基于轻量级卷积神经网络的心脏杂音分级方法  

A heart murmur grading method based on lightweight convolutional neural networks

作　　者：黄昭涵 何培宇[1] 李世龙 李莉[2] 赵启军[3] 潘帆[1] HUANG Zhaohan;HE Peiyu;LI Shilong;LI Li;ZHAO Qijun;PAN Fan(School of Electronics and Information Engineering,Sichuan University,Chengdu 610065,China;Division of Pediatric Cardiology,West China Second University Hospital,Sichuan University,Chengdu 610041;School of Computer Science,Sichuan University,Chengdu 610065)

机构地区：[1]四川大学电子信息学院,成都610065 [2]四川大学华西第二医院儿童心血管科,成都610041 [3]四川大学计算机学院,成都610065

出　　处：《生物医学工程研究》2024年第6期423-431,共9页Journal Of Biomedical Engineering Research

基　　金：国家自然科学基金项目(62066042);四川省重点研发项目(2022YFG0045);中央高校基本科研业务费专项资金(2022SCU12008)。

摘　　要：针对人工听诊存在经验依赖和主观性问题,本研究提出了一种基于轻量级卷积神经网络的心脏杂音分级方法。首先,采用滑动窗口和Gammatone滤波器组对心音信号进行预处理,将心音片段的对数耳蜗谱图作为网络的输入;其次,设计初始卷积模块和选择性卷积模块捕获全局特征和多尺度特征,并使用深度可分离卷积降低网络参数量;最后,基于提出的决策规则对多个听诊区的预测结果进行综合判断,得到患者的心脏杂音等级。本研究在CirCor DigiScope PCG数据集上的实验结果显示,测试集上的未加权平均召回率、加权平均召回率和未加权F1分数分别达到81.63%、88.46%和79.85%。本研究方法具有较好的杂音分级性能,不仅适用于终端设备,还可为心脏疾病的自动分析提供重要依据。Aiming at the issues of experience dependence and subjectivity in artificial auscultation,a heart murmur grading method based on lightweight convolutional neural networks was proposed.Firstly,the sliding window method and Gammatone filter bank were used to preprocess the heart sound signal,and the log-cochleagram of the heart sound fragment was obtained,which was used as the input feature of the network.Secondly,the initial convolutional module and selective convolutional module were designed to capture global features and multi-scale features,and the depth-separable convolution was used to reduce the number of network parameters.Finally,based on the proposed decision rules,the prediction results of multiple auscultation sites were combined to obtain the heart murmur grade of patients.The validation experiment results on the CirCor DigiScope PCG dataset showed that the unweighted average recall rate,weighted average recall rate and unweighted F1 score on the test set reached 81.63%,88.46%and 79.85%,respectively.This method has better performance of murmur grading,which is not only suitable for terminal equipment,but also provides an important basis for automatic analysis of heart disease.

关 键 词：心脏杂音 杂音分级 对数耳蜗谱 轻量级网络 选择性卷积 临床决策 

分 类 号：R318[医药卫生—生物医学工程] TN912.3[医药卫生—基础医学] TP39[电子电信—通信与信息系统]