Deep neural network models for COVID-19 diagnosis from CT-Scan, explainability and analysis using trained models

Abstract

The world is going through a severe viral pandemic which is caused by COVID- 19. People infected with this virus, experience severe respiratory illness. The virus spreads through particles of saliva or droplets from an infected person. There are ways of identifying COVID-19 based on the symptoms such as fever, dry cough, tiredness, but these symptoms are similar to other existing viral or respiratory infections. There is no quick approach in diagnosing if a patient is infected or not. To overcome the drawbacks mentioned, a faster diagnosis is needed which leads us to the objective of this study. we intend to construct a diagnostic approach that uses pre-existing data mostly on COVID-19, as well as take datasets from other respiratory diseases. We will apply deep learning models to the acquired datasets enabling us to obtain more accurate and efficient results. We aim to use Deep Neural Network models namely Convolutional Neural Network models (CNN) such as VGG19, Inception v3, MobileNetV2, and ResNet-50. These four models are pre-trained and they classify the CT-Scan images based on the trained learning approaches. The result of each model is compared among the models to get faster and more accurate results. This paper also proposes a "Hybrid" model which is composed of a Convolutional Neural Network (CNN) and a Support Vector Machine (SVM). The Hybrid Model is shallow and just as accurate as the pre-trained models. In light of the exactness of the result and the minimal measure of time needed for image classi cation, we will be able to diagnose more accurately and effectively.