Research On Lung Sound Recognition Based On Convolutional Recurrent Neural Network

According to the survey data of the World Health Organization,the incidence and mortality rate of lung diseases have increased year by year,which has had a serious impact on the lives of the general population.In clinical practice,auscultation of lung sounds is an important method for doctors to diagnose the lung health of patients.With the rise of artificial intelligence technology,the use of deep neural networks to analyze lung sounds can help doctors get the right diagnosis more quickly.However,because the traditional single network model cannot deal with the complex characteristics of abnormal lung sound signals and the problem of few data records,the task of lung sound recognition faces great challenges.In order to solve this problem,a lung sound recognition method based on convolutional recurrent neural network is designed,and the following work is conducted to enhance the performance of the lung sound recognition mode:(1)In terms of lung sound preprocessing and feature extraction,a method for highpass filtering to remove low-frequency noise and a wavelet threshold method for removing heart sounds are designed to remove noise interference in lung sounds.At the same time,lung sounds are segmented and labeled according to the respiratory cycle,so that multiple lung sound samples are obtained,including normal lung sounds and other three types of abnormal lung sounds.Finally,the spectral pattern obtained by short-term Fourier transform and the characteristics of Mel frequency cepstral coefficient are extracted for lung sounds.(2)For the purpose of solving the limitations of a single network model in lung sound signal recognition tasks,an improved CNN-Bi LSTM model is constructed,which made full use of the advantages of CNN being able to extract locally important feature information and Bi LSTM being able to learn the temporal characteristics of abnormal lung sound signals.In addition,an improved method is proposed to alleviate the problem of low recognition accuracy caused by unbalanced dataset,and the focal loss function is used in this module for the first time to balance the weight of difficult lung sound samples and easy lung sound samples in the cost function,so as to improve the recognition accuracy of abnormal lung sounds with small sample size.(3)Since the CNN-Bi LSTM model can only learn a single feature of the sample,the difference between different lung sound features is small,resulting in the overall recognition accuracy of the model being low.To solve this problem,a dual-channel CNNBi GRU model is proposed to further capture the detailed information of lung sound characteristics.The parallel CNN network is used to simultaneously learn two different features of the samples to obtain the spatial fusion characteristics of the lung sound signal,and then the temporal correlation of the lung sound signal is learned by the Bi GRU network with relatively few parameters.In this study,a series of related experiments are carried out based on the ICBHI 2017 lung sound dataset.The effects of multiple variables on the lung sound are compared,such as lung sound denoising,feature extraction method and loss function selection.In addition,the effectiveness of CNN-Bi LSTM model and dual-channel CNN-Bi GRU model on lung sound recognition tasks is verified.At the same time,the experimental results of the existing research methods are compared to verify the reliability of the study in the lung sound recognition task.