CT影像识别的卷积神经网络模型  

Convolutional Neural Network Model of CT Images Recognition

作　　者：管姝 张骞予 谢红薇[1] 强彦[1] 程臻 Guan Shu;Zhang Qianyu;Xie Hongwei;Qiang Yan;Cheng Zhen(College of Computer Science and Technology,Taiyuan University of Technology,Taiyuan 030024;Health Care Center,Shanxi Dayi Hospital,Taiyuan 030024)

机构地区：[1]太原理工大学计算机科学与技术学院,太原030024 [2]山西大医院健康体检部,太原030024

出　　处：《计算机辅助设计与图形学学报》2018年第8期1530-1535,共6页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61373100);山西省回国留学人员科研资助项目(2016-038)

摘　　要：针对传统分类方法分割精度低、特征提取耗时等问题,构建一个适用于CT肺结节良恶性分类的卷积神经网络模型.首先确定网络深度、卷积核数目和卷积核大小等参数,构建卷积神经网络初始模型;然后选择激活函数类型、学习率和学习率衰减策略等训练参数;最后提出对感兴趣区域划分局部子区域的方式增强样本进行训练.在LIDC-IDRI数据集上进行实验的结果表明,准确率、特异性、敏感性及AUC值分别达到92.50%,0.91,0.94和0.93;对恶性结节的识别能力明显优于其他网络模型.Aiming at the problems of low segmentation accuracy and time-consuming feature extraction intraditional classification methods, a convolutional neural network（CNN） for identifying benign and malignantnodules in lung CT images is constructed. Firstly, the network depth, the number and size of convolutionkernel were determined, and the initial model of CNN was constructed. Secondly, selected the activationfunction, learning rate, learning rate decay strategy and other training parameters. Finally, the region of interestwas divided into a large number of local sub regions, and the enhanced data samples were used fortraining. On the LIDC-IDRI dataset, the accuracy, specificity, sensitivity and AUC value were 92.50%, 0.91,0.94 and 0.93 respectively. The recognition ability of malignant nodules is superior to other models.

关 键 词：卷积神经网络 肺部CT 图像分类 局部子区域 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]