A demystifying convolutional neural networks using Grad-CAM for prediction of coronavirus disease (COVID-19) on X-ray images

The 2019 novel coronavirus (COVID-19) has spread quickly among people living in different countries and is impending 26,27,630 cases worldwide according to the statistics of European Center for Disease Prevention and Control. To control the spread of COVID-19, testing large numbers of alleged cases for proper quarantine and treatment is of at most important. Because of the rapid spread of the virus among people, there is limited number of testing kits available in hospitals. Since, the doctors cannot depend only on these kits, it is necessary to find an alternate way to prevent this pandemic disease. To overcome this situation, it is necessary to implement the recognition of quick substitute diagnosis options for the prevention of COVID-19 among the people. In the proposed work, the main concentration is made by considering the COVID-19 chest X-ray images and the normal chest X-ray images to build the customized deep learning classification model. Then the images obtained from the data sources both from COVID-19 chest X-ray images and from normal chest X-ray images are combined together. Roughly 20% of the images were randomly selected and used for validation, and the remaining 80% of images were used for training purpose. The dataset was carefully screened to have only relevant chest X-ray images removing images of other type or the images which lacked resolution to get the final dataset. To obtain a better accuracy, a customized convolutional neural network architecture was built for this purpose. The model was compiled with Adam optimizer along with binary cross-entropy as a loss function. Image data augmentation such as zoom, shear, normalization, and horizontal flip was used to negate the effects of using a small dataset. The highest validation accuracy obtained after a series of epochs was 98% along with a training accuracy of 96%.