Abstract

Info

Title:: COVID-19 lesion segmentation based on multi-scale feature fusion and reverse attention

Author(s):: LI Bicao1; 2; WANG Jing1; GUO Xuwei3; HUANG Jie1; WEI Miaomiao1; LI Panpan1; 1. School of Electronic and Information, Zhongyuan University of Technology, Zhengzhou 450007, China 2. School of Information Engineering, Zhengzhou University, Zhengzhou 450001, China 3. Department of Pediatrics, the First Affiliated Hospital of Henan University of Science and Technology, Luoyang 471000, China

Keywords:: Keywords: COVID-19 lung lesion segmentation global context aggregation multi-scale feature fusion reverse attention

Abstract:: Abstract: A COVID-19 lesion segmentation network based on multi-scale feature fusion and reverse attention (MFFRA) is proposed to overcome the problems of high variability and low contrast between lesion and background in COVID-19 segmentation. The residual network is used as the backbone network to extract features, and the global context aggregation strategy is adopted to integrate different hierarchical features for obtaining rough segmentation results. In addition, the multi-scale feature fusion module is added at the bottleneck of the network to enable the ability to segment lesions of different scales using atrous convolutions and multi-kernel pooling. Finally, a novel cascaded reverse attention module is designed to improve the low contrast between normal tissue and lesions based on the detailed features of complementary regions. The proposed method has an accuracy, specificity and sensitivity of 0.714, 0.700, 0.958 on the COVID-19 CT test set, reduces the areas of misdetection and missed detection, and enhances the segmentation ability of fine lesions.