Multiple Diseases Classification Of X-ray Chest Film Based On SERes-RACNN

Chest X-ray is a mandatory item in checkups.During the diagnosis of chest radiograph,the long time,high load and the interference of human subjective factors may affect the diagnosis accuracy.In addition,there are some problems in our country,such as high pressure of hospital reception in developed areas and low accuracy of hospital diagnosis in underdeveloped areas.Therefore,the efficient and intelligent auxiliary diagnosis method has become one of the urgent research topics.However,the hotspots in this field is mainly focused on the algorithm of diagnosis single disease,while there are few research on the algorithm of simultaneous diagnosis of multiple diseases.In this thesis,a SERes-RACNN network based on the Recurrent Attention Convolutional Neural Network(RACNN)is proposed for the fine-grained classification algorithm of multi disease classification,which visualizes the lesion area of the disease.Finally,a set of auxiliary diagnosis algorithms that can diagnose eight categories include pulmonary emphysema,pulmonary congestion,pneumothorax,heart shadow enlargement,hydrothorax,pneumonia and pulmonary effusion,normal is realized,which can visualize the lesion area at the same time.The main work and contributions of this thesis are as follows:(1)In view of the problem of low classification accuracy of traditional classification algorithm,this thesis puts forward that the diagnosis classification of chest radiograph can regarded as fine-grained classification and a SERes-RACNN network was improved to make the average receiver operating characteristic curve(AUC)value of the eight categories in the X-ray chest radiography classification of 0.9004,which is 1.97% higher than the traditional method.(2)In view of the problem of low generalization performance of medical data sets,the model training strategy and optimization strategy were improved.The three methods of increasing batch size,weighting loss function and model fusion were used to accelerate convergence,prevent overfitting and improve generalization accuracy.Finally,the AUC was further increased to 0.9114.(3)In view of the problem of single algorithm feedback,the multi classification results of chest films are visualized by grad cam(Grad class activation mapping).The visualization results are presented by heat map.Finally,the coarse lesion area is segmented by setting a mask threshold,and realize the automatic generation of the mask of the lesion area.In this thesis,the disease classification on the X-ray chest radiography data set has achieved the expected results,and the multi-class visualization of the disease focus area of the X-ray chest radiography has been successfully realized.This will help to relieve the pressure of hospital reception,save medical resources,improve the accuracy of diagnosis,and has important practical significance for the realization of intelligent medical and telemedicine diagnosis in the future.