Conditional cascaded network (CCN) approach for diagnosis of COVID-19 in chest X-ray and CT images using transfer learning

Abstract

The COVID-19 pandemic has caused substantial global health and economic damage, with over five million confirmed cases worldwide. The importance of the rapid, accurate diagnosis of infected patients has been underscored. However, due to the shortage of testing kits and the time-consuming, troublesome nature of the manual RT-PCR test, an automated, efficient diagnosis system using chest medical images for the early screening of COVID-19 is crucial. In this paper, an automated approach for the rapid, accurate diagnosis of COVID-19 in chest X-ray and computed tomography (CT) images using transfer learning is presented. The proposed technique leverages the conditional cascaded network approach, which employs multiple levels of networks to analyze images with high confidence. Transfer learning is employed with seven commonly used existing convolutional neural network architectures and four datasets for X-ray and CT images. Various regularization, optimization, dropout, and data augmentation techniques are examined through a series of experiments with three optimizers. Our technique is compared with other state-of-the-art techniques, and the achieved results demonstrate highly promising performance metrics. Additionally, occlusion specificity and gradient-weighted class activation mapping techniques are employed to understand the network output better. The proposed technique is highly adaptable and scalable and does not require manual hyperparameter tuning.