PSACOVID: Parallel Segmentation Architecture for Localization and Contagion Quantification of Covid-19 Pneumonia in Chest Computed Tomography Imaging

Early and precise detection of Coronavirus related anomalies in lung computed tomography scans is necessary for implementing preventive measures and initiating proper treatment. However, the procedure of visually examining lung computed tomography (CT) images is demanding particularly when it comes to detecting tiny anomalies. This work intends to use artificial intelligence (AI) to help radiologists improve their illness diagnostic efficacy in reading computed tomography images by localizing, detecting, and quantifying COVID-19 pulmonary lesion severity to predict patient outcomes. A public dataset has been acquired from 20 Corona patients, which includes manually annotated lung and infection masks, to train a parallel quantification system (PSACOVID Net) that consists of two networks. The first one (UCGAN) is dedicated to identifying lung parenchyma, while the second one (ECGANCOVID) focuses on identifying areas affected by COVID-19 lesions to predict the disease severity stage. An additional data augmentation method in the preprocessing step is introduced in the proposed method to improve the diagnostic performance and provide a way to accurately quantify the infected area of the lung.Consequently, our framework has outperformed various baseline methods by achieving the highest sensitivity of 98.80% in COVID-19 detection performance and Enhance- Mean Absolute Error of 0.53% in quantifying infected lung regions.