基于卷积神经网络的CT图自动检测深度学习  

Automatic detection deep learning of CT images based on convolutional neural network

作　　者：张少宇 ZHANG Shao-yu(Experimental Teaching Center,Guangdong University of Finance,Guangzhou 510521,China)

机构地区：[1]广东金融学院实验教学中心

出　　处：《信息技术》2019年第11期121-124,共4页Information Technology

摘　　要：针对肺结节良恶性诊断的难度,采用自定义卷积神经网络对其进行建模分析。通过多次实验分析,构建得到了可以实现7层肺结节检测的卷积神经网络模型,在每层中都含有训练参数,得到13×13的卷积核。测试肺结节的算法通过4个指标进行评价,分别为准确性和特异性、敏感性和假阳率。PndCnn-7参数优化结果得到,学习率lr介于[0.4,1.05]范围内,需要处理的批次达到最少0.75。13×13卷积核能够实现网络的快速收敛,并且不会引起震荡的现象。当epoch增大至50,则会引起误差的明显减小,使网络达到良好的收敛状态。In view of the difficulty in the diagnosis of benign and malignant pulmonary nodules,a custom convolutional neural network is adopted for modeling and analysis.Through multiple experimental analyses,a convolutional neural network model that can realize the detection of 7-layer pulmonary nodules is constructed.Training parameters are contained in each layer,and the convolution kernel of 13×13 was obtained.The algorithm for testing pulmonary nodules is evaluated by four indicators,namely accuracy,specificity,sensitivity and false positive rate.The Pndcnn-7 parameter optimization results show that the learning rate lr is within the range of[0.4,1.05],and the batches to be processed reached the minimum of 0.75.The convolution kernel of 13×13 can achieve fast convergence of the network without causing oscillation.If the epoch increased to 50,it would cause a significant decrease in error and make the network reach a good convergence state.

关 键 词：卷积神经网络 CT图 自动检测 深度学习 

分 类 号：TP31[自动化与计算机技术—计算机软件与理论]