An Investigation of COVID-19 Diagnosis and Severity Detection Using Convolutional Neural Networks

Abstract

AbstractSevere acute respiratory syndrome coronavirus 2 (SARS-CoV-2) causes coronavirus disease 2019 (COVID-19). It is a contagious disease that has infected more than millions of people around the globe. COVID-19 can be diagnosed based on the amount of infection in the lungs. Apart from the gold standard reverse transcription-polymerase chain reaction test, X-rays and CT scans can be used to diagnose and detect COVID-19 severity. Due to technological advancements, deep learning models play a crucial role in COVID-19 analysis because of their efficiency and accuracy. Hence, the present work investigates various research works to detect COVID-19 using a convolutional neural network. It compares architectures such as Inception, MobileNet, DenseNet, and new architectures like CovidNet and CovidSDNet were developed. Some of the investigations in COVID-19 detection were conducted by performing data augmentation and ensembling techniques with and without transfer learning. Most research works used accuracy, precision, recall, and F1-score as the performance metrics for evaluation. The numerical comparison analysis shows that the earlier works achieved an accuracy of about 89 to 98 percent. However, in most investigated research, multiclass classification is performed to classify the given CT scan or X-ray image into COVID-19, normal or pneumonia. In the current situation, there is a possibility that some people with COVID-19 might have other respiratory diseases as well. Hence, the investigations suggest that multi-label classification with convolutional neural networks can be suitable to determine the combination of respiratory problems present along with COVID-19.KeywordsReverse transcription-polymerase chain reactionGround-glass opacitiesCOVID-19Computerized tomography scansX-rayConvolutional neural networks