Study On EBTB Assisted Diagnosis Based On Convolutional Neural Network

Endobronchial Tuberculosis(EBTB)is a special type of Tuberculosis.In less developed areas,doctors often do not have enough clinical experience to make a correct diagnosis of EBTB,resulting in early missed diagnosis and misdiagnosis,which worsens the condition and seriously affects the quality of life of patients.In recent years,although AI technology has been widely used in clinical disease diagnosis,there is currently no research on the auxiliary diagnosis of EBTB.Therefore,this paper establishes the world’s first EBTB dataset and develops an EBTB auxiliary diagnosis system based on Convolutional Neural Network(CNN),which fills the gap in the field of EBTB intelligent diagnosis and has great significance in the early prevention and treatment of diseases.This paper constructed the world’s first standard EBTB dataset,with a total of20,469 EBTB images,including 18161 normal images,453 inflammatory infiltrating images(type I),1275 ulcer necrotic images(type II),and 580 granulomatous images(type III).All images were screened and labeled by chief physicians with more than 15 years of clinical experience.In order to improve the real-time performance and accuracy of EBTB assisted diagnosis,Mobilenet-NP and Mobilenet-TC were obtained based on MobilenetV2 network using transfer learning strategy training,which were the binary classification model of normal and pathological images and the three-classification model of pathological lesions.The experimental results showed that the accuracy of MobilenetNP model was 89.37%,and the image detection speed was 0.023 seconds.The accuracy of Mobilenet-TC model for EBTB classification was higher than 80%,reaching the diagnosis standard of attending doctors.In Mobilenet-TC lesion classification,due to the insignificant feature difference between tuberculosis type II and type III,it is easy to identify type III lesions as type II.To solve this problem,this paper proposes an improved CNN algorithm based on the mechanism of location-embedded attention.In Mobilenet-TC network,the location embedding function block is added to decompose channel attention into two parallel one-dimensional feature coding processes,and the spatial attention is effectively integrated into the generated attention matrix.Experimental results showed that this method improved the accuracy of TB type III from 83.2% to 90.8%.Considering that the attention mechanism can improve the classification accuracy,this mechanism is also introduced into the normal and pathological image binary classification model Mobilenet-NP.The experimental results show that the binary classification sensitivity is improved from 85.03% to 88.57%.Finally,a set of EBTB auxiliary diagnostic system which can perform bronchoscopy is developed.The system interface will display the video stream of the bronchoscope host synchronously,and the images collected by the doctor through the medical foot will be transmitted into the system in real time to realize the intelligent detection of EBTB lesions.The clinical application of this system can reduce the workload of doctors,and it can greatly improve the detection rate of EBTB in hospitals in underdeveloped areas,so as to achieve early detection and early treatment of EBTB.