Pathogen Microscopic Image Recognition Methods And Detection Management System Based On Deep Learning

Public health events caused by pathogenic infections have been increasing in recent years,and the outbreak of COVID-19 in particular has brought new challenges to public health emergency management.One of the priorities in the emergency management of pathogen-triggered public health events is the rapid and accurate identification and detection of pathogen.Microscopy remains the most cost-effective means of pathogen detection.With the rapid development of deep learning technology,it continues to be active in the field of medical image analysis and shows excellent processing capabilities.However,the current medical image analysis technology often focuses on the diagnosis and application of cancer and other diseases,but the computer aided diagnosis technology for pathogens classification has not been deeply studied.Full and efficient use of a large number of pathogen microscopic image resources,research and development of deep learning-based pathogen microscopic image analysis technology to achieve rapid automatic detection,and design data and model management mechanism,is undoubtedly of great significance for the enhancement of the emergency management capacity of public health events in China..In this paper,deep learning technology is used to solve the problem of micro-image detection of pathogens as the breakthrough point,and several key problems of micro-image detection of pathogens in the field of medically assisted diagnosis are studied.The classification and determination of several pathogens that are prone to cause major infectious diseases are solved with high accuracy and precision.Micro-image detection algorithms of pathogens based on deep learning are presented,and an intelligent management system with functions of information analysis,decision support and interactive management is constructed.The main contributions are as follows:(1)In order to save cost-expensive labeling work and effectively use unlabeled data to improve the recognition ability of different leukocyte categories,a fine-grained interactive attention learning method based on semi-supervised learning framework is proposed in this thesis.It includes semi-supervised teacher-student module(SSTS)and fine-grained interactive attention mechanism(FIGA).In detail,SSTS uses limited leukocyte images to predict and generate a large number of probability vectors of unlabeled data in a way similar to human eye observation.After screening these probability vectors by Top-k,effective data with high confidence can be selected from a large number of unlabeled data for training.Using a small number of annotated data,FIAL can achieve 93.22% classification accuracy on BCCD dataset,and only 75 annotated images of each type are required,which fully shows the excellent ability of the model in semi-supervised leukocyte classification task.(2)In order to solve the problem of multi-stage malaria parasite recognition,a novel neighbor correlated graph convolutional network(NCGCN)is proposed to solve this challenging task of microscopic image classification.NCGCN is composed of CNN(convolutional neural network)feature learning,neighbor correlation mining and graph feature embedding modules.This method first extracts CNN features from each parasite image,and then combines k-nearest neighbor(KNN)and ò-radius construction algorithm,CNN features and their correlation are introduced into graph convolution network(GCN),and the neighborhood correlation between CNN features is established.On the multi-stage malaria dataset,the model can achieve 94.17% accuracy,94.84% precision,94.17% recall and94.20% F1-score,which verifies that NCGCN model has strong multi-stage malaria recognition ability.(3)In order to use less labeled data to identify a variety of apicomplexan parasites,a semi-supervised graph learning(SSGL)framework is proposed in this thesis,which is composed of CNN(convolutional neural network)feature embedding,learnable graph construction and graph feature learning.Firstly,CNN features are extracted,and then the correlation between samples is mined by using the learnable adjacency matrix.Finally,parasite cells are classified by graph convolution network.In order to verify the performance of SSGL model,we use multiple metrics to evaluate the model.The SSGL model can achieve91.75% high accuracy,91.83% AUC,91.75% sensitivity and 97.25% specificity with only a small amount of labeled data(20%).(4)In order to solve the task of virus micro-image classification,an enhanced graph convolution network(EGCN)is proposed,which is composed of pixel-level feature extraction and group super classification graph embedding.Specifically,the CNN model is firstly used to extract the pixel-level features of virus images,then the obtained CNN features are used to construct the adjacency matrix,and the relationship between samples is mined by GCN combined with group super classification loss.In order to test the ability of EGCN algorithm to diagnose virus morphology,this thesis employs several evaluation metrics to measure the model.The EGCN model can achieve 3.40% top-1 error rate,1.88% top-2 error rate,96.65% accuracy and 96.60% recall.A series of comparative experiments show that each module in this method plays an important role in virus morphological classification.(5)Based on the developed pathogen microscopic image classification algorithms from deep learning,the subsystem schemes of automatic detection,manual intervention,self-learning optimization and auxiliary decision-making are given.For the actual needs of hospital users,the relevant sample(patient)database and the information management subsystem with the functions of query and outgoing diagnosis report are constructed.Through the organic correlation of the two subsystems,the research of microscopic image information analysis,decision-making and management system based on deep learning is successfully completed.