Investigation of Deep Convolutional Neural Network (CNN) approaches’ accuracy for the detection of COVID-19

Abstract— the world those days focuses on protecting human health and combating the irruption of coronavirus patients (COVID-19). As results of its extra ordinarily contagious infection that have caused a disturbance in everyone's lives in various ways. For early screening, Reverse Transcription Protein Chain Reaction (RT-PCR) test is used to examine the onset of the patients by detecting the RNA material of the virus among the patients’ samples. Recent results indicate that the applying of X-ray images and X-radiation (CT) improves the detection accuracy of this disease. However, the classification task of medical images is tough due to several factors such as lack of dataset for COVID-19, and difficulty in identifying type of infection. Recent research works have been proposed for COVID-19 detection that has been applied on specific datasets. Thus, it is vital to validate their performance on various datasets with different imaging disease conditions. The paper presents a comparison study between top performer CNN models that recorded the very best detection accuracy in image detection and classification: COVID-Net, VGG16, ResNet, Bayesian, DenseNet, and DarkNet. Such CNN approaches can assist medical staff in the early detection of infection. Additionally, we improved dataset in terms of quality, clarity, and quantity using augmentation technique. The quantitative results show that Darknet and COVID-net yield high detection accuracy when applied on CT and X-ray dataset. We validated our results by training the models on multiple different datasets, using CPU and GPU with various bach sizes and optimizers.