基于残差网络的血管内超声图像识别  

Intravascular Ultrasound Image Recognition Based on Residual Network

作　　者：杨云[1] 张立泽清 齐勇[1] 张文天 YANG Yun;ZHANG Li-zeqing;QI Yong;ZHANG Wen-tian(College of Electrical and Information Engineering,Shanxi University of Science and Technology,Xi’an Shanxi 710021,China)

机构地区：[1]陕西科技大学电气与信息工程学院,陕西西安710021

出　　处：《计算机仿真》2020年第4期269-273,共5页Computer Simulation

基　　金：陕西省重点研发计划项目(202021711);陕西省教育厅项目基金(15JK1086)。

摘　　要：为提高血管内超声(Intravenous Ultrasound,IVUS)图像在动脉粥样硬化识别准确率,实现更高效的计算机辅助诊断,提出综合使用图像增强、特征提取和基于批量归一化(Batch Normalization,BN)优化残差网络的血管内超声图像识别方法。使用Sobel算子在原图像水平和垂直方向进行边缘检测,在此基础上获得锐化增强图像,结合使用灰度共生矩阵提取纹理特征信息;为丰富网络的特征信息,防止梯度消失,使用残差学习对卷积神经网络进行改进。批量归一化通过拟合数据特征分布减少内部协变量转移加速网络收敛。实验结果表明上述方法相比较传统机器学习与改进前的卷积神经网络识别错误率平均降低了58.23%。In order to improve the accuracy of intravascular ultrasound(IVUS) images in atherosclerosis recognition and achieve more efficient computer-aided diagnosis, an IVUS image recognition method with a comprehensive use of image enhancement, feature extraction and batch normalization(BN) optimization is proposed. The Sobel Operator is used to detect the edge in the horizontal and vertical directions of the original image. On this basis, the sharpened enhanced image is obtained, and the texture feature information is extracted by using the gray level co-occurrence matrix. To enrich the feature information of the network and prevent the gradient from disappearing, the residual learning is used. The convolutional neural network is improved;and the batch normalization reduces the internal covariate transfer to accelerate network convergence by fitting the data feature distribution. The simulation results show that, compared with the traditional machine learning and unimproved convolutional neural network, the proposed method reduces the error rate of the method by 58.23%.

关 键 词：血管内超声 残差网络 灰度共生矩阵 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]