Researchon Symmetric Registration Algorithm Of Multimodal Medical Image Based On Deep Learning

Medical image registration is a fundamental problem in the field of medical image analysis.The main task of medical image registration is to associate the corresponding pixels of images obtained from different devices or different times,different depths,and different angles of view and then convert them into the same coordinate system.Medical image registration technology plays an important role in medical image analysis.More and more medical image registration algorithms are integrated into medical imaging equipment.Advanced medical image registration algorithms are integrated into medical image engineering equipment,which further improves the intelligent application level of image engineering equipment.At present,medical image registration technology is widely used in computer-aided diagnosis,computer-aided therapy,surgical navigation,radiotherapy planning,and examination of treatment effects.The research on medical image registration has vital clinical value and practical significance for improving the intelligence of medical imaging equipment.Researchers around the world have done a lot of work over the past few years in medical image registration and achieved remarkable results.Early image registration methods are mainly divided into feature point matching and intensity matching.After years of research and improvement,some scientific research achievements have been successfully applied to clinical practice.However,the early registration methods usually obtain the spatial correspondence between images by iterative optimization,which has high computational complexity and is difficult to be used in real-time surgical navigation.Medical images also have the characteristics of complex structure and difficult acquisition,resulting in the following problems to be solved in the field of medical image registration:(1)The most of the existing registration technologies often directly predict the unidirectional deformation field,which can not guarantee the existence of inverse transformation,and large errors will occur in the areas with large structural differences in the unidirectional deformation process;(2)The similarity between multimodal images is difficult to explore,which affects the accuracy of multimodal image registration;(3)It is very difficult to obtain labeled medical image data,which hinders the research of new registration algorithm.These are urgent problems to be solved in the research of medical image registration.In recent years,the development of computer technology and the emergence of deep learning have brought significant changes to the fields of computer vision and medical image analysis.The registration problems based on the deep learning methods are defined as functions with parameters,and the convolutional neural networks are used to model the processes.The images in the data set used to train the model form a training set in pairs to optimize the parameters of the network,that is,the weight of the convolutional kernel in the network.Given a pair of new 2D or 3D input images,the trained model can directly predict the spatial mapping relationship between all voxels from subject image to template image.The algorithms based on deep learning show great potentials in the research of medical image registration,which provide new research ideas and methods for medical image registration,which have a very broad application prospect in the application of medical image engineering,and which are of great significance to improve the efficiency and intelligence of medical imaging equipment.This dissertation takes the clinical application of medical image analysis as the background,and takes medical images as the research object,and takes image registration based on deep learning as the research goal,and carries out a series of research work around the symmetric registration of single-mode/multimodal medical image,discussing how to use deep learning algorithms to improve the performance of medical image registration and solve the problem of medical image registration.This dissertation focuses on the problems of symmetric image registration in medical image registration,the difficulty of similarity measurement in the process of multimodal medical image registration,and the difficulty of obtaining labeled data in the research process of the medical image registration algorithm.The contributions of this dissertation mainly include the following aspects:(1)Aiming at the problem of symmetric registration of single-mode medical images,a symmetric registration network of single-mode medical images based on deep convolutional neural networks is proposed,which solves the problem of symmetric registration of single-mode medical images and enhances the accuracy and robustness of image registration with large differences in anatomical structures.A symmetric registration network of single-mode medical images based on deep convolutional neural networks is proposed.Firstly,a new end-to-end image symmetric registration network is built to predict the distortion of the target image and the template image into the two deformation fields in the pseudo central template space;Secondly,image pair data training is used to optimize the parameters of image symmetric registration network model with symmetry strategy.Finally,the test set of images are used to compare the proposed algorithm with the algorithms presented by others.Through experiments on brain MRI images,the experimental results show that the smoothness of the two deformation fields in the symmetrical direction predicted by the proposed single-mode medical image symmetrical registration algorithm based on deep convolutional neural network has also been significantly improved,and the registration in the image areas with large anatomical structure differences has better accuracy and robustness.(2)Aiming at the problem of single-mode/multimodal medical image registration,a singlemode/multimodal medical image symmetric registration network algorithm based on generative adversarial networks is proposed.The algorithm solves the problem of single-mode/multimodal medical image symmetric registration and uses semi-supervised learning strategies to train and optimize the proposed network model using unlabeled data.A single-mode/multimodal medical image symmetric registration network algorithm based on generative adversarial networks is proposed.Firstly,in the process of building the network algorithm,the symmetric registration problem of single-mode/multimodal images in medical images is modeled as a conditional generative adversarial network model;Secondly,in the training process,semi-supervised learning strategies are used to train the algorithm model.Semi-supervised learning strategies can make full use of valuable labeled data and many unlabeled data to train the algorithm model;Finally,the symmetrical loss function is introduced into the algorithm model to stimulate the cyclic network composed of geometric deformation and inverse deformation from one image to another.The medical images are restored to the original images after two transformations.Through experiments on four single-mode medical image data sets and three multi-mode medical image data sets,the comparative experimental results show that the medical image symmetric registration algorithm based on generative adversarial networks proposed in this chapter has advantages over other existing registration methods.(3)Aiming at the problem of multispectral fundus image registration,multimodal medical image adversarial segmentation and registration algorithm based on end-to-end deep learning model is proposed,which solves the problems of registration between multispectral fundus image sequences and fundus vascular segmentation.A multimodal medical image adversarial segmentation and registration network based on deep learning algorithms is proposed.Through the end-to-end adversarial learning process,the distribution of fundus blood vessels and multispectral fundus image registration can be predicted simultaneously.The whole network is divided into two parts: the segment-driven registration module and the segmentation module.Based on the segmentation-driven registration network,the semi-supervised adversarial learning strategies are used to train the segmentation network and registration network at the same time.In the process of registration of multispectral image pairs,the fundus vessels on the image are segmented simultaneously.The experimental results show that the proposed adversarial segmentation and registration networks for multispectral fundus images can meet the ideal accuracy requirements.