Research On Multi-scale Registration Of Medical Images Based On Deep Learning

Image registration is the premise of medical image analysis,which can make the corresponding structure of two images consistent in spatial position.Specifically,it is to register the corresponding structure of one image to the spatial position coordinate system of another image.Medical image registration is a key technology of computer-aided medicine,which plays an important role in medical fields such as disease detection,surgical simulation and navigation,atlas-based structure segmentation,adjuvant therapy and so on.Although medical image registration technology has achieved good development in recent years,it still faces some problems in the accuracy and quality of registration.From the perspective of data,the image registration method based on deep learning requires a large amount of real annotation data for training,but the open annotation data is less,and the data annotation process is time-consuming and laborious.From the perspective of registration strategy,the single scale registration strategy adopted by current registration methods is difficult to capture complex deformation in medical images.From the perspective of network structure,the structure of simple stacked convolution blocks is difficult to extract discriminant features that are more discriminative for structural displacement information in image.All these problems affect the results of medical image registration.In order to improve the quality and accuracy of medical image registration,aiming at the low quality of medical image registration caused by the lack of real annotation data and single-scale registration strategy,this thesis proposes an unsupervised multi-scale medical image registration method based on deep learning;Aiming at the problem that the existing registration methods do not fully extract the features with more discriminative for structural displacement information in image,this thesis proposes an unsupervised medical image registration method based on feature representation match volume.The main research contents of this thesis are as follows:(1)An unsupervised medical image registration method based on deep learning is proposed,which adopts multi-scale registration strategy.This method extracts the multiple scales feature from image,and predicts the registration field on each scale,and warps the image to be registered into the registered image by combining the spatial transformation module.Then,the he weighted sum of difference values between the registered image and the target image at all scales is used as the loss value training network to realize the gradual refinement of the registered field and the registered image from coarse to fine.This method was analyzed and evaluated on cardiac MRI dataset.Experimental results show that the proposed multi-scale registration strategy enhances the learning effect of unsupervised image registration network and improves the registration effect.(2)An unsupervised medical image registration method based on feature representation match volume is proposed.Firstly,match volume structure is added on the basis of the previous registration method.The match volume stores the matching costs associated with the corresponding points on the two images,and the discriminant features that are more discriminant to the displacement information are extracted by constructing match volume.Secondly,the separable four-dimensional convolution is used to process the match volume,and the displacement information contained in the match volume is fully utilized to improve the registration effect.The proposed method is evaluated on the combined cardiac MRI data set.Experimental results show that the multi-scale unsupervised image registration network based on match volume can learn the displacement differences of corresponding anatomical structures between images,and achieve high precision and high reliability image registration.