Research On U-net Segmentation Algorithm And Classification Of Heart Sound Signal

The research on segmentation and classification of heart sound signals was a research hotspot.On the one hand,the information of each state can be obtained by segmenting the heart sound signal,and the health status of the heart can be preliminarily evaluated;On the other hand,aiming at the research of heart sound signal classification,segmentation was an important breakthrough in classification research.Hidden Semi-Markov Model(HSMM)was a model commonly used for heart sound segmentation,but this method needs to add the prediction time of each state,and errors will occur in the case of long period and irregular sinus rhythm.The method of deep learning directly used the network structure to learn the sound features that can minimize the segmentation errors from the heart sound signal itself or the features extracted from the heart sound signal,and then completed the heart sound segmentation.Aiming at the research of heart sound signal,this topic has carried out the related work of heart sound segmentation and classification,mainly as follows:(1)We analyzed the physiological mechanism and time-frequency domain characteristics of heart sound signals,and transformed single-channel heart sound data into four-channel data by filtering,peak elimination,envelope extraction and data expansion,which laid the foundation for further analysis and processing.(2)792 cases of heart sound segmentation based on Convolutional Neural Network(CNN)U-net model were studied,and 121 cases of clinical data were used to test the performance of the model.By optimizing the model,a U-net segmentation model with good segmentation performance was obtained on the public data set,in which PA was0.994,CPA was [0.986 0.993 0.994 0.996],and MPA was 0.992.Further,the U-net model was applied to the clinical data,and the segmentation results of PA was 0.863,CPA was [0.909 0.774 0.818 0.912],and MPA was 0.853.(3)The research on the classification of heart sounds was carried out,including the classification of normal abnormalities and different diseases.Firstly,the U-net model was used to segment the heart sounds of the data set.Then,the Adaboost classifier,CNN-based classifier and CNN-Long Short-Term Memor(LSTM)-based classifier were used on 301 cases of public data sets,and the CNN-LSTM model with good performance was screened out.The results were as follows: Acc is 0.806,Se was 0.809,Sp was 0.802,F1 score was 0.807;Furthermore,the model with good performance was applied to 1000 signals in another open database with different disease categories,and the classification effect of Acc was 0.92,Se was 1.0,Sp was 0.90,and F1 score was0.99.Finally,the model was used to classify different diseases on this data set,and the classification results were as follows: the positive predictive value of aortic stenosis,mitral insufficiency,mitral stenosis and systolic murmur were 0.864,0.85,0.75 and0.793 respectively..The research showed that the heart sound segmentation algorithm based on U-net combined with the heart sound classification algorithm based on CNN-LSTM can show good performance in the research of heart sound signals.