融合注意力特征及动态卷积的肺结节辅助诊断  

Fusion of Attention and Dynamic Convolution for Computer Aided Diagnosis of Pulmonary Nodules

作　　者：谷宇[1] 刘佳琪 杨立东[1] 张宝华[1] 张祥松[1,2] 贾成一 GU Yu;LIU Jia-qi;YANG Li-dong;ZHANG Bao-hua;ZHANG Xiang-song;JIA Cheng-yi(Inner Mongolia Key Laboratory of Pattern Recognition and Intelligent Image Processing,School of Information Engineering,Inner Mongolia University of Science and Technology,Baotou 014010,China;BOE Technology Group Co.,Ltd.,Beijing 100176,China;China Second Metallurgy Group Corporation Limited,Baotou 014010,China)

机构地区：[1]内蒙古科技大学信息工程学院内蒙古自治区模式识别与智能图像处理重点实验室,包头014010 [2]京东方科技集团股份有限公司,北京100176 [3]中国二冶集团有限公司,包头014010

出　　处：《科学技术与工程》2023年第16期6834-6844,共11页Science Technology and Engineering

基　　金：国家自然科学基金(62001255,61962046,62161040,62262048,61841204);中央引地方科技发展资金项目(2021ZY0004);内蒙古自治区高等学校青年科技英才支持项目(NJYT23057);内蒙古科技大学基本科研业务费专项资金(042,019);内蒙古自治区自然科学基金(2019MS06003,2022MS06017,2015MS0604);内蒙古自治区科技计划项目(2020GG0315,2021GG0082,2021GG0023);草原英才创新人才团队、内蒙古自治区高等学校科学研究项目(NJZY145);教育部“春晖计划”合作科研项目(教外司留[2019]1383号)。

摘　　要：针对肺结节关键影像征象信息不易获取,部分卷积神经网络(convolutional neural networks,CNN)模型对肺结节的识别率不高的问题,提出一种融合注意力特征的动态卷积残差网络(dynamic convolutional residual networks incorporating attention features,DcANet),并在有效实现肺结节良恶性分类的基础上对所提模型的诊断结果进行可视化分析。此网络以适应肺结节三维小尺寸输入特点的残差网络为基本框架,在DcABlock部分使用可以自适应调整卷积参数的动态卷积以及迭代注意特征融合模块,使模型能够更准确地获取肺结节信息,提高模型的表征能力。此外,还使用类激活映射将三维图像的各层切片进行可视化分析。实验在最终测试集上的准确率为85.87%,平衡F分数(F1)值为82.67%,敏感度和特异性的综合指标Gmean值为85.51%。实验结果表明:该网络可以提升对肺结节良恶性分类的准确性,诊断结果具有可信性,有一定的临床应用价值。Aiming at the problem that the key imaging sign information of pulmonary nodules is difficult to obtain and the recognition rate of some convolutional neural networks(CNN)models for pulmonary nodules is low,an dynamic convolutional residual networks incorporating attention features(DcANet)was proposed.This network can not only realize the classification of benign and malignant pulmonary nodules,but also visually analyzed the diagnosis results of the proposed model.This network was based on a residual network that adapts to the three-dimensional small-size input characteristics of lung nodules.In the DcABlock part,the dynamic convolution and iterative attention feature fusion module were used,so that the model can accurately obtain the information of pulmonary nodules,thereby improving the representation ability.In addition,class activation mapping was used to visually analyze the slices of each layer of the 3D image.The accuracy of this experiment on test set is 85.87%,balanced F score(F1)value is 82.67%,comprehensive index of sensitivity and specificity Gmean value is 85.51%.The experimental results show that the DcANet has satisfactory accuracy in classifying benign and malignant pulmonary nodules,and the diagnosis results are reliable,which has certain clinical application value.

关 键 词：肺结节辅助诊断 动态卷积 迭代注意特征融合模块 深度学习 类激活映射 

分 类 号：R563[医药卫生—呼吸系统] TP391.4[医药卫生—内科学]