A Semi-Supervised Domain Adaptation Approach for Diagnosing SARS-CoV-2 and its Variants of Concern (VOC)

Since December 2019 the world has been facing an unprecedented crisis in handling the outbreak of the biological disaster COVID-19 putting a lot of pressure on the healthcare systems globally. Poor collection of swab samples, delayed testing, or variations that have possibly changed the disease patterns may be the reasons that have led to the increased false-negative results of Reverse Transcription Polymerase Chain Reaction (RT-PCR) tests recently which is considered the gold standard to detect Severe Acute Respiratory Syndrome Coronavirus-2 (SARS-CoV-2). At such times the doctors have to depend on X-rays or CT scans of the chest to establish the diagnosis. Artificial Intelligence (AI) based diagnosis through deep learning models for the classification of medical radiological images can play a very important role in the present scenario. However, for new infections like SARS-CoV-2 and its Variants of Concern (VOC) the labeled data may not be readily available for training a deep learning model while at the same time, the labeled images may be available for similar previously existing infections like viral or bacterial pneumonia. This study aims to propose a novel Semi-Supervised Domain Adaptation neural network, CoVSSDA to handle this limitation. CoVSSDA is an end-to-end deep convolutional neural network that is trained on the labeled images of the related previous infection with two classes {Normal, Pneumonia}, unlabeled images of the new infection with three classes {COVID19, Normal, Pneumonia} and a small batch of labeled images of the new infection such that the trained model acquires the knowledge about the novel class COVID19 and adapts to achieve an accuracy of 93.92% when tested for the new infection on the target domain and also outperforms other models based on domain adaptation.