SD-UNet: A Novel Segmentation Framework for CT Images of Lung Infections

• DOI: 10.3390/electronics11010130
• Publication Date: 2022-01-04T03:51:50Z
• Authors: Shuangcai Yin; Hongmin Deng; Zelin Xu; Qilin Zhu; Junfeng Cheng
• Keywords: Pooling; Pyramid (geometry)

Due to the outbreak of lung infections caused by coronavirus disease (COVID-19), humans have face an unprecedented and devastating global health crisis. Since chest computed tomography (CT) images COVID-19 patients contain abundant pathological features closely related this disease, rapid detection diagnosis based on CT is great significance for treatment blocking spread disease. In particular, segmentation lung-infected area can quantify evaluate severity However, due blurred boundaries low contrast between infected non-infected areas in images, manual lesion laborious places high demands operator. Quick accurate lesions from deep learning has drawn increasing attention. To effectively improve effect infection, a modified UNet network that combines squeeze-and-attention (SA) dense atrous spatial pyramid pooling (Dense ASPP) modules) (SD-UNet) proposed, fusing context multi-scale information. Specifically, SA module introduced strengthen attention pixel grouping fully exploit information, allowing better mine differences connections pixels. The Dense ASPP utilized capture information lesions. Moreover, eliminate interference background noise outside lungs highlight texture area, we extract advance pre-processing stage. Finally, our method using binary-class multi-class infection datasets. experimental results show metrics Sensitivity, Dice Similarity Coefficient, Accuracy, Specificity, Jaccard are 0.8988 (0.6169), 0.8696 (0.5936), 0.9906 (0.9821), 0.9932 (0.9907), 0.7702 (0.4788), respectively, (multi-class) task proposed SD-UNet. result segmented SD-UNet closer ground truth compared several existing models such as CE-Net, DeepLab v3+, UNet++, other models, which further proves more be achieved method. It potential assist doctors making quantitative assessment COVID-19.