A Soft-Voting Ensemble Classifier for Detecting Patients Affected by COVID-19

COVID-19 is an ongoing global pandemic of coronavirus disease 2019, which may cause severe acute respiratory syndrome. This disease highlighted the limitations of health systems worldwide regarding managing the pandemic. In particular, the lack of diagnostic tests that can quickly and reliably detect infected patients has contributed to the spread of the virus. Reverse Transcriptase-Polymerase Chain Reaction (RT-PCR) and antigen tests, which are the main diagnostic tests for COVID-19, showed their limitations during the pandemic. In fact, RT-PCR requires several hours to provide a diagnosis and is not properly accurate, thus generating a high number of false negatives. Unlike RT-PCR, antigen tests provide rapid diagnosis but are less accurate in detecting COVID-19 positive patients. Medical imaging is an alternative diagnostic test for COVID-19. In particular, chest computed tomography allows detecting lung infections related to the disease with high accuracy. However, visual analysis of a chest scan generated by computed tomography is a demanding activity for radiologists, making widespread use of this test unfeasible. Therefore, it is essential to lighten their work with automated tools able to provide accurate diagnosis in a short time. To deal with this challenge, in this work, an approach based on 3D Inception CNNs is proposed. Specifically, 3D Inception-V1 and Inception-V3 models have been built and compared. Then, soft-voting ensemble classifier models have been separately built on these models to boost the performance. As for the individual models, results showed that Inception-V1 outperformed Inception-V3 according to different measures. As for the ensemble classifier models, the outcome of experiments pointed out that the adopted voting strategy boosted the performance of individual models. The best results have been achieved enforcing soft voting on Inception-V1 models.