Research On Classification And Visualization Of Covid-19 X-rays Based On Deep Learning

COVID-19 is a highly contagious disease.Since December 2019,it has killed more than two million people worldwide,causing great harm to people around the world.Therefore,timely and accurate identification of COVID-19 patients can not only provide timely treatment for patients,but also effectively prevent the spread of the epidemic.The main test for COVID-19 at this stage is nucleic acid test,but nucleic acid test has a high false negative rate,and multiple tests are needed to confirm the diagnosis.Chest X-ray is the standard for diagnosis of common pneumonia and can also be the basis for diagnosis of COVID-19.Due to the high visual similarity between COVID-19 and common pneumonia medical images,it is difficult for many professional doctors to make a judgment.Therefore,the accurate classification of COVID-19 X-ray films is of great significance in clinical practice.In this paper,the COVID-19 Chest X-Ray Database open dataset was used as the research object.The packet contains four categories of COVID-19 X-rays,ordinary pneumonia x-rays,other viral pneumonia x-rays and normal chest x-rays.In the case of relatively small data volume,the network model is pre-trained on the Chest X-ray14 dataset by using the transfer learning method,and then migrated to the dataset for fine-tuning.The main research contents and results are as follows:(1)The paper compares the classification effects of three convolutional neural network models,VGG19,Res Net50 and Che XNet,on this dataset,and the results show that the accuracy of Che XNet model is the highest,reaching 94.79%.(2)The paper compares the classification accuracy of three different models on the SVM classifier,and the results show that the classification effect of Che XNet+SVM model is better,with the accuracy reaching 96.33%.In this paper,the prediction results of Che XNet network model were visualized,and the focus area of chest X-ray was displayed by generating heat map,which increased the persuasions of classification results.In this paper,CAM,Grad-CAM,Grad-CAM++ and Score-CAM methods were used to conduct a visualization study on the predicted COVID-19 X-ray films,and the focus areas of the images were shown by generating heat maps.At the same time,the visualization effects of the four methods were compared,and the results showed that the heat maps generated by the four methods were not significantly different in visual sense,but the Score-CAM method performed relatively well in the detailed texture structure,and the lesion area was more accurate.The visualization method can clearly express the decision of the convolutional network model,and better assist doctors to complete the diagnosis and treatment of COVID-19.