Research On The Labeled Neuron Counting System Of Mouse Brain Slices Based On Image Representation And Registration

In the field of brain science,studying animal brain images is the most direct way to understand the physiological functions of the brain.An important step in image analysis of brain slices is to achieve specific brain region segmentation by matching brain slices to a standard brain reference atlas and perform statistical analysis of labeled neurons in each brain region.Taking mouse fluorescently labeled brain slices as an example,due to the noise and distortion introduced in the preparation of brain slice images,as well as the modal difference with the standard brain atlas images,the brain slice images cannot directly establish an accurate correspondence with the brain atlas images.This in turn affects the accuracy of the number of labeled neurons in each brain region.At the same time,the current research on the localization of brain slice images and the counting of labeled neurons is mainly done manually,which requires a lot of manpower and material resources.Therefore,in this paper,a mouse brain slice labeled neuron counting system based on image representation and registration is developed,which can fully automatically realize accurate brain region localization of mouse brain slice images and accurate counting of labeled neurons in each brain region.The main work completed is as follows:(1)In order to solve the problem of subjective differences between artificially matching brain slice images and brain atlas images and save manpower and material resources,this paper proposes an automatic matching algorithm of brain slice images and brain atlas images based on VGG-Net.The model transforms the task of matching brain slice images and brain atlas images into the classification task of brain slice images,which not only can index the correct brain atlas images for a given brain slice image,but also becomes an important step in the labeled neuron counting system.(2)In order to solve the problem of large direct registration error caused by the modal difference between brain slice images and brain atlas images,this paper adopts the modality conversion method of image representation before registration,and proposes an image representation algorithm based on UNet++.The model is based on the encoder-decoder,and uses deep aggregation as a skip connection path to narrow the semantic gap between decoder and encoder features and reduce the difficulty of network learning.The proposed image representation network can not only reduce the modal difference between brain slice images and brain atlas images,but also enhance the feature correspondence between them,which plays an important role in achieving accurate registration of the two.(3)In order to achieve accurate registration of brain slice images and brain atlas images,and to ensure that the topological structure of the images remains unchanged during the registration process,this paper proposes a brain image registration algorithm based on symmetric differential homeomorphism.The model has symmetric differential homeomorphism,symmetry similarity,and local orientation consistency.Differential homeomorphism guarantees the reversibility and topological structure of image deformation,symmetric similarity can maximize the similarity between deformed images in a bidirectional manner,and local orientation consistency ensures that the topological structure of the image is invariant as it deforms by imposing a regularization constraint on the deformation field.The proposed image registration algorithm can quickly and automatically achieve the ideal image registration effect under the condition that the image topology remains unchanged during deformation.(4)In order to improve the detection and counting effect of labeled neurons.especially to achieve accurate detection of small and multi-type neurons,this paper proposes a labeled neuron detection algorithm,which uses decoupling and anchor-free prediction to detect labeled neurons.Decoupled prediction separates the classification task from the regression task,reduces the training difficulty of the model,and improves the detection effect of small targets;anchor-free prediction reduces the complexity of the detection model and improves the detection speed of the model.The mouse brain slice labeled neuron counting system integrated in this paper can not only process brain slice images automatically and in large batches to achieve accurate counting of labeled neurons in each brain region,but also effectively improve the work efficiency of brain science researchers.