基于深度卷积神经网络的心音分类算法

Heart sound classification algorithm based on deep convolutional neural network

摘要:

针对现有心音分类算法普适性差、依赖于对基本心音的精确分割、分类模型结构单一等问题,提出采用大量未经过精确分割的心音二维特征图训练深度卷积神经网络(Convolutional neural networks, CNN)的方法。首先采用滑动窗口方法和梅尔频率系数对心音信号进行预处理,得到大量未经过精确分割的心音特征图；然后利用深度CNN模型对心音特征图进行训练和测试。根据卷积层间连接方式的不同,设计了三种深度CNN模型：基于单一连接的卷积神经网络、基于跳跃连接的卷积神经网络、基于密集连接的卷积神经网络。实验结果表明基于密集连接的卷积神经网络比其他两种网络具备更大的潜力。与其他心音分类算法相比,该算法不依赖于对基本心音的精确分割且在分类准确率、敏感性和特异性方面均有提升。

Abstract:

Existing heart sound classification algorithms based on convolutional neural networks have the disadvantages of relying on precise segmentation of basic heart sounds, single classification model structure, and poor universality. So a method of training deep convolutional neural networks using a large number of two-dimensional heart sound feature maps that have not been accurately segmented is proposed. Firstly, the heart sound signal is preprocessed by the sliding window method and the Mel frequency coefficient to obtain a large number of heart sound feature maps that have not been accurately segmented. Then the deep CNN model is used to train and test the heart sound feature maps. According to the different connection modes between convolutional layers, three deep CNN models are designed: convolutional neural network based on single connection, convolutional neural network based on skip connection, and convolutional neural network based on dense connection. The experimental results show that the convolutional neural network based on dense connections has greater potential than based on single or skip connection. Compared with other heart sound classification algorithms, the algorithm we proposed does not rely on precise segmentation of basic heart sounds and has improved the accuracy, sensitivity and specificity of classification.