Endobronchial tuberculosis(EBTB)is a special type of pulmonary tuberculosis,which refers to tuberculosis that occurs in the trachea,bronchial mucosa and submucosa.With the development of intelligent medicine,the medical intelligent auxiliary diagnosis system based on Convolutional Neural Network(CNN)has achieved good clinical results,but the research survey found that the medical auxiliary diagnosis system for EBTB has not been developed yet.Unfortunately,the EBTB bronchoscopy image database has not yet been established.Therefore,this study established the world’s first EBTB image database,and proposed to use the densely connected convolutional network(Dense Net)in CNN combined with the convolutional attention module(CBAM)to construct the CBAM-Dense Net network model,and designed an EBTB intelligent auxiliary diagnosis system based on CBAM-Dense Net.The main work is as follows:Firstly,the EBTB bronchoscopy image database was constructed.All EBTB cases were obtained from Jiangxi Chest Hospital,and the screening and labeling of EBTB images were completed by two director-level bronchoscopy physicians.A total of 19,721 valid images have been collected so far,including 17,585 type 0 normal images,418 type I(inflammation and infiltration)images,1166 type II(ulcer and necrosis)images,and 552 type III(granulation proliferation)images.Secondly,the Dense Net network is used as the benchmark network to realize the intelligent classification of EBTB bronchoscopy images,which can basically correctly distinguish normal,type I,type II and type III.The experimental results show that the accuracy of the optimal model Dense Net-OS on the test set is 0.8905,and the image prediction time is 0.093 seconds.Thirdly,in order to improve the accuracy of EBTB lesion detection and classification,the CBAM module is used to optimize the Dense Net network,and a new network CBAM-Dense Net is proposed.The experimental results show that the accuracy of the optimal model CBAM-Dense Net-RI on the test set is 0.9213,and the image prediction time is 0.031 seconds,and the overall performance is better than the Dense Net-OS model.Finally,based on the CBAM-Dense Net-RI model,this paper develops and designs an EBTB intelligent auxiliary diagnosis system.In the clinical test,the accuracy of the system was 0.8557,which reached the diagnostic standard of the deputy chief physician.Therefore,the system can effectively help doctors to make early diagnosis of EBTB patients.This study makes up for the lack of EBTB computer detection system in the medical field,and is more conducive to further promoting the application of deep learning in the medical field. |