A Novel Architecture of Convolutional Neural Network to Diagnose COVID-19 Disease

Abstract

COVID-19 is a rapidly spreading viral infection with its rapid spread and increasing numbers of patients that lasted until the time of writing this research caused which has led to a great loss all over the world, economically and socially, andit has become a necessity to rapid diagnose the condition and contain it from spreading. We are aware that at the beginning of the pandemic, researchers in the field of artificial intelligence proposed a large number of automatic diagnosing models in an effort to aid radiologists and improve diagnosing accuracy based on X-ray images and Computed Tomography (CT) images, which have since been widely adopted to confirm positive COVID-19 RT-PCR tests, this was done due to the time-consuming nature of the Reverse Transcription Polymerase Chain Reaction (RT-PCR) tests and the false-negative rate. In this work, the chest CT scan images are classified and detected to COVID-19 and non-COVID-19 classes by using six automated DCNN architectures (VGG19, Inception-V3, Resnet-50, Inception-ResNet-V2, DenseNet121) and comparison between them by using some different activation functions. our results demonstrate that Inception-V3 and DenseNet-121 produce superior outcomes with other activation functions instead of ReLU.For this, we suggest using these methods to aid the physician in making resolution in clinical practice, especially in impoverished areas with limited availability of radiologists with sufficient training in COVID-19 imaging. Finally, we conclude that the model and the dataset's behavior relative to the model determine which activation function is optimal. Additionally, we intend to compare simple activation functions with complex, adaptive activation functions using the newest application-specific architectures in our next work.