An Efficient Deep Learning Framework of COVID-19 CT Scans Using Contrastive Learning and Ensemble Strategy

Abstract

Since the outbreak of COVID-19 in 2019, more than 200 million individuals have been infected worldwide, resulting in over four million deaths. Although large-scale nucleic acid test is an effective way to diagnose COVID-19, the possibility of false positives or false negatives means that the chest CT scan remains a necessary tool in COVID-19 diagnosis for cross-validation. A lot of research has been carried out using deep learning methods for COVID-19 diagnosis using CT scans. However, privacy concerns result in very limited datasets being publicly available. In this research, we propose a novel framework based on the centripetal contrastive learning of visual representations (CeCLR) method with stacking ensemble learning to represent features more efficiently so as to achieve better performance on a limited COVID-19 dataset. Experimental results demonstrate that our deep learning system is superior to other baseline models. Our method achieves an F1 score of 0.914, AUC of 0.952, and accuracy of 0.909 when diagnosing COVID-19 on CT scans.