Attention-Based Residual Learning Network for COVID-19 Detection Using Chest CT Images

COVID-19 is a deadly viral infection that is highly contagious and has brought a great loss of human lives and economic resources. Hence, it is very critical to detect the virus at an early stage with high accuracy. Deep learning algorithms are very effective in learning the discriminative features of medical images. It can facilitate the rapid diagnosis of the disease. In this research, different deep learning architectures have been evaluated on a balanced CT image dataset for COVID-19. From the analysis performed, this research proposes a deep learning architecture that performs substantially better than the reviewed models. This research aims to analyze and identify an effective baseline deep architecture on which a model can be built for advanced COVID-19 detection. The architectures analyzed in this research include, AlexNet, DenseNet, GoogLeNet, InceptionV4, ResNet, ShuffleNet, SqueezeNet, and Visual Geometric Group (VGG16). Experimental results indicated that GoogLeNet based network was able to detect COVID-19 with an accuracy of 83.27%. This backbone model was further modified to design a better performing network exclusively for COVID-19. This work proposes an attention-based residual learning block that is integrated with the GoogLeNet backbone. The proposed model performed obtained an accuracy of 91.26%.