基于三维卷积神经网络降低肺结节检测假阳性的方法  

Detection of false positive reduction of pulmonary nodules based on three-dimensional convolutional neural network

作　　者：侯智超 杨杨[1] 李晓琴[1] 王晓曦[1] 高斌[1] HOU Zhichao;YANG Yang;LI Xiaoqin;WANG Xiaoxi;GAO Bin(Faculty of Environment and Life,Beijing University of Technology,Beijing 100124)

机构地区：[1]北京工业大学环境与生命学部,北京100124

出　　处：《北京生物医学工程》2022年第4期352-359,380,共9页Beijing Biomedical Engineering

基　　金：国家自然科学基金(61931013、81701644、11832003);国家重点研发计划(2017YFC0111104)资助。

摘　　要：目的为降低在计算机断层扫描图像中筛查肺结节的假阳性率,提出了一种基于三维卷积神经网络降低肺结节检测假阳性的方法。方法选用美国2016年肺结节分析挑战赛提供的两个版本的开源数据集,分别用于模型的训练和测试。首先应用图像增强技术解决数据集中正负样本分布不均衡的问题,并基于多视角采样技术扩充正样本;基于自动编码器及K-means无监督聚类方法将负样本分为5类,并分别与正样本组合得到了5个训练集,该方法既减少了每个数据集中负样本的样本量又保证了负样本的多样性。然后搭建三维卷积神经网络,并分别使用构建的5个训练集训练网络,在此过程中不断调整和优化网络结构和参数,得到5组降低肺结节假阳性检测模型,接着利用简易集成法对肺结节进行综合判决。结果经测试,模型的敏感性和特异性分别为0.966和0.996,通过FROC曲线计算得出CPM得分为0.886。结论本文提出的方法可以有效降低肺结节检测假阳性,可以为肺癌筛查工作提供有效帮助。Objective In order to reduce the false-positive rate of pulmonary nodules in the preliminary screening of CT images,a method was proposed based on the 3D convolutional neural network(3D-CNN).Methods Two versions of open source datasets provided by the 2016 pulmonary nodule analysis challenge in the United States were selected for training and testing of the model.To balance the distribution of the positive and negative samples,we augmented the positive samples by the multi-view sampling and classified the negative samples into five categories by the auto-encoder and the K-means unsupervised clustering.Then we got five training sets through combining the negative samples with the positive samples.This method reduced the size of negative samples and ensured its diversity.Next,we constructed the 3D convolutional neural network and respectively trained it with the five training sets,and adjusted the structure and parameters until obtaining a better model.Finally,the pulmonary nodules were judged by synthesizing the results of the five models.Results The sensitivity and specificity of the model were 0.966 and 0.996 respectively,and the CPM score was 0.886 calculated by the FROC curve.Conclusions The method we proposed can effectively reduce the false-positive detection of pulmonary nodules and provide help for lung cancer screening.

关 键 词：图像处理 深度学习 三维卷积神经网络 计算机辅助检测 肺结节 

分 类 号：R318.04[医药卫生—生物医学工程] TP391[医药卫生—基础医学]