计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (16): 182-189.DOI: 10.3778/j.issn.1002-8331.2005-0037

• 模式识别与人工智能 • 上一篇    下一篇

深度卷积神经网络在心音分类方法中的应用

陈伟,孙强,齐月月,徐晨   

  1. 1.南通大学 医学院(护理学院),江苏 南通 226001
    2.南通大学 信息科学技术学院,江苏 南通 226019
  • 出版日期:2021-08-15 发布日期:2021-08-16

Deep Convolutional Neural Networks for Heart Sound Classification

CHEN Wei, SUN Qiang, QI Yueyue, XU Chen   

  1. 1.Medical School, Nantong University, Nantong, Jiangsu 226001, China
    2.School of Information Science and Technology, Nantong University, Nantong, Jiangsu 226019, China
  • Online:2021-08-15 Published:2021-08-16

摘要:

通过分析心音信号对心脏早期的病理状态进行确诊具有重要的意义。提出了一种基于深度卷积神经网络的心音分类方法。将心音信号转化成具有时频特性的梅尔频谱系数(Mel Frequency Spectral Coefficient,MFSC)特征图,将其作为深度卷积神经网络模型的输入;利用深度卷积神经网络对MFSC特征图进行训练,引入中心损失函数建立最优的深度学习模型;测试阶段,先将心音信号转换成多张二维MFSC特征图,然后利用训练好的深度学习模型对其分类,最后利用多数表决原则判断心音信号的类别。针对人工标注的训练样本有限,导致模型训练正确率不高的问题,以心音的二维MFSC特征图为对象分别从时间域和频率域进行随机屏蔽处理进而扩充训练样本。实验结果表明,该方法在PASCAL心音数据集上进行测试,对正常、杂音、早搏三种心音的分类性能明显优于现有最好的方法。

关键词: 心音分类, 深度卷积神经网络(DCNN), 数据扩充

Abstract:

It is of great significance to diagnose the early pathological state of the heart by analyzing the heart sound signals. This paper presents a heart sound classification method based on Deep Convolutional Neural Network(DCNN). Firstly, the heart sound signal is transformed into Mel feature maps with time-frequency characteristics, which are used as the input of the DCNN model. Then, the DCNN model is used to train the Mel feature maps, and the center loss function is introduced to establish the optimal deep learning model. In the testing stage, the heart sound signal is first converted into several two-dimensional Mel feature maps. Then, the feature maps are classified by the pre-trained deep learning model. Finally, the classification of heart sound signal is judged by the principle of majority voting. Due to the limited number of labeled samples, the accuracy of model is not high. In this paper, two-dimensional Mel feature maps of heart sound are randomly shielded in time domain and frequency domain in order to augment the training datasets. The experimental results show that the performance of this method is better than the state-of-the-art methods in the PASCAL heart sound datasets, which aims to classify normal, murmur and extrasystole heart sounds in the test samples.

Key words: heart sound classification, Deep Convolutional Neural Network(DCNN), data augment