电子学报 ›› 2022, Vol. 50 ›› Issue (10): 2425-2432.DOI: 10.12263/DZXB.20200713

• 学术论文 • 上一篇    

基于一维卷积神经网络与循环神经网络串联的心音分析方法

肖斌1, 陈嘉博1, 毕秀丽1, 张俊辉2, 李伟生1, 王国胤1, 马旭3   

  1. 1.重庆邮电大学图像认知重庆市重点实验室,重庆 400065
    2.重庆医科大学附属第一医院,重庆 400042
    3.国家卫生健康委员会科学技术研究所,北京 100081
  • 收稿日期:2020-07-14 修回日期:2021-06-10 出版日期:2022-10-25 发布日期:2022-10-11
  • 作者简介:肖 斌 男,1982年出生,重庆人.博士,教授,博士生导师,CCF会员.主要研究领域为图像增强与复原.E-mail: xiaobin@cqupt.edu.cn
    陈嘉博 男,1996年出生,浙江人.硕士研究生.主要研究领域为图像识别与分析.
  • 基金资助:
    国家自然科学基金(61806032);国家重点研发计划(2016YFC1000307-3);重庆市基础与前沿项目(cstc2018jcyjAX0117);重庆市教委科学技术研究计划重点项目(KJZD-K201800601)

A Method of Heart Sound Analysis Based on One-Dimensional Convolutional Neural Network and Recurrent Neural Network

XIAO Bin1, CHEN Jia-bo1, BI Xiu-li1, ZHANG Jun-hui2, LI Wei-sheng1, WANG Guo-yin1, MA Xu3   

  1. 1.Chongqing Key Laboratory of Image Cognition,Chongqing University of Posts and Telecommunications,Chongqing 400065,China
    2.The First Affiliated Hospital of Chongqing Medical University,Chongqing 400042,China
    3.Institute of Science and Technology,National Health Commission,Beijing 100081,China
  • Received:2020-07-14 Revised:2021-06-10 Online:2022-10-25 Published:2022-10-11

摘要:

面向心脏疾病计算机辅助诊断,本文提出一种基于一维卷积神经网络和循环神经网络混合深度学习结构的心音分析方法.本结构首先利用卷积神经网络学习心脏病症在心音信号上的表征,然后通过循环神经网络处理心音信号中的时序信息进行分类,在提升心音分类正确率的同时,大幅度降低了网络参数.为验证本深度学习结构所学特征的有效性,除已有的成人心音数据集外,本文还专门构建了一个面向婴幼儿先天性心脏病的心音数据集,并通过端到端的类别响应图证明了本方法在室缺诊断时学习到的心音信号特征符合临床医师的心音听诊经验.实验结果表明,本文方法能在3 153例成人心音数据分类上达到92.56%的正确率,在528例婴幼儿心音数据分类上达到97.48%正确率,模型参数仅有0.05 M.

关键词: 心音听诊, 一维卷积神经网络, 循环神经网络, 类别响应图

Abstract:

For the computer-aided heart disease diagnosis, this paper proposes a method of heart sound analysis based on the mixed structure of one-dimensional convolutional neural network and recurrent neural network. The proposed structure uses the convolutional neural network to learn the representation of heart disease on the heart sound signal, and then processes the time sequence information in the heart sound signal through the recurrent neural network for classification, which greatly reduces the network parameters while improving the accuracy of heart sound classification. The experimental results show that the proposed method can achieve the accuracy of 92.56% on the classification of 3153 cases of normal and abnormal heart sounds for adults, and 97.48% on the classification of 528 cases of normal and abnormal heart sounds for infants and children. The parameter of the proposed method is 0.05M, which is suitable for portable application situations.

Key words: heart sound analysis, one-dimensional convolutional neural network, recurrent neural network, class activation map

中图分类号: