一种用于肺结节恶性度分类的生成对抗网络  

Generative Adversarial Networks for the Classification of Lung Nodules Malignant

作　　者：徐久强 洪丽萍 朱宏博 赵海 XU Jiu-qiang;HONG Li-ping;ZHU Hong-bo;ZHAO Hai(School of Computer Science&Engineering,Northeastern University,Shenyang 110169,China)

机构地区：[1]东北大学计算机科学与工程学院,辽宁沈阳110169

出　　处：《东北大学学报（自然科学版）》2018年第11期1556-1561,共6页Journal of Northeastern University(Natural Science)

基　　金：国家科技支撑计划项目(2012BAH82F04)

摘　　要：针对肺结节数据集中良恶性样本数比例失衡的问题,首次引入深度卷积生成对抗网络(deep convolutional generative adversarial networks,DCGAN)模型,该模型根据输入的肺结节图像,生成与输入图像具有相似纹理特征的肺结节并将生成图像用于训练DCGAN模型.此外,将图像来源分类问题改为图像来源分类和肺结节等级1～5分类问题,从而增强了DCGAN模型的抗噪能力和实现了DCGAN模型对肺结节的等级分类.实验表明,改进的DCGAN中G模型在生成图像时具有良好的抗噪能力且生成图像中大约有90. 42%的图像判别为真实图像,D模型对肺结节图像的等级分类具有较好的判别能力且肺结节等级分类准确率为70. 89%,肺结节良恶性分类准确率为80. 13%.In order to solve the proportion of benign and malignant lung nodule,a novel model named deep convolutional generative adversarial networks(DCGAN)was introduced.The model generates lung nodule images with similar texture feature from the input lung nodules images,and then using them to train the DCGAN model.In addition,the classification of image source is changed to the classification of image source and lung nodules grade 1~5.thus,the noise immunity of DCGAN model is enhanced and the classification of lung nodules by DCGAN model is realized.Experiments show that the model G in improved DCGAN enhances the performance of anti-noise capability with 90.42%images are distinguished true images when it generating images,and the model D has a good discriminant ability for the classification of lung nodule images and the classification accuracy of lung nodules is 70.89%,the recognition rate of malignant lung nodules is 80.13%.

关 键 词：肺结节 深度卷积生成对抗网络(DCGAN) 纹理特征 改进DCGAN 肺结节等级分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]