Hippocampus Registration Based On 3D-convolution Neural Network

Hippocampus is an important part of the limbic nervous system,which is mainly responsible for human memory,learning,and spatial positioning.Many neurological diseases,such as temporal lobe epilepsy,Alzheimer's disease,schizophrenia,and depression,are associated with changes in the shape,volume and function of the hippocampus.Magnetic resonance imaging(MRI)can provide three-dimensional brain tissue information with rich contrast and high resolution.It is essential data for the study of hippocampal morphology.Image registration is very important for hippocampal analysis.Usually,the captured hippocampal images are registered into the standard template space,and then we can realize the statistics of hippocampal morphology,gray level,and other information.However,the traditional image registration methods have the problems of a large amount of computation and time-consuming.In recent years,with the development of deep learning and the emergence of the end-to-end network,it provides a new way to the above problems of image registration.For hippocampus registration,this paper proposes a hippocampus registration model based on 3D convolution neural network,which can estimate the spatial correspondence between images without iteration during the test and solve the problem of time-consuming in the traditional registration algorithm.Our method can be training to estimate the displacement between the two input images through end-to-end learning from the ground-truth deformation field.The ground-truth deformation field can be used as the initial value of fully convolution network,and then further optimized the parameters of the network using a similarity matrix similarity to the traditional registration method.(1)We propose a hippocampus registration model based on 3D convolutional neural network and successfully solve the hippocampal registration task in the MRI brain image.The method is based on an end-to-end 3D convolutional neural network which using the dual supervision,including known deformation field and similarity matrix,the hippocampus registration model is further optimized in an unsupervised manner while retaining the accuracy of learning from the traditional image registration methods.The trained model can directly estimate the 3D spatial correspondence between the two input images.The experimental results on the public data set prove that the method can effectively estimate the spatial correspondence between images and achieve the alignment of the hippocampus of the brain.(2)We propose a global brain image registration model based on 3D convolutional neural network,which can effectively realize the registration of deep learning methods on whole brain images.Two implementation schemes are proposed:(1)Based on the image patch method and(2)Modifying the convolutional layer.The patch-based method divides the whole brain image into equal-sized image patches,and there are overlaps between the image patches.Image registration is performed separately for each image patch.Finally,the spatial correspondence of the whole brain image is obtained by calculating and splicing the image patch deformation.For the methodof modifying the convolution layer,we adjust the number of convolution channels by calculating the required memory for the whole brain image.Through experiments on the open brain image data set,both schemes can effectively achieve brain image registration.Also,we analyze the advantages and disadvantages of these two methods.(3)We propose a deep-learning-based registration method for multispectral fundus images,which provides a new approach for solving the problem of multimodal image registration.With FCN,the spatial correspondence between images is studied by recording the convolution parameters with the training process.For multispectral fundus images,retinal vessel alignment can represent image alignment.Therefore,in the model training,the image corresponding vascular map is used to calculate the similarity to help the training of the network.The experimental results on the fundus multispectral data show that the weak supervised image registration method guided by segmentation map can effectively realize the registration of fundus multispectral images.