RGU-Net: Residual Guided U-Net Architecture for Automated Segmentation of COVID-19 Anomalies Using CT Images

The outbreak of novel coronavirus disease(COVID-19) has spread worldwide as a severe respiratory disease. The public’s health as well as the global financial system are also in serious jeopardy. Computer-aided fast and automated diagnosis toolkits are required to reduce the prevalence of this deadly virus. However, one of the main challenges is to differentiate between COVID-19 and typical pneumonia cases, as most clinical manifestations are very similar. So, it is vital to detect the region of infections caused by COVID-19 rather than the mere disease prediction. In this study, a residual guided deep learning-based segmentation framework named ‘RGU-Net’ was proposed to quantify COVID-19 cases accurately using chest CT slices. However, deeper networks often suffer from vanishing-gradient problems. A stack of residual guided skip-connections was utilized with the ‘U-Net’ architecture to develop the proposed ‘RGU-Net’. The ‘RGU-Net’ was intended to retain the spatial and location information as well as alleviate the vanishing-gradient issue. A modified segmentation loss called ‘Dice Focal’ loss was also introduced in this study to reduce the impact of noisy ground truth labels and class imbalance issue. The segmentation results demonstrate the outstanding performance of the proposed framework compared to the recent competing segmentation models. It produced a mean dice score of 90.23% and a mean IoU of 93.19%.