Segmentation of COVID-19 Infections on CT: Comparison of Four UNet-Based Networks

Abstract

Diagnosis and staging of COVID-19 are crucial for optimal management of the disease. To this end, novel image analysis methods need to be developed to assist radiologists with the detection and quantification of the COVID-19-related lung infections. In this work, we develop and evaluate four Artificial intelligence (AI) based lesion segmentation and quantification methods from chest CT, using U-Net, Attention U-Net, R2U-Net, and Attention R2U-Net models. These models are trained and evaluated using a dataset consisting of 8739 CT images of the lungs from 147 healthy subjects and 150 patients infected by COVID-19. The results show that the Attention R2U-Net model is superior to the others with a Dice value of 0.79. The lesion volumes estimated by the Attention R2U-Net model are highly correlated with those of the manual segmentations by an expert, with a correlation coefficient of 0.96.