Deep Features Comparative Analysis for COVID-19 Detection from the Chest Radiograph Images

Abstract

The COVID-19 pandemic needs present resources for the mitigation of its distressing effects. Suspected cases of COVID-19 need rapid, early and accurate detection to prevent the spread on a large scale. Existing tests for COVID-19 diagnosis are slow and need a few hours to generate the results. The situation is even worsened in developing countries such as Pakistan due to the lack/limited facilities of reliable COVID detection test. Moreover, existing tests like PCR are not highly reliable and often result in an incorrect diagnosis. Early and reliable COVID-19 diagnosis is vital to reduce the complications in the treatment of COVID-19 patients. This paper presents a comparative analysis of different deep learning-based feature extraction models for COVID-19 detection. Towards this end, we have employed twelve pretrained deep learning models, i.e., Mobilnetv2, Darknet19, Densenet201, Squeeznet, Alexnet, Googlenet, Inceptionv3, Inceptionresnetv2, Resnet50, Resnet18, Resnet101 and Shufflenet for features extraction from the chest radiograph images. The support vector machine (SVM) was then trained using these features for classification of chest radiograph images into COVID or non-COVID/normal. We evaluated the performance of each deep learning model on a standard COVID-19 Radiography Database that consists of 13808 Chest X-Ray images including 3616 COVID-19 positive images and 10192 X-ray COVID-19 negative images. Our experimental results illustrate that the Mobilnetv2 model provides the most effective deep features for image representation. The accuracy of98.5% signifies the effectiveness of the Mobilnetv2 model for reliable detection of COVID-19.