffeomorphism in the deformation process,a reciprocal training method is proposed to introduce the mutual interaction between the floating image and the fixed image in the image registration process.Inverse deformation,in order to ensure the reduction of folding phenomenon in the deformation process,and restrict the one-to-one mapping relationship of the deformation field.At the same time,local regularization constraints are imposed on the deformation field to replace the global regularization,so that the loss function focuses on the folded part instead of all deformation positions.In order to verify the effect of the model,the ADNI public data set is obtained and the software is used to obtain segmentation semantic information.The experimental results show that the model method in this article is compared with the traditional method(Elastix)and other deep learning methods(Voxel Morph,FAIM,etc.))The accuracy has been improved by nearly 10%,and the folding phenomenon(diffeomorphism)is more reasonable.At the same time,the time cost is also improved to a certain extent(30%).In addition,through ablation experiments,the results prove that the introduction of segmentation semantic information promotes the superiority of the model in feature extraction,and the reciprocal consistency loss and local regularization have obvious constraints on the deformation field.The results of visualization can more intuitively verify the effect of the model proposed in this article.