A Study of U-Net-Based Architectures for Segmenting COVID-19 Lung Lesions

Early detection of lung lesions in patients with COVID-19 allows for better management and treatment, with Computed Tomography (CT) images playing a crucial role. This highlights the need to develop and refine tools that help medical experts identify interstitial lung anomalies such as Ground-Glass Opacities (GGOs) in these images. Using open datasets of CT images of patients with COVID-19, we studied four Deep Convolutional Neural Network architectures: U-Net, Attention U-Net, Recurrent Residual U-Net, and Attention Recurrent Residual U-Net in the context of lung lesion segmentation. We obtained a Dice-Sørensen coefficient score of up to 0.823 using techniques that include data augmentation, mixing datasets and limiting the area to segment using a previous lung segmentation. Overall, we demonstrated how certain deep learning architectures benefit greatly from extending their training dataset through data augmentation and mixing images from different sources.