SAARSNet: A Deep Neural Network for COVID-19 Cases Diagnosis

Abstract

The global spread of the COVID-19 is a continuously evolving situation and it is still a major risk on the health of people around the world. A huge number of people are infected by this deadly virus and the number is still getting increased day by day. At this time, no specific vaccines or treatments of COVID-19 are found. Numerous ways are offered to detect COVID-19 such as swab test, CDC and RT-PCR tests. All of them can detect corona virus in different ways but they are not recommended by the reason of their limited availability, inaccurate results, high false-negative rate predicates, high cost and time consuming. Hence, medical radiography and Computer Tomography (CT) images were suggested as the next best alternative of RT -PCR and other tests for detecting Covid-19 cases. Recent studies found that patients with COVID-19 cases are present abnormalities in chest X-Ray images. Motivated by this, many researchers propose deep learning systems for COVID-19 detection. Although, these developed AI systems have shown quite promising results in terms of accuracy, they are closed source and unavailable to the research community. Therefore, in the present work, we introduced a deep convolutional neural network design (SAARSNet) designed to detect COVID-19 cases from chest X-Ray images. 1292 X-Ray images have been used to train and test the proposed model. the images have been collected from two open-source datasets. The input images are progressively resized into (220 by 150 by 3) in order to decrease the training time of the system and improve the performance of the SAARSNet architecture. Furthermore, we also investigate how SAARSNet makes predictions under three different scenarios with the aim of distinguishing COVID-19 class from both Normal and Abnormal classes as well as gaining deeper perceptions into critical factors related to COVID-19 cases. We also used the confusion metrics for evaluating the performance of SAARSNet CNN in an attempt to measure the true and false identifications of the classes from the tested images. With the proposed architecture promising results has been achieved in all of the three different scenarios. Although, there are some misclassified cases of COVID-19, the corresponding performance was best in detecting both Normal and Abnormal cases correctly. Furthermore, in the three classes scenario, normal class has been achieved 100% positive predictive value while optimistic results have been investigated in detecting COVID-19 and abnormal classes.