Chapter 12 - Classification of COVID-19 and non-COVID-19 lung computed tomography images using machine learning

Abstract

A novel coronavirus disease 2019 (COVID-19) was first reported in Wuhan, China in late December 2019. In March 2020, World Health Organization (WHO) declared this sudden epidemic as a global pandemic. It is highly contagious and can cause serious lung inflammation. The typical symptoms are fever, cough, shortness of breath, headache and sore throat. Till 23 August 2021, a total of more than 211 million cases of COVID-19 have been reported to WHO worldwide, with a total of more than 4.4 million of deaths. Hence early detection is crucial to control the spread. Currently, the key diagnosis method is the reverse transcription polymerase chain reaction (RT-PCR) test using swab samples. However, it is subject to certain limitations, such as low sensitivity and shortage of kits. To address these issues, lung computed tomography (CT) scan can be the alternative as it is fast, easy, and proven to be sensitive in detecting COVID-19 cases. This study presents an automated method to differentiate the COVID-19 CT images from the Non-COVID-19 images using different convolutional neural networks (CNN) through three stages procedures. In the first stage, the dataset which consists of 746 images of COVID-19 and Non-COVID-19 was split into 3 parts for training, validation, and testing, respectively. The training and validation data were then applied with different augmentation techniques to increase the dataset, while the testing data remained with no augmentation. In stage 2, 10 different pretrained CNNs were initialized to train and classify the binary class. In stage 3, gradient descent class activation mapping (GradCAM) was used for abnormality localization. The best performance was achieved by ResNet152, ResNeXt, GoogleNet, and DenseNet201 with the highest overall accuracy of 98.51%. ResNet152, GoogleNet, and DenseNet201 had achieved a sensitivity of 100%, and specificity of 97.06%, whereas ResNeXt had achieved a sensitivity of 96.97%, and specificity of 100%.