摘要
目的介绍卷积神经网络方法,并将其应用于肺部多模态图像进行肺结节良恶性分类。方法基于肺部PET-CT多模态图像,分析临床信息与卷积神经网络的结合并与机器学习支持向量机方法作效果对比。结果52名患者的323张结节图像纳入研究。研究表明,模型中纳入临床信息能够改进模型的分类效果,准确率、灵敏度、特异度分别为0.913、0.942、0.417;在与支持向量机模型的对比分析中,卷积神经网络模型特异度较低为0.417,而灵敏度达到0.942;支持向量机模型灵敏度较低为0.570,而特异度达到0.927。结论基于肺部PET-CT多模态图像进行肺结节良恶性判别,卷积神经网络能够获得较高的灵敏度。
Objective To establish a diagnosis model to classify the detected nodules into benign and malignant groups based on the lung PET-CT images.Methods We analyzed the influence of clinical information on the model's performance and compared the convolutional neural network with supporting vector machine.Results 52 patients and 323 lung nodule images were involved in this study.It showed involving the clinic information(demographic variables and nodule morphological features)in model could improve the performance as accuracy,sensitivity and specificity were increased to 0.913,0.942,0.417 Respectively.The convolutional neural network showed markedly perfect sensitivity(94.2%vs 57.0%)while the supporting vector machine showed better specificity(92.7%vs 41.7%).Conclusion For the diagnosis of lung cancer,convolutional neural network could achieve higher sensitivity.
作者
武志远
马圆
唐浩
姚二林
郭秀花
Wu Zhiyuan;Ma Yuan;Tang Hao(Department of Public Health,Capital Medical University(100069),Beijing)
出处
《中国卫生统计》
CSCD
北大核心
2019年第6期806-808,813,共4页
Chinese Journal of Health Statistics
基金
国家自然科学基金(81773542)
北京市自然科学基金资助项目(Z160002)
