Cascaded 3D UNet architecture for segmenting the COVID-19 infection from lung CT volume

Abstract

Abstract World Health Organization (WHO) declared COVID-19 (COronaVIrus Disease 2019) as pandemic on March 11, 2020. Ever since then, the virus is undergoing different mutations, with a high rate of dissemination. The diagnosis and prognosis of COVID-19 are critical in bringing the situation under control. COVID-19 virus replicates in the lungs after entering the upper respiratory system, causing pneumonia and mortality. Deep learning has a significant role in detecting infections from the Computed Tomography (CT). With the help of basic image processing techniques and deep learning, we have developed a two stage cascaded 3D UNet to segment the contaminated area from the lungs. The first 3D UNet extracts the lung parenchyma from the CT volume input after preprocessing and augmentation. Since the CT volume is small, we apply appropriate post-processing to the lung parenchyma and input these volumes into the second 3D UNet. The second 3D UNet extracts the infected 3D volumes. With this method, clinicians can input the complete CT volume of the patient and analyze the contaminated area without having to label the lung parenchyma for each new patient. For lung parenchyma segmentation, the proposed method obtained a sensitivity of 93.47%, specificity of 98.64%, an accuracy of 98.07%, and a dice score of 92.46%. We have achieved a sensitivity of 83.33%, a specificity of 99.84%, an accuracy of 99.20%, and a dice score of 82% for lung infection segmentation.