基于RSE-Vnet卷积网络的肺结节分割方法研究  

Research on Lung Nodule Segmentation Method Based on RSE-VnetConvolutional Network

作　　者：闫永强 秦斌[2] YAN Yongqiang;QIN Bin(School of Computer,Hunan University of Technology,Zhuzhou Hunan 412007,China;School of Transportation and Electrical Engineering,Hunan University of Technology,Zhuzhou Hunan 412007,China)

机构地区：[1]湖南工业大学计算机学院,湖南株洲412007 [2]湖南工业大学交通与电气工程学院,湖南株洲412007

出　　处：《湖南工业大学学报》2025年第5期46-51,共6页Journal of Hunan University of Technology

基　　金：湖南省自然科学基金资助项目(2023JJ50166)。

摘　　要：针对在细粒度图像的分割任务中容易出现欠分割与漏检的问题,提出一种改进的端到端3D分割算法——RSE-Vnet。加入Res2net网络捕获不同结节的多尺度细粒特征,为网络馈送更多精准的结节位置信息;同时残差连接避免了网络退化问题,建立了结节数据驱动模型;注意力机制能够有效为重要特征通道自适应加权,减少背景图像的干扰。构建了的方法在一定程度上解决了多类型结节欠分割和漏检问题,最终在LUNA16数据集中得以验证,模型DSC提升了7%,检测灵敏度提升了6%。In view of the flaws of under-segmentation and missed detection in fine-grained image segmentation tasks,an improved end-to-end 3D segmentation algorithm,RSE-Vnet,has thus been proposed.With Res2net network incorporated,multi-scale fine-grained features of different nodules can be captured,thus feeding more accurate nodule location information to the network.Meanwhile,residual connections help to avoid network degradation issues,thus establishing a data-driven model for nodules.The attention mechanism can effectively weight important feature channels so as to reduce the interference of background images,with the constructed method solving the problem of under-segmentation and missed detection of multiple types of nodules to some extent.Finally,it can be verified in the LUNA16 dataset,with a 7%increase in model DSC and a 6%increase in detection sensitivity specifically.

关 键 词：计算机辅助诊断 多尺度细粒特征 注意力 Res2net网络 多类型结节 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]