An automatic COVID-19 CT segmentation based on Progressive encoder and decoder U-Net++ with attention mechanism

Abstract

The coronavirus disease (COVID-19) pandemic has contribute to a harsh effect on the global public health. Computed Tomography (CT) is an effective tool in the screening of COVID-19. It is greater important to rapidly and accurately segment COVID-19 from CT to help diagnostic and monitor patients. In this paper, we propose a Progressive encoder and decoder U-Net++ based segmentation network using attention mechanism. In terms of COVID-19 lesion segmentation problems with highly imbalanced dataset and small regions of interests (ROI), we will use a progressive encoder and decoder combined with dilated convolution to form a deeper network structure, which can extract more and lower level semantic features while ensuring spatial information features. We propose to incorporate an attention mechanism to a progressive encoder and decoder U-Net++ architecture to capture rich contextual relationships for better feature representations. Meanwhile, the focal tversky loss is enhanced to address the small lesion segmentation. In addition, after combining the advantages of multiple modules, the network parameters will increase abruptly. According to the performance of the model in the validation set, we cut the redundant branch of the network model to do the final segmentation test, which can not only reduce the segmentation accuracy, but also reduce the network parameters and calculation cost. The experiment results, evaluated on a small dataset where only 3520 CT images are available, prove the enhanced model can achieve an accurate result on COVID-19 segmentation. The obtained Dice Score, Sensitivity and Specificity are 70.1%, 82.1%, and 92.3%, respectively.