Diagnosing Chest X-Rays For Early Detection Of COVID-19 And Distinguishing It From Other Pneumonia Using Deep Learning Networks

ABSTRACT

Chest X-rays of COVID-19 patients helped detect and diagnose the virus early and assess the severity of the infection. Therefore, assessing the severity of Covid-19 infection plays an important role in determining the patient's condition and distinguishing cases that need intensive clinical care. But there are challenges facing doctors and radiologists because of vital signs, different areas of infection, and the wide differences between many cases. Therefore, deep learning techniques play an important role in solving these challenges for early detection of Covid-19 disease in X-ray images and distinguishing it from other pneumonias. In this study, three CNN models, AlexNet, ResNet-18 and GoogleNet, are proposed to diagnose a data set collected from multiple sources. Each model diagnosed a multi-class data set (four classes) and a two-class data set. All dataset images were processed and removed from the data before they were entered into CNN networks. Because the data set is unbalanced, a data augmentation technique was applied to balance the data set between collecting classes. Characteristics were extracted in a hybrid way between CNN models and Gray-level Cooccurrence Matrix (GLCM), Local Binary Pattern (LBP) and Discrete Wavelet Transform (DWT) algorithms and combined all the algorithms into a single vector for each image. All networks achieved superior performance in diagnosing COVID-19 and distinguishing it from other pneumonias. GoogleNet reached an accuracy, sensitivity, specificity, and AUC of 94.10%,95%, 97.75% and 96.13%, respectively with the dataset of multiple classes. while ResNet-18 achieved an accuracy, sensitivity, specificity, and AUC of 98.60%, 98%, 98%, and 97.10%, respectively with two-class (COVID-19 and normal).