Covid-19 Detecting in Computed Tomography Lungs Images Using Machine and Transfer Learning Algorithms

Abstract

Coronavirus disease 2019 (COVID-19), a rapidly spreading infectious disease, has led to millions of deaths globally and has had a significant impact on public healthcare due to its association with severe lung pneu- monia. The diagnosis of the infection can be categorized into two main approaches, a laboratory-based approach and chest radiography approach where the CT imaging tests showed some advantages in the pre- diction over the other methods. Due to restricted medical capacity and the fast-growing number suspected cases, the need for finding an immediate, accurate and automated method to alleviate the overcapacity of radiology facilities has emerged. In order to accomplish this objective, our work is based on developing machine and deep learning algorithms to classify chest CT scans into Covid and non-Covid classes. To obtain a good performance, the accuracy of the classifier should be high so the patients may have a clear idea about their state. For this purpose, there are many hyper parameters that can be changed in order to improve the performance of the artificial models that are used for the identification of such illnesses. We have worked on two non-similar datasets from different sources, a small one consisting of 746 images and a large one with 14486 images. On the other hand, we have proposed various machine learning models starting by an SVM which contains different kernel types, KNN model with changing the distance measure- ments 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 then two consecutive con- volution layers followed by a single pooling layer each time. The machine learning models showed better performance compared to CNN on the small dataset, while on the larger dataset, CNN outperforms these algorithms. In order to improve the performance of the models, transfer learning has also been used 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% speci- ficity and 86% F1-score on the small dataset while the respective scores on the large dataset were 97.52%, 97.28%, 97.77% and 98%. Experimental interpretation advises the potential applicability of ResNet50V2 transfer learning approach in real diagnostic scenarios, which might be of very high usefulness in terms of achieving fast testing for COVID19. Povzetek: Raziskava se osredotoča na razvoj algoritmov strojnega in globokega učenja za razvrščanje CT posnetkov prsnega koša v razrede Covid in ne-Covid. Rezultati kažejo, da je pristop prenosa učenja ResNet50V2 najbolj učinkovit za hitro testiranje COVID-19.