Automatic Segmentation of COVID-19 CT Images using improved MultiResUNet

Abstract

Corona Virus Disease 2019 (COVID-19) has seriously threatened human life and health in just a few months. The global economy, education, transportation and other aspects have been affected. In order to solve the problems caused by COVID-19 as soon as possible, it is important to quickly and accurately confirm whether people are infected. In this paper, we take MultiResUNet as the basic model, introduce a new "Residual block" structure in the encoder part, add Regularization and Dropout to prevent training overfitting, and change the partial activation function. Propose a model suitable for COVID-19 CT image sets, which can automatically segment four parts of COVID-19 CT images (leftright lung , disease and background) by deep learning. The segmentation results are evaluated and the expected results are achieved. It is helpful for medical workers to recognize the infection area quickly.