用于肺部病灶图像分割的多尺度稠密融合网络  

Multiscale dense fusion network for lung lesion image segmentation

作　　者：卢小燕 徐杨 袁文昊 LU Xiaoyan;XU Yang;YUAN Wenhao(College of Big Data and Information Engineering,Guizhou University,Guiyang Guizhou 550025,China;Guiyang Aluminum-Magnesium Design and Research Institute Company Limited,Guiyang Guizhou 550009,China)

机构地区：[1]贵州大学大数据与信息工程学院,贵阳550025 [2]贵阳铝镁设计研究院有限公司,贵阳550009

出　　处：《计算机应用》2023年第10期3282-3289,共8页journal of Computer Applications

基　　金：贵州省科技计划项目(黔科合支撑[2021]一般176)。

摘　　要：针对主流的深度学习网络难以完整分割肺部病灶、区域边界预测模糊的问题,提出一种基于U-Net的多尺度稠密融合网络(MDF-Net)。首先,引入多分支密集跳层连接以捕获多级上下文信息,并在网络末端引入信息加权融合(IWF)模块进行逐级融合,以解决网络中的特征损失问题;其次,设计一种自注意力金字塔模块,使用各金字塔层对特征图进行不同规模的切分处理,并使用自注意力机制计算像素关联度,从而增强局部与全局区域的感染特征显著性;最后,设计一种区别于传统U-Net的上采样模式的上采样残差(UR)模块,多分支的残差结构与通道特征激励使网络能够还原更加丰富的微小病灶特征。在两个公开数据集上的实验结果显示,与UNeXt相比,所提网络的准确度(ACC)分别提升了1.5%和1.4%,平均交并比(MIoU)分别提升了3.9%和1.9%,实验结果验证了MDF-Net具有更好的肺部病灶分割性能。Aiming at the problems of incomplete segmentation of lung lesions and fuzzy prediction of regional boundaries in mainstream deep learning networks,a Multiscale Dense Fusion Network(MDF-Net)based on U-Net was proposed.Firstly,multi-branch dense skip connections were introduced to capture multi-level contextual information,and Information Weighted Fusion(IWF)module was introduced at the end of the network for level-by-level fusion to solve the feature loss problem in the network.Secondly,a self-attention pyramid module was designed.Each pyramid layer was used to segment the feature map in different scales,and the self-attention mechanism was applied to calculate the pixel correlation,thereby enhancing the saliency of the infection features in local and global regions.Finally,unlike the up-sampling form in traditional U-Net,a Up-sampling Residual(UR)module was designed.The multi-branch residual structure and channel feature excitation were used to help the network restore more abundant features of micro lesions.Experimental results on two public datasets show that compared with UNeXt,the proposed network improves the ACCuracy(ACC)by 1.5%and 1.4%respectively,and the Mean Intersection over Union(MIoU)by 3.9%and 1.9%respectively,which verify that MDF-Net has better lung lesion segmentation performance.

关 键 词：肺部疾病 密集跳层连接 自注意力金字塔 上采样残差 信息加权融合 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391[自动化与计算机技术—控制科学与工程]