DUDA-Net: a double U-shaped dilated attention network for automatic infection area segmentation in COVID-19 lung CT images

Xie, Feng; Huang, Zheng; Shi, Zhengjin; Wang, Tianyu; Song, Guoli; Wang, Bolun; Liu, Zihong

PURPOSE: The global health crisis caused by coronavirus disease 2019 (COVID-19) is a common threat facing all humankind. In the process of diagnosing COVID-19 and treating patients, automatic COVID-19 lesion segmentation from computed tomography images helps doctors and patients intuitively understand lung infection. To effectively quantify lung infections, a convolutional neural network for automatic lung infection segmentation based on deep learning is proposed. METHOD: This new type of COVID-19 lesion segmentation network is based on a U-Net backbone. First, a coarse segmentation network is constructed to extract the lung areas. Second, in the encoding and decoding process of the fine segmentation network, a new soft attention mechanism, namely the dilated convolutional attention (DCA) mechanism, is introduced to enable the network to focus on better quantitative information to strengthen the network's segmentation ability in the subtle areas of the lesions. RESULTS: The experimental results show that the average Dice similarity coefficient (DSC), sensitivity (SEN), specificity (SPE) and area under the curve of DUDA-Net are 87.06%, 90.85%, 99.59% and 0.965, respectively. In addition, the introduction of a cascade U-shaped network scheme and DCA mechanism can improve the DSC by 24.46% and 14.33%, respectively. CONCLUSION: The proposed DUDA-Net approach can automatically segment COVID-19 lesions with excellent performance, which indicates that the proposed method is of great clinical significance. In addition, the introduction of a coarse segmentation network and DCA mechanism can improve the COVID-19 segmentation performance.