Pre-trained Deep Learning Models for COVID19 Classification: CNNs vs. Vision Transformer

Abstract

The fast proliferation of the coronavirus disease 2019 (COVID19) has pushed many countries' healthcare systems to the brink of disaster. It has become a necessity to automate the screening procedures to reduce the ongoing cost to the healthcare systems. Although the use of the Convolutional Neural Networks (CNNs) is gaining attention in the field of COVID19 diagnosis based on medical images, these models have disadvantages due to their image-specific inductive bias, which contradict to the Vision Transformer (ViT). This paper conducts comparative study of the use of the three most established CNN models and a ViT to deal with the classification of COVID19 and Non-COVID19 cases. This study uses 2481 computed tomography (CT) images of 1252 COVID19 and 1229 Non-COVID19 patients. Confusion metrics and performance metrics were used to analyze the models. The experimental results show all the pre-trained CNNs (VGG16, ResNet50, and IncetionV3)outperformed the pre-trained ViT model, with InceptionV3 as the best performing model (99.20% of accuracy).