Covid-19 Detection from CT-scan Images: Empirical Evaluation and Explainability

Abstract

Covid-19 has been a great disaster for the entire world. It is caused by the novel coronavirus, which is highly contagious. Detection of Covid-19 can be done either through saliva or through a CT scan. Given the scale at which this Covid-19 can spread, an automated detection is required which can be adopted at large scale. In this work, we focus on the detection of Covid-19 through CT scan images. Our work evaluates well-known CNN architecture-based models in different experimental settings: fine-tuning, removal of pre-trained layers, and data augmentation. For evaluation, we use the dataset of images comprising Covid-19 CT scans. We analyze the performance of VGG-16, InceptionNet, and ResNet. After rigorous experiments, the InceptionNet model performs the best with 0.99 AUC outperforming the prior work (which claimed 0.98 AUC), with the training accuracy and testing accuracy of 99.94% and 96.43%, respectively. Furthermore, we also perform explainability experiments on both Covid and Non-Covid CT-Scan images.