多分支卷积神经网络肺结节分类方法及其可解释性  

Multi-branch Convolutional Neural Network for Lung Nodule Classification and Its Interpretability

作　　者：张佳嘉 张小洪[1] ZHANG Jia-jia;ZHANG Xiao-hong(School of Big Data&Software Engineering,Chongqing University,Chongqing 400000,China)

机构地区：[1]重庆大学大数据与软件学院,重庆400000

出　　处：《计算机科学》2020年第9期129-134,共6页Computer Science

基　　金：国家自然科学基金(61772093);重庆市重大主题专项项目(cstc2018jszx-cyztzxX0017)。

摘　　要：肺结节CT图像表征复杂且多样,导致对肺结节进行分类较为困难。虽然越来越多的深度学习模型被应用到计算机辅助肺癌诊断系统的肺结节分类任务中,但这些模型的“黑盒”特性无法解释模型从数据中学习到了哪些知识,以及这些知识是如何影响决策的,导致诊断结果缺乏可信性。为此,文中提出了一种可解释的多分支卷积神经网络模型来判别肺结节的良恶性。该模型利用医生诊断时所用的肺结节语义特征信息来辅助诊断肺结节的良恶性,并将这些特征与肺结节良恶性判别网络融合成多分支网络,在完成肺结节良恶性诊断任务的同时,得到肺结节相关语义特征的预测结果,为医生提供可信的诊断依据。在LIDC-IDRI数据集上的实验结果表明,与现有方法相比,所提模型不仅可以得到可解释的诊断结果,而且实现了更好的肺结节良恶性分类效果,其准确率可达97.8%。The characteristics of lung nodules are complex and diverse,which make it difficult to classify lung nodules.Although more and more deep learning models are applied to the lung nodule classification task of computer-aided lung cancer diagnosis systems,the“black box”characteristics of these models cannot explain what knowledge the model has learned from the data and how the knowledge influences the decision,leading to a lack of reliability in the diagnosis results.To this end,an interpretable multi-branch convolutional neural network model is proposed to identify the benign and malignant lung nodules.The model uses the semantic features of the pulmonary nodules which radiologists use in diagnosis to assist identifying the benign and malignant lung nodules.These characteristics are combined with the branch of malignancy classification into a multi-branch network.Then beyond the malignancy classification,the model can predict nodule attributes,which could potentially explain the diagnosis result.Experimental results on the LIDC-IDRI dataset show that,compared with the existing methods,the proposed model can not only obtain interpretable diagnostic results,but also achieve better classification of lung nodules with an accuracy rate of 97.8%.

关 键 词：计算机辅助诊断 卷积神经网络 多分支 可解释性 肺结节恶性程度分类 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]