Research On Automatic Hierarchical Recognition Algorithm For COVID-19 Based On Deep Learning

At the end of 2019,COVID-19 broke out worldwide and became a pandemic.Every day,a large number of newly confirmed patients poured into hospitals,resulting in a severe shortage of medical resources and staff.In order to allocate medical resources and staff appropriately,it is significant to accurately classify COVID-19 patients into mild and severe patients.At present,most of the relevant studies are based on single CT images,and the information of single CT images is used to classify COVID-19.This will lead to low universality of the model and high requirements on the input CT images.Once the input CT images do not have a strong representation of the disease,it may cause great errors.The research content of this paper is devoted to improving the universality of the model and the effect of classification,mainly doing the following three works.(1)In order to conduct experimental verification of the hierarchical model proposed in this study,we need to make CT image data sets and clinical indicator data sets.According to the characteristics of wide lung field and clear lesion in the middle part of CT section sequence,we selected the CT images of the middle part of each patient and made a CT image data set.In the process of making clinical indicators dataset,we to multiple may be associated with severity COVID-19 clinical indexes of Pearson significance analysis and correlation analysis,identify the lymphocyte count,lymphocyte,neutrophil count and the age and percentage COVID-19,there exists a correlation between the severity was made according to the clinical indicators data set.(2)Considering that three-dimensional information of patients’ lungs can be represented by CT image sequences,this study proposed an automatic classification model of COVID-19 severity based on CT image sequences,called TAS.The model takes multiple CT images of the same patient as input,uses traditional image processing operations to extract characteristic information,calculates the proportion of average and maximum lesion area of patients,and finally classifies them by SVM.Finally,the classification accuracy and recall rate of this model on CT image test set are 87.65% and 86.67%,both better than the two deep learning models(VGG16 model and Res Net50 model)in the comparative experiment,which verifies that the use of CT sequence input can improve the identification ability of the model for severe patients.Thus,the recall rate of the model can be improved.(3)In order to further improve the performance of the classification model,we proposed the idea of integrating CT images and clinical indicators.Under the guidance of this idea,we improved Res Net50 model and TAS model,and proposed two data fusion classification models,named Res＿SVM and TAS+ respectively.Res＿SVM model integrates CT image data and clinical index data of the same patient,greatly improving the recall rate of Res Net50.Finally,in the experimental stage,the classification accuracy of Res Net50 model is 92.94%,and the recall rate is 87.67%,and all evaluation indexes have achieved the optimal results.These results indicated that fusion CT images and related clinical indicators can improve the performance of the model to a certain extent.In addition,we found through the experimental results of TAS+ model that the performance of TAS+ model decreased with the increase of the number of clinical indicators of fusion.Through the analysis of the comprehensive experimental results of the two models,it was found that although the fusion of four clinical indicators could solve the contingency problem of CT image selection,it would reduce the distinction between mild and severe patients,which was beneficial and disadvantageous to the improvement of model performance.This study was based entirely on data provided by the hospital after a rigorous desensitization procedure,without the participation of doctors,including the production of two data sets.However,the classification models proposed in this study still achieve excellent results,which indicates that these models not only have excellent results in the automatic classification of mild and severe patients,but also have good universality,with almost no requirements for input data.