U-Net-based COVID-19 CT Image Semantic Segmentation: A Transfer Learning Approach

ABSTRACT

Deep learning (DL) algorithms are widely applied in many disciplines such as medical imaging, bioinformatics, and computer vision. DL models have been used in medical imaging to perform image segmentation, classification, and detection. During the outbreak of the COVID-19 pandemic, DL has been extensively used to develop COVID-19 screening systems. The reverse transcription-polymerase chain reaction (RT-PCR) is the gold standard method for COVID-19 screening. However, DL has been proposed to detect patients infected with COVID-19 through radiological imaging in Chest X-rays and chest computed tomography (CT) images. This paper proposes transfer learning to train modified U-Net models to segment the COVID-19 chest CT images into two regions of lung infection (ground-glass and consolidation). The proposed modified U-Net models were constructed by replacing the encoder part with a pre-trained convolutional neural network (CNN) model. Three pre-trained CNN models, namely, EfficientNet-b0, EfficientNet-b1, and EfficientNet-b2 were used. The proposed models were evaluated on the COVID-19 CT Images Segmentation dataset available in an open Kaggle challenge. The obtained results show that the proposed EfficientNet-b2_U-Net model yielded the highest FScore of 0.5666. © 2022 IEEE.