Image Registration Of Microscope Slice Images Based On Convolutional Neural Network

Objective: Image registration is a basic task in medical image analysis.Traditional image registration tasks,especially non-rigid registration algorithms,require higher computational costs and longer execution times.However,with the development of computer technology and the revival of deep learning,convolutional neural network have shown excellent results in processing image classification,object detection and other tasks,and also have the ability to handle image registration tasks.In this study,an unsupervised registration model based on a convolutional neural network is used to obtain the transformed images one step at a time,effectively reducing the registration time of a pair of microscope slice images.At the same time,for many cases in which the data containing histological artifacts are directly discarded and data information is lost,this article processes images containing cracked areas,and the processed images can be used for subsequent registration and other tasks.Methods: The methods in this article mainly includes three parts.First,including the image normalization,affine registration,and encapsulating the data.Second,the images containing crack regions will be Gaussian filtered,binarized,targets region extracted,and targets merged to reduce the crack area.Finally,the convolutional neural network is used to model the spatial deformation field,the input images get the spatial deformation field through the coder decoder structure with the skip connections,and then the moving image will be interpolated through the spatial transformation function to obtain the registered image.This process does not require any artificial ground truth.Results:1.The registration results are compared with SimpleITK and PyElastix based on the ITK toolkit.It is also compared with the registration results based on VGG16 network.The registration results of the method used in this article are also displayed in three dimensions.At the same time,the mutual information value and cosine similarity are calculated quantitatively for the registration results.2.Processed 20 of the 141 images in the E14 B group and 7 of the 75 images in the E14 C group with cracked areas,retaining the information contained in the data.Conclusion: The registration model based on convolutional neural network used in this article shows its superiority over the ITK toolkit and the registration model based on VGG16.At the same time,a well processing of the image containing the crack region can make this part of the data continue to be used in subsequent experiments.