A preliminary investigation on a deep learning convolutional neural networks based pulmonary tuberculosis CT diagnostic model

Abstract

Objective: To evaluate the clinical value of a pulmonary tuberculosis CT diagnostic model based on deep learning convolutional neural networks (CNN). Methods: From March 2017 to March 2018,a total of 1 764 patients with positive sputum for tuberculous bacterium and had received high-resolution chest CT scan in radiology department of Hebei province chest hospital were enrolled. Among them, 937 were male, and 827 were female, aging from 17-73 years (average 38.4). A total of 20 139 CT images (17 kinds of image features) classified by 4 radiologists were used as training dataset to create a tuberculosis CT CNN diagnostic model. The top 5 image features in training set were: infiltrative pulmonary tuberculosis, cavitary pulmonary tuberculosis, pleural thickening, caseous pneumonia and pleural effusion. A total of 302 images were randomly selected from the marked images as testing dataset. The diagnosis of 2 senior radiologists was taken as "golden standard". The differences of sensitivity and accuracy in CT diagnosis between the CNN diagnostic model and the radiologists were compared. The classification error types and numbers of the CNN diagnostic model were recorded. FROC(free response operating characteristic curve)curve was drawn and the highest diagnostic efficiency of the model was measured. Results: The diagnostic accuracy of infiltrative pulmonary tuberculosis, cavitary pulmonary tuberculosis, pleural thickening, caseous pneumonia and pleural effusion by the CNN diagnostic model were 95.33%(10 982/11 520), 73.68%(2 151/2 920), 73.07%(1 128/1544), 83.33%(1 020/1225)and 94.11%(814/865), respectively. The overall diagnostic sensitivity and accuracy of the CNN model were 95.49%(339/355)and 90.40%(339/375), respectively, and the corresponding values ​​of radiologists were 93.80%(348/371)and 92.80%(348/375), respectively, and there was no statistical difference between the CNN model and the radiologists(sensitivity χ2=1.022,P=0.312;accuracy χ2=1.404,P=0.236). FROC curve showed that when sensitivity of the CNN model was 78% and FPI value was 2.48, it reached the highest diagnostic efficiency. The classification error of CNN diagnostic models was mainly confusion of fiber stripe components, cavitary pulmonary tuberculosis, caseous pneumonia and infiltrative pulmonary tuberculosis. Conclusions: The CNN-based pulmonary tuberculosis CT diagnostic model exhibited high sensitivity and accuracy (95.49% and 90.40% respectively). It could assist radiologists in CT diagnosis of pulmonary tuberculosis and deserve further clinical application.