Study On The Registration Of Standard Brain Atlas And MRI Images Of Human Brain

Brain atlases are important research tools for exploring the structural and functional information of the human brain,which accurately locate some important functional areas of the human brain.Research on the registration of brain atlas and MRI images of human brain has important significance for brain tumor surgery and neurosurgery.Precise registration of the two can ensure that important brain tissue structural areas are accurately mapped from the brain atlas to the human brain images during the operation,thereby assisting the doctor in planning the operation.Currently,traditional non-linear registration methods are mostly used for the registration of brain atlas and human brain images,which have slow registration speed and poor inverse consistency.When there is huge gray scale or deformation difference between the images to be registered,the deformation field tends to fold at some local voxels,which will destroy the topological structures of the deformation field and result in the reduction of inverse consistency of the deformation field.In view of the above problems,a non-linear registration method based on unsupervised learning is proposed to register brain atlas and human brain images.This method effectively solves the problems of slow registration speed and poor inverse consistency,and uses the Jacobian loss function to punish the folding voxels,thereby improving the topology-preserving property and inverse consistency of the deformation field.The main work and innovative results of this article are as follows:1.The spatial and intensity regularization of the images to be registered.The former mainly changes the spatial positions of the voxels in the images without changing the voxel values.While the latter is the opposite,which changes the voxel values without changing the spatial positions of the voxels.Among them,the spatial regularization mainly includes the unification of image size and resolution,acquisition of registration marks,rigid registration,affine registration and image cropping,etc.Intensity regularization is mainly divided into intensity value normalization and histogram matching.The existing histogram matching algorithm is improved in this paper.By adding a threshold,the algorithm reduces the grayscale difference of the images to be registered and ensures that the voxel values of the background areas remain unchanged at the same time,so as not to interfere with the registration of the regions of interest.2.The construction of an unsupervised nonlinear registration algorithm.The algorithm consists of an encoder-decoder network module,a velocity field integration module and a grid sampling module.The encoder-decoder network module uses the inception module to extract and fuse input image features of different scales and uses transposed convolution to learn the parameters of the upsampling process.The output of the network is the stationary velocity field,which is integrated by the scaling and squaring method to obtain the registration deformation field.The grid sampling module uses the deformation field to transform the source image to obtain the registered image.3.The training and testing of the unsupervised registration algorithm.During the training process,the mean square loss is computed for the registered image and the target image,and the spatial gradient loss and the Jacobian loss are calculated for the velocity field.The total loss of the weighted sum of the three loss functions is backpropagated to update the registration network parameters.During the test,the registration algorithm of this paper is compared with two algorithms with good registration performance on several registration performance metrics.The results show that the algorithm of this paper achieves the best registration performance compromise.4.The performance analysis of key parts of the unsupervised registration algorithm.The novel encoder-decoder network can significantly improve the inverse consistency,but it will lead to the increase of folding voxels.Therefore,this paper designs the Jacobian loss function to reduce the folding voxels and to further improve the inverse consistency.In addition,small increase in the depth and width of the registration network cannot significantly improve the registration performance.The proportion of folding voxels,the inverse consistency error,and the registration accuracy increase first and then remain unchanged as the integration step of the SS method increases.The registration time gradually increases with the growth of the integration step.