基于多特征广义深度自编码的肺结节诊断方法  

Lung nodules diagnosis using multi-features generalized deep auto-encoder based on extreme learning machine

作　　者：罗嘉滢 赵涓涓[1] 强彦[1] 唐笑先[2] LUO Jia-ying;ZHAO Juan-juan;QIANG Yan;TANG Xiao-xian(College of Computer Science and Technology,Taiyuan University of Technology,Taiyuan 030024,China;Department of PET/CT Center,Shanxi Provincial People’s Hospital,Taiyuan 030012,China)

机构地区：[1]太原理工大学计算机科学与技术学院,山西太原030024 [2]山西省人民医院PET/CT中心,山西太原030012

出　　处：《计算机工程与设计》2019年第1期154-160,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61373100);虚拟现实技术与系统国家重点实验室开放基金项目(BUAA-VR-16KF-13;BUAA-VR-17KF-15);山西省回国留学人员科研基金项目(2016-038)

摘　　要：针对基于深度学习的肺结节良恶性分类诊断中由于学习到的肺结节特征不够全面引起的分类准确率不高问题,提出一种基于多特征广义深度自编码的肺结节诊断方法。通过预处理构建肺结节图像的3种数据集作为输入;将流形学习引入基于极限学习机的深度自编码中,形成无监督广义深度自编码,利用该网络逐层提取特征;通过不同的融合策略对肺结节,进行良恶性分类。实验结果表明,该方法可以有效提高分类性能,肺结节分类的准确率达到94.72%。To improve classification accuracy of benign and malignant lung nodules when using deep learning models which cannot fully learn the complex feature of lung nodules,a lung nodules diagnosis using multi-features generalized deep auto-encoder based on extreme learning machine was proposed.Three datasets of lung nodules were built as input.Manifold learning was combined into stacked auto-encoder based on extreme learning machine(SAE-ELM)called unsupervised stacked auto-encoder based on generalized extreme learning machine(SAE-GELM).This model was used to extract features.The fusion strategy was used to classify the benign and malignant lung nodules.Experimental results show that the proposed method can effectively improve the accuracy and the classification accuracy rate of lung nodules reaches 94.72%.

关 键 词：肺结节 多特征 极限学习机 流形学习 自编码 良恶性分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]