摘要
目的:针对目前计算机辅助诊断中肺结节特征提取单纯依靠人工设计,分类结果存在很大差异这一问题,提出一种新的肺结节特征自动提取及良恶性分类方法。方法:首先通过阈值概率图从肺部CT图像中分割肺结节图像,然后通过一个2层的稀疏自编码神经网络自动提取肺结节图像的特征,最后利用Logistic回归分类器对提取到的特征进行良恶性分类。结果:肺部图像数据库联盟(1ung image database consortium,LIDC)数据库上的实验结果表明,与目前基于人工设计的特征提取方法相比,该提取方法获得了最高的分类精度与曲线下面积(area under curve,AUC)值。结论:稀疏自编码神经网络能够直接从肺结节图像本身自动提取肺结节特征,避免了人工提取及选择的差异性,提高了肺结节良恶性分类的准确度,能够为临床诊断提供参考依据。
Objective To put forward a method for feature extraction and benign or malignant classification of lung nodules to solve the problems due to manual operation. Methods Firstly, lung nodule images were segmented using the threshold probability-maps method from lung CT images. Secondly, sparse auto-coding neural network was applied in feature extraction and high level features of lung nodules were extracted through multiple auto-encoder. Finally, logistic regression was used to classify benign or malignant nodules by the extracted features. Results Experimental results based on lung image database consortium （LIDC） data set showed that the proposed method had higher classification accuracy and area under curve （AUC） value when compared with manual method. Conchuslon Sparse auto-encoder neural network can extract the features from image level directly, avoid the difference due to manual extraction and selection, improve the classification accuracy and provide references for clinical diagnosis. [Chinese Medical Equipment Journal,2015,36 （12）：7-10, 14]
出处
《医疗卫生装备》
CAS
2015年第12期7-10,14,共5页
Chinese Medical Equipment Journal
基金
徐州市科技计划项目(KC14SM089)
江苏省大学生科研训练计划项目(201410313062X)
徐州医学院振兴计划项目
关键词
肺结节
特征提取
稀疏自编码神经网络
良恶性分类
lung nodule
feature extraction
sparse auto-encoder neural network
benign or malignant classification
