Proposing a New Feature Clustering Method in Order to the Binary Classification of COVID-19 in Computed Tomography Images

Abstract

Since the outbreak of coronavirus (COVID-19), different image classification algorithms have been proposed to detect suspected cases based on computed tomography scan (CT) images. Due to the lack of a valid database and the lack of CT scan images, the results of most of these algorithms have not been reliable and they have not been trained with the standard number of CT images. So, in this paper, a new classification algorithm is proposed in which a valid database of CT scans with 6322 images is used in order to train the proposed algorithm. This algorithm includes three main steps: In the first step, the features are extracted using two pre-trained convolutional neural networks called ResNet18 GoogleNet. Due to generating many features using the ResNet18 and GoogleNet, a new dimension reduction method is proposed in which a new metaheuristic algorithm called Curling Optimization Algorithm (COA) is proposed as a part of the dimension reduction process. The main idea of the new dimension reduction algorithm is the combination of COA and DBSCAN algorithms. In the third step, binary classification is done by the Support Vector Machine (SVM). According to the results of the classification of the CT images, the proposed algorithm has achieved high accuracy more than 93% in ResNet18 features and more than 91% in GoogleNet features.