Deep Feature Extraction for Detection of COVID-19 Using Deep Learning

Abstract

SARS-CoV-2, a severe acute respiratory syndrome that is related to COVID-19, is a novel type of influenza virus that has infected the entire international community. It has created severe health and safety concerns all over the globe. Identifying the outbreak in the initial phase may aid successful recovery. The rapid and exact identification of COVID-19 limits the risk of spreading this fatal disease. Patients with COVID-19 have distinctive radiographic characteristics on chest X-rays and CT scans. CXR images can be used for people with COVID-19 to diagnose their disease early. This research was focused on the deep feature extraction, accurate detection, and prediction of COVID-19 from X-ray images. The proposed concatenated CNN model is based on deep learning models (Xception and ResNet101) for CXR images. For the extraction of features, CNN models (Xception and ResNet101) were utilized, and then these features were combined using a concatenated model technique. In the proposed scheme, the particle swarm optimization method is applied to the concatenated features that provide optimal features from an overall feature vector. The selection of these optimal features helps to decrease the classification period. To evaluate the performance of the proposed approach, experiments were conducted with CXR images. Datasets of CXR images were collected from three different sources. The results demonstrated the efficiency of the proposed scheme for detecting COVID-19 with average accuracies of 99.77%, 99.72%, and 99.73% for datasets 1, 2 and 3, respectively. Moreover, the proposed model also achieved average COVID-19 sensitivities of 96.6%, 97.18%, and 98.88% for datasets 1, 2, and 3, respectively. The maximum overall accuracy of all classes—normal, pneumonia, and COVID-19—was about 98.02%.