Abstract:

Abstract The coronavirus 2019 (COVID-19) pandemic carrying on to seriously affect the health and well-being of the world-wide public. Effectively screening those who are infected with COVID-19 is the first step in the battle against it, and having examinations of chest X-rays. The research aimed to evolve a deep-learning method for the early identification of pneumonia and COVID-19 lung disease using chest X-rays. In this paper, a deep learning method has been proposed using an improved 53-layer residual network model. The COVIDx dataset with 13975 CXR images and the Kermany [17] dataset with 5856 CXR images has been used to evaluate the proposed models. In these image collections, a 4:1 aspect ratio is used for training and testing. The experimental results are performed using Python and results are compared and analyzed with pre-trained models such as GoogLeNet, ResNet50, and DenseNet121.The proposed model outperforms the most sophisticated models on the Kermany dataset with accuracy of 97.9%, sensitivity of 98.1%, specificity of 97.6%, and precision of 97%. The proposed model performance on the COVIDx dataset is 97.1%, 98.9%, 95.7% and 94.5% for accuracy, sensitivity, specificity and precision, respectively. Apart from that, we have also incorporated three more layers to ResNet50, creating it a ResNet50+3 layer design which resolves the vanishing gradient issue and, makes training easier. The result of the whole analysis shows that the proposed model not only outperforms most classifiers but is also a very generic system that is adjustable to a various healthcare datasets.

Keywords: