基于多尺度级联注意网络的肺实质分割
Lung parenchymal segmentation based on multiscale cascaded attention network
许圳兴 1余耀 2赵东 2陈园 3范圣旺3
作者信息
- 1. 南京信息工程大学电子与信息工程学院 南京 210044
- 2. 无锡学院电子与信息工程学院 无锡 214105
- 3. 杭州海康威视数字技术股份有限公司 上海 200023
- 折叠
摘要
针对肺实质分割任务中不同尺度特征的全局上下文信息利用率低、分割精度低、分割细节模糊等问题,提出一种多尺度级联注意网络(multiscale cascaded attention networks,MCANet).该网络主要由多尺度特征提取网络(multi-scale feature extraction network,MSFENet)、多尺度注意力引导模块(multi-scale attention guidance module,MSAG)、解码特征整合器(decoding feature integrator,DFI)组成.首先,设计 MSFENet以提高特征信息在不同通道维度上的空间交互能力,在采样过程中最大限度地保留图像的关键特征,丰富全局上下文信息.然后,设计 MSAG提高模型在解码过程中对多尺度特征信息的利用率,并最大限度地融合两种注意力机制的优势.最后设计DFI,重新整合解码器生成的解码特征,以提高模型对边缘信息的分割性能.在LUNA16数据集上对模型性能进行实验验证,得到了0.993的Dice和3.864的HD,实验结果证明了 MCA-Net与其他主流医学分割模型相比有更优异的分割性能,能更准确地分割肺实质.
Abstract
A multiscale cascaded attention network(MCANet)is proposed to solve the problems of low utilization rate of global context information,low segmentation accuracy and fuzzy segmentation details of different scale features in lung parenchymal segmentation task.The network is mainly composed of a multi-scale feature extraction network(MSFENet),a multi-scale attention guidance(MSAG)module,and a decoding feature integrator(DFI).Firstly,the MSFENet network is designed to improve the spatial interaction ability of feature information in different channel dimensions,retain the key features of the image to the greatest extent in the sampling process,and enrich the global context information.Then,MSAG is designed to improve the utilization of multi-scale feature information in the decoding process of the model,and maximize the integration of the advantages of the two attention mechanisms.Finally,DFI was designed to reintegrate the decoding features generated by the decoder to improve the segmentation performance of the model for edge information.In this paper,the performance of the model was verified by experiments on the LUNA16 dataset,and the Dice of 0.993 and HD of 3.864 were obtained,and the experimental results proved that MCANet has better segmentation performance and can segment the lung parenchyma more accurately than other mainstream medical segmentation models.
关键词
肺实质分割/多尺度级联注意网络/多尺度特征提取网络/多尺度注意力引导模块/解码特征整合器Key words
lung parenchymal segmentation/MCANet/MSFENet/MSAG/DFI引用本文复制引用
基金项目
江苏省高等学校基础科学(自然科学)研究面上省资助经费项目(22KJB140017)
无锡学院人才启动基金(2019r015)
无锡学院人才启动基金(2022r006)
出版年
2024
NETL
NSTL
万方数据