基于改进残差和注意力的CT肺癌辅助诊断  

Lung cancer CT-assisted diagnosis based on improved residual and attention

作　　者：谷宇[1] 迟靖千 张宝华[1] 杨立东[1] 李建军[1] 唐思源[1,2] GU Yu;CHI Jingqian;ZHANG Baohua;YANG Lidong;LI Jianjun;TANG Siyuan(Key Laboratory of Pattern Recognition and Intelligent Image Processing of Inner Mongolia Autonomous Region,School of Digital and Intelligent Industry,Inner Mongolia University of Science and Technology,Baotou 014010,China;School of Computer Science and Technology,Baotou Medical College,Inner Mongolia University of Science and Technology,Baotou 014040,China)

机构地区：[1]内蒙古科技大学数智产业学院内蒙古自治区模式识别与智能图像处理重点实验室,内蒙古包头014010 [2]内蒙古科技大学包头医学院计算机科学与技术学院,内蒙古包头014040

出　　处：《传感器与微系统》2024年第9期30-34,共5页Transducer and Microsystem Technologies

基　　金：国家自然科学基金资助项目(62001255,61962046,62161040,62262048,62066036,61841204);中央引导地方科技发展资金资助项目(2021ZY0004);内蒙古自治区自然科学基金资助项目(2024MS06008,2024LHMS06006,2019MS06003,2022MS06017,2020MS06009,2020MS06001,2015MS0604);内蒙古自治区直属高校基本科研业务费项目优秀青年基金资助项目(042);内蒙古自治区高等学校青年科技英才计划资助项目(NJYT23057,NJYT22074);内蒙古自治区科技计划资助项目(2020GG0315,2021GG0082,2021GG0023);内蒙古自治区卫生健康科技计划资助项目(202201395);草原英才“智能交通大数据分析与应用创新人才”创新人才团队。

摘　　要：为了提高结节检测和肺癌诊断的性能,本文提出一种改进的卷积神经网络(CNN)模型用于这2个步骤的结节特征提取。该模型以一种改进U-Net网络作为基础网络,并引入了ResNeXt模块和注意力机制,在保持模型复杂度的同时,提高网络学习多种形状和大小的结节特征的能力。实验结果表明:该模型在DSB数据集上表现良好,结节检测的灵敏度和特异性分别达到了99.15%和99.99%,肺癌诊断的准确率和AUC值分别达到了80.43%和0.86。可见,本文方法对于多种多样的结节特征具有高度敏感性,具有一定的临床价值。In order to improve the performance of lung nodule detection and lung cancer diagnosis,an improved convolutional neural network(CNN)model for feature extraction of nodules is proposed in these two steps.An improved U-Net is taken as the basic network,and ResNeXt module and attention mechanism are introduced to improve the ability of the network to learn the characteristics of nodules of various shapes and sizes while maintaining the complexity of the model.Experimental results show that the model performs well on DSB dataset.The sensitivity and specificity of lung nodule detection are 99.15%and 99.99%,respectively.The accuracy and AUC of lung cancer diagnosis are 80.43%and 0.86,respectively.It is concluded that the method is highly sensitive to a variety of nodule characteristics and has certain clinical value.

关 键 词：肺结节辅助检测 肺癌辅助诊断 CT图像 卷积神经网络 ResNeXt模块 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]