基于多尺度改进的V-Net肺结节分割方法研究  

Research on Segmentation Method of Lung Nodules Based on Multi-scale Improved V-Net

作　　者：门靖茹 王泽荣 张富春[1] 白宗文[1] MEN Jingru;WANG Zerong;ZHANG Fuchun;BAI Zongwen(School of Physics and Electronic Information,Yan’an University,Yan’an 716000,China)

机构地区：[1]延安大学物理与电子信息学院,陕西延安716000

出　　处：《延安大学学报（自然科学版）》2022年第1期115-120,共6页Journal of Yan'an University：Natural Science Edition

基　　金：国家自然科学基金项目(61761042);延安市信息处理与测控技术科技创新团队项目(2017CXTD-01);延安大学研究生教育教学改革研究项目(YDYJG2019016)。

摘　　要：近年来基于深度模型分割已成为肺结节分割的主要方法,但多数深度模型的精度与轻量性难以共存,且大模型不利于方便部署。为了得到一种轻量级且尽可能不损失精度的模型,提出了一种基于M-VNet的肺结节分割方法。该网络总体设计继承V-Net结构,并添加了不同深度路线平衡细节信息和语义信息,使用路线注意力机制进行高效融合。M-Block组件设计将残差信息纳入卷积计算,在有效缩小模型的同时保留模型的优异分割性能。研究结果显示,M-VNet在参数量仅为V-Net的13%的情况下,骰子系数较V-Net提高4%。使用LIDC-IDRI肺结节公开数据集对基线模型和改进模型进行性能评估,结果表明M-VNet的性能优异,对不同形态的肺结节分割效果良好且性能稳定。该方法在肺结节分割和提高诊断速度、准确率方面具有一定的临床应用价值。In recent years,segmentation based on deep models has become the main method for segmentation of lung nodules,but the accuracy and lightness of most deep models are difficult to coexist,and large models are not conducive to easy deployment.In order to obtain a lightweight model without loss of accuracy as much as possible,this paper proposes a lung nodule segmentation method based on M-VNet.The overall design of the network inherits the V-Net structure,and adds different depth routes to balance detailed information and semantic information,and uses the route attention mechanism for efficient integration.The M-Block design incorporates the residual information into the convolution calculation,which effectively reduces the model size while retaining the excellent segmentation performance of the model.The research results show that when the parameter amount of M-VNet is only 13%of V-Net,the dice coefficient is 4%higher than that of V-Net.The paper uses the LIDC-IDRI lung nodule public data set to evaluate the performance of the baseline and the improved model.The results show that the performance of M-VNet is excellent,and has a good segmentation effect and stable performance for different forms of lung nodules.This method has very important clinical application value in segmenting lung nodules and improving the speed and accuracy of diagnosis.

关 键 词：肺结节分割 深度学习 多尺度特征 V-Net 路线注意力 

分 类 号：R318[医药卫生—生物医学工程]