CoroTrans-CL: A Novel Transformer-Based Continual Deep Learning Model for Image Recognition of Coronavirus Infections

The rapid evolution of coronaviruses in respiratory diseases, including severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), poses a significant challenge for deep learning models to accurately detect and adapt to new strains. To address this challenge, we propose a novel Continuous Learning approach, CoroTrans-CL, for the diagnosis and prevention of various coronavirus infections that cause severe respiratory diseases using chest radiography images. Our approach is based on the Swin Transformer architecture and uses a combination of the Elastic Weight Consolidation (EWC) and Herding Selection Replay (HSR) methods to mitigate the problem of catastrophic forgetting. We constructed an informative benchmark dataset containing multiple strains of coronaviruses and present the proposed approach in five successive learning stages representing the epidemic timeline of different coronaviruses (SARS, MERS, wild-type SARS-CoV-2, and the Omicron and Delta variants of SARS-CoV-2) in the real world. Our experiments showed that the proposed CoroTrans-CL model achieved a joint training accuracy of 95.34%, an F1 score of 92%, and an average accuracy of 83.40% while maintaining a balance between plasticity and stability. Our study demonstrates that CoroTrans-CL can accurately diagnose and detect the changes caused by new mutant viral strains in the lungs without forgetting existing strains, and it provides an effective solution for the ongoing diagnosis of mutant SARS-CoV-2 virus infections.