Using different deep learning methods for Covid-19 classification from CT scans segmented by generative adversarial networks and UNet

The new coronavirus, which emerged at the end of 2019, caused a major global health crisis in 7 continents. An essential step towards fighting this virus is limiting the spread of the disease through effective screening, early detection and isolation of infected persons. While computed tomography (CT) scans have become an effective method to detecting the diagnosis, radiological examination of CT has led to a large increase in the workload for radiologists. Computer Aided Detection (CAD) systems for Covid-19 from CT images can therefore reduce the burden on radiologists and improve efficiency of diagnosis. In this study, a new system based on UNet-based Conditional Generative Adversarial Network (cGAN) that automatically segments infected regions from chest CT slices is proposed. For the purpose of segmentation, the generator learns to create binary masks that outline boundaries around these infected regions while the discriminator learns to distinguish between the real and synthetic masks. This adversarial training therefore forces the generator to detect infected regions by creating very realistic binary masks. Experimental results show that the proposed method achieves a segmentation success with a dice score of 92.32% and IoU score of 86.41%. Furthermore, 3 classifiers which include a Convolutional Neural Network (CNN), a PatchCNN and a Capsule Neural Network (CapsNet) are proposed to classify the generated masks as either Covid-19 or not. Success of these classifiers was 99.20%, 92.30% and 73.83%, respectively. According to these results, the highest success was achieved in the system where cGAN_Unet and CNN are used together. Keywords: Covid-19 Segmentation, Covid-19 Classification, Generative Adversarial Network, Convolutional Neural Network, PatchCNN, Capsule Neural Network