基于多尺度特征融合与反向注意力的COVID-19病灶分割  

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

作　　者：李碧草 王晶 郭旭伟[3] 黄杰[1] 魏苗苗 李盼盼 LI Bicao;WANG Jing;GUO Xuwei;HUANG Jie;WEI Miaomiao;LI Panpan(School of Electronic and Information,Zhongyuan University of Technology,Zhengzhou 450007,China;School of Information Engineering,Zhengzhou University,Zhengzhou 450001,China;Department of Pediatrics,the First Affiliated Hospital of Henan University of Science and Technology,Luoyang 471000,China)

机构地区：[1]中原工学院电子信息学院,河南郑州450007 [2]郑州大学信息工程学院,河南郑州450001 [3]河南科技大学第一附属医院儿科,河南洛阳471000

出　　处：《中国医学物理学杂志》2023年第4期403-409,共7页Chinese Journal of Medical Physics

基　　金：国家自然科学基金(61901537);河南省高校科技创新人才支持计划(23HASTIT030);河南省留学人员科研择优项目;中国博士后科学基金(2020M672274);中原工学院学科实力提升青年硕导培育计划(SD202207)。

摘　　要：针对新型冠状病毒肺炎(COVID-19)分割问题中感染区域具有高变异性以及病灶与背景对比度低等问题,提出一种基于多尺度特征融合与反向注意力的COVID-19感染分割网络。首先,利用残差网络作为主干网络进行特征提取,并使用全局上下文聚合策略对不同层次特征进行融合得到粗略的分割结果;其次,在网络瓶颈处添加多尺度特征融合模块,利用空洞卷积与多核池化增强网络分割不同尺度病变的能力;最后,设计一种级联结构的反向注意力模块,利用互补区域的细节特征增强背景与目标的对比度。本文方法在COVID-19 CT分割测试集上的准确率、特异性、灵敏度分别达到0.714、0.700和0.958,误检和漏检区域明显减少,细小病灶的分割能力显著提升。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 multiscale 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.

关 键 词：COVID-19 肺炎感染分割 全局上下文聚合 多尺度特征融合 反向注意力 

分 类 号：R318[医药卫生—生物医学工程] TP391[医药卫生—基础医学]