Deep Learning-Based Approaches to Improve Classification Parameters for Diagnosing COVID-19 from CT Images

Abstract

Patients infected with the COVID-19 virus develop severe pneumonia, which generally leads to death. Radiological evidence has demonstrated that the disease causes interstitial involvement in the lungs and lung opacities, as well as bilateral ground-glass opacities and patchy opacities. In this study, new pipeline suggestions are presented, and their performance is tested to decrease the number of false-negative (FN), false-positive (FP), and total misclassified images (FN + FP) in the diagnosis of COVID-19 (COVID-19/non-COVID-19 and COVID-19 pneumonia/other pneumonia) from CT lung images. A total of 4320 CT lung images, of which 2554 were related to COVID-19 and 1766 to non-COVID-19, were used for the test procedures in COVID-19 and non-COVID-19 classifications. Similarly, a total of 3801 CT lung images, of which 2554 were related to COVID-19 pneumonia and 1247 to other pneumonia, were used for the test procedures in COVID-19 pneumonia and other pneumonia classifications. A 24-layer convolutional neural network (CNN) architecture was used for the classification processes. Within the scope of this study, the results of two experiments were obtained by using CT lung images with and without local binary pattern (LBP) application, and sub-band images were obtained by applying dual-tree complex wavelet transform (DT-CWT) to these images. Next, new classification results were calculated from these two results by using the five pipeline approaches presented in this study. For COVID-19 and non-COVID-19 classification, the highest sensitivity, specificity, accuracy, F-1, and AUC values obtained without using pipeline approaches were 0.9676, 0.9181, 0.9456, 0.9545, and 0.9890, respectively; using pipeline approaches, the values were 0.9832, 0.9622, 0.9577, 0.9642, and 0.9923, respectively. For COVID-19 pneumonia/other pneumonia classification, the highest sensitivity, specificity, accuracy, F-1, and AUC values obtained without using pipeline approaches were 0.9615, 0.7270, 0.8846, 0.9180, and 0.9370, respectively; using pipeline approaches, the values were 0.9915, 0.8140, 0.9071, 0.9327, and 0.9615, respectively. The results of this study show that classification success can be increased by reducing the time to obtain per-image results through using the proposed pipeline approaches.