Residual Dense U-Net for Segmentation of Lung CT Images Infected with Covid-19

The novel coronavirus disease 2019 (Covid-19) has been declared as a pandemic by the World Health Organization which in the current global scenario has brought everything from economy to education to a halt. Due to its rapid spread around the globe, even the most developed countries are facing difficulties in diagnosing Covid-19. For efficient treatment and quarantining of the exposed population it is important to analyse Lung CT Scans of the suspected Covid-19 patients. Computer aided segmentation of the suspicious Region of Interest can be used for better characterization of infected regions in Lung. In this work a deep learning-based U-Net architecture is proposed as a framework for automated segmentation of multiple suspicious regions in a CT scan of Covid-19 patient. Advantage of Dense Residual Connections has been taken to learn the global hierarchical features from all convolution’s layers. So, a better trade-off in between efficiency and effectiveness in a U-Net can be maintained. To train the proposed U-Net system, publicly available data of Covid-19 CT scans and masks consisting of 838 CT slices has been used. The proposed method achieved an accurate and rapid segmentation with 97.2%, 99.1% and 99.3% as dice score, sensitivity and specificity respectively.