Modified U-Net Based Covid-19 Lesion Segmentation Using CT Scans

Abstract

Computed Tomography (CT) based analysis will assist doctors in a prompt diagnosis of the Covid-19 infection. Automated segmentation of lesions in chest CT scans helps in determining the severity of the infection. The presented work addresses the task of automated segmentation of Covid-19 lesions. A U-Net framework incorporated with spatial-channel attention modules (contextual relationships), Atrous Spatial Pyramid Pooling module (a wider receptive field) and Deep Supervision (lesion focus, less error propagation) is proposed. Focal Tversky Loss is used to evaluate the outputs at coarser scales while Tversky loss evaluates the final segmentation output. This combination of losses is used to enhance segmentation of the small lesions. The framework is trained on CT scans of 20 subjects of COVID19 CT Lung and Infection Segmentation Dataset and tested on Mosmed dataset of 50 subjects, where infection has affected less than 25% of lung parenchyma. The experimental results show that the proposed method is effective in segmenting the hard ROIs in Mosmed data resulting in a mean Dice score of 0.57 (9% more than the state-of-the-art).