Applications Of Deep Neural Network In Some Medical Problems

In recent years,deep neural networks have made great progress,and it has been applied in many healthcare fields,such as medical image recognition,electronic health record management,drug discovery.Its application on healthcare has become more and more eyecatching,and it has great significance in assisting doctors in diagnosis and planning treatments,personal health management and personalized customized treatment of patients,research and development in the pharmaceutical industry,the prediction and monitoring of epidemics by government regulatory agencies,etc.Different medical problem has different application scenarios.In real applications,it still has improving room for most neural networks due to they are not special designed for the application scenarios.The performance of networks will be enhanced if they are designed case by case.In this thesis,the analysis method of breast cancer early detection,gastric disease diagnosis,and Corona Virus Disease 2019(COVID-19)epidemic trend prediction is proposed separately.1.For breast cancer early detection,a mammogram benign-malignant classification method is proposed based on deep convolutional neural networks.The proposed method combines multi-instance learning and transfer learning to analyze the original whole mammogram.It is using Region of Interest(ROI)to pre-train a Res Net50,and using this pretrained Res Net50 as features extraction based on transfer learning method.Then using the features extraction to extract features from different patches of the original whole mammogram,these features are concatenated and input to fully connected layers to finish the classification task at last.The experimental results on an open dataset show that the proposed method has some practicality.2.For gastric disease diagnosis,a gastroscope image classification method of white light gastroscopy is proposed based on deep convolutional neural networks.The proposed method changes the Dense Net structure by borrowing ideas of the bottom-top pathway and the topbottom pathway of FPN(Feature Pyramid Network).In the proposed method,multi-scale features can be reused,and high dimension features are merged with low dimension features to enrich the semantics of all scale features.Through experiments,the performance of the proposed gastroscope image classification method is better than all other classic classification neural networks with a little price of the increased scale of parameters.3.For COVID-19 epidemic trend prediction,a dynamic forecasting method of the rate of COVID-19 virus transmission is derived based on LSTM(Long Short-Term Memory networks),and then LS-Net(LSTM-SEIR Network)is proposed based on LSTM and SEIR epidemic model.LS-Net mitigate the problem that SEIR model can’t predict the epidemic trend dynamically because its parameters are not real-time adaptively.In experiments,results of LSNet show that LS-Net can predict epidemic trends validly and they are much better than that of SEIR model.