Cardiorespiratory Sound Separation Method Based On Dictionary Learning Network

Cardiopulmonary disease has become the "number one killer" threatening human health.According to the statistics of the World Health Organization in 2018,about 41 million people die of non-communicable diseases every year,of which about 53% die of cardiovascular diseases and chronic respiratory diseases.Stethoscope is a common medical instrument for preliminary screening of cardiopulmonary disease.But the crosstalk of cardiac sound and respiratory sound in temporal and frequency domain reduces the effectiveness of clinical auscultation and traditional signal processing methods such as band-pass filtering cannot separate cardiac sound and respiratory sound completely.Aiming at this problem and on the basis of the research on the technology of cardiorespiratory sound separation at home and abroad,this paper proposes a method of cardiorespiratory sound separation based on dictionary learning network:1)Based on NMF model and LSTM network,the methods of cardiorespiratory sound separation have achieved good results.They first obtain the time-frequency spectrum of cardiorespiratory sound by STFT,then separate the time-frequency spectrum of cardiorespiratory sound to obtain the estimated time-frequency spectrum of the cardiac sound and respiratory sound signal,and finally reconstruct the cardiac sound and respiratory sound signal in the temporal domain with the phase of the mixed signal.However,the set of trigonometric basis functions fixed by STFT is not necessarily the most suitable for cardiorespiratory sound separation.In order to solve this problem,a cardiorespiratory sound separation model embedded with a basic functions(dictionary)learning network is established in this paper,which realizes the end-to-end network optimization from the mixed signal in the time-domain to the cardiac sound and respiratory sound signal in the time-domain.The simulation results show that different from the existing technology based on STFT,this model can adaptively adjust the transform basis functions according to the training data which can enhance the performance of cardiorespiratory sound separation.2)The existing methods of cardiorespiratory sound separation assume that the energy ratio of the cardiac sound and the respiratory sound of the mixed signal is known,and then choose the separation model.However,the energy ratio of the cardiac sounds and the respiratory sounds of clinical auscultation signals is not known in advance.This will make the selection of separation model difficult and restrict the performance of cardiorespiratory sound separation.To solve this problem,on the basis of the model based on dictionary learning network,this paper proposes a multi-SNR model integrated network,which can estimate the best weight of different SNR separation models according to the mixed signal to output the cardiac sound and the respiratory sound corresponding to the actual SNR.The simulation results show that different from the existing technology which assumes that the energy ratio of the cardiac sound and the respiratory sound of the mixed signal is known,the integrated network of multiple SNR separation models solves the problem of adaptive matching between the cardiorespiratory sound separation model and the mixed signal under clinical conditions.Finally,this paper also proves the validity and necessity of cardiorespiratory sound separation under the condition of cardiac sound signal polluted by respiratory sound signal through the heart sound segmentation algorithm.