Covid-19 classification using deep learning

Abstract

ABSTRACT Coronavirus disease 2019 (COVID-19) is a fast-spreading infectious disease that causes lung pneumonia which killed millions of lives around the world and has a significant impact on public healthcare. The diagnostic approach of the infection is mainly divided into two broad categories, a laboratory-based and chest radiography approach where the CT imaging tests showed some advantages in the prediction over the other methods. Due to the limited medical capacity and the dramatical increase of the suspected cases, the need for finding a quick, accurate and automated method to mitigate the overloading of radiologists’ efforts for diagnosis has emerged. In order to achieve this goal, our work is based on developing machine and deep learning algorithms to classify chest CT scans into Covid or non-Covid classes. We have worked on two non-similar datasets from different sources, a small one of 746 images and a larger one with 14486 images. In the other hand, we have proposed various machine learning models starting by an SVM which contains different kernel types, K-NN model with changing the distance measurements and an RF model with two different number of trees. Moreover, two CNN based approaches have been developed considering one convolution layer followed by a pooling layer for the first approach, then two consecutive convolution layers followed by a single pooling layer each time for the second approach. The machine learning models showed better performance comparing to the CNN on the small dataset. While on the large dataset, CNN outperforms these algorithms. In order to improve performance of the models, transfer learning also have been used in this project where we trained the pre-trained InceptionV3 and ResNet50V2 on the same datasets. Among all the examined classifiers, the ResNet50V2 achieved the best scores with 86.67% accuracy, 93.94% sensitivity, 81% specificity and 86.11% F1-score on the small dataset while the respective scores on the large dataset were 97.52%, 97.28%, 97.77% and 97.60%. Experimental observations suggest the potential applicability of ResNet50V2 transfer learning approach in real diagnostic scenarios, which could be of very high utility in terms of achieving fast testing for COVID-19.