Research On Detection Of B End.3 Cells Based On Convolutional Neural Network

In recent years,neurological diseases such as Alzheimer's disease have seriously affected people's lives,among which the main cause of Alzheimer's disease is the change in the structure and function of the blood-brain barrier.Therefore,the detection of the function and active state of cerebral microvascular endothelial cells,the main component of the blood-brain barrier,is of great significance for the treatment of Alzheimer's disease.Since the pathogenesis of many diseases has been established in mice,it is of great significance to reproduce the pathogenesis of Alzheimer's disease in vitro by using rat brain microvascular endothelial cells(bEnd.3)for rapid screening of relevant therapeutic drugs.Traditional bEnd.3 cell detection requires artificial design features,and then classification recognition through machine learning classification algorithms,but because the detection method based on artificial design features can not use the deep-level features of the image and is easily affected by factors such as light changes,Which in turn leads to false detections and missed detections.The detection method based on deep learning can integrate the extracted features to extract more abstract features,especially in the medical field of image classification,object detection and other studies have achieved very good results.This thesis takes bEnd.3 cells as the research object and constructs a neural network model based on convolutional neural network for object detection of bEnd.3 cells.The main contents are as follows:First of all,this thesis designs a neural network model based on regression and adopts a multi-convolution parallel structure.Multi-convolution parallel structure that contains both the 7× 7 convolution kernels and 3× 3 convolution kernels and the cascaded cross-channel parameter pool layer,7× 7 convolution kernels should check the feature information strong ability of semantic representation,3 × 3 convolution kernels should check the feature information geometric detail representation ability,the multi-convolution parallel structure improves the detection performance of the network model for small-size bEnd.3 cells by fusing feature information of different scales.At the same time,the cascaded cross-channel parameter pool layer also reduces the calculation amount of the network model and improves the calculation efficiency of the network model.Secondly this thesis introduced the attention mechanism,it performs global average pooling operation on the output feature map of the multi-convolution parallel structure to get global characteristics of different channels,by gating mechanism to assign different weights to different feature channels according to their importance,to enhance the key feature information,weaken the irrelevant information,to solve the data redundancy problem caused by multi-scale data fusion.Then,the experimental data set used in this thesis is constructed,and the image brightness and contrast are improved by multi-scale retinal enhancement algorithm with color restoration.Secondly,in view of the small amount of data contained in the data set,this thesis uses spatial geometric transformation to expand the data amount.Finally,mark the target area in the data image,complete the corresponding construction of the training image data and the target detection area,to pave the way for the next comparative experiment on the data level.Experimental results show that the convergence value of the loss function of the neural network model designed is 0.28 lower than the original network structure's after 5000 iterations of training,the accuracy value is increased to 91.2%,the mean average precision increased from 64.3% to 70.1%,at the same time higher than the mean average precision value of traditional object detection algorithm based on regional proposal,to provide a more scientific evaluation basis for bEnd.3 cell detection.