Separation And Recognition Of Cardiopulmonary Sounds Based On Complex-valued Neural Network

In recent years,cardiopulmonary diseases have gradually leapt to primary factors threatening human health.Now auscultation has been overlooked in the consultation process,and the problem is becoming more serious by the shortage of medical resources and the exposure of physicians to danger in traditional auscultation due to the COVID-19.Electronic stethoscopes,with their advantages of wireless connectivity,real-time transcription,and online analysis,have become a new means of solving the above problem.Specifically,the deep learning methods emerging in recent years have excellent ability to extract data features,making it possible to automatically analyze the cardiac and lung sounds online.When using an electronic stethoscope to collect signals,cardiac sounds and lung sounds often appear mixed,which will seriously affect the subsequent recognition tasks,so it is necessary to study cardiopulmonary sound separation method and hence improve the practicability of electronic stethoscopes.Traditional monaural speech/acoustic signal separation methods only deal with the amplitude of short-time Fourier transform(STFT)results,but the ignored phase would also include noise interference,and they also do not take the characteristics of cardiac and lung sounds into account.The current cardiac or lung sounds recognition methods usually use pure cardiac and lung sounds,and no study has analyzed the cardiac and lung sounds recognition performances before and after separation.In this dissertation,diagnosis of cardiopulmonary disease in a practical scenario where cardiac and lung sounds are highly overlapping,which has been neglected by previous studies,is studied.The complex-valued artificial neural network and the characteristics of cardiac and lung sounds explored to develop the methods of cardiac and lung sounds separation and recognition.The work in this dissertation has two parts that followed.The first part of the work focuses on the cardiopulmonary sound separation problem.In contrast to that conventional monaural speech separation methods only focus on processing the real-valued amplitudes of STFT time-frequency spectrum,resulting in information loss,in this dissertation complex-valued neural networks for cardiopulmonary sounds separation are designed.In addition,the quasi-periodicity of cardiac sound signals was particularly considered in the process of cardiopulmonary sound separation,quantified by cyclostationarity.The cost function was built by calculating the cyclostationary frequency as a part of it,for the optimization of cardiopulmonary sound separation network.In this dissertation,the complex-valued neural network and cyclostationarity involved separation method makes full use of all the amplitude and phase information of the time-frequency spectrum,and the introduction of cyclostationarity to the cost function not only optimizes the separation result of cardiac sounds,but also improves the separation result of lung sounds.The proposed cardiopulmonary sound separation method can be easily implemented online,and its performance has been verified in various simulation or real cardiopulmonary sound separation tasks.Compared to the recently published cardiopulmonary sound separation methods,the proposed method achieves higher signal artifact ratio(SAR),signal distortion ratio(SDR),and signal interference ratio(SIR)for both separated cardiac and lung sounds.In the second part,the cardiopulmonary sound recognition method was studied.Combined with the method of cardiopulmonary sound separation,the final diagnostic performance of cardiopulmonary disease after the entire process of cardiopulmonary sounds separation and recognition was evaluated.Due to the diversity of cardiopulmonary diseases,the task in this study was specialized to identify five categories of cardiac diseases/states:aortic stenosis(AS),mitral regurgitation(MR),mitral stenosis(MS),mitral valve prolapse(MVP),and normal cardiac sounds.Inspired by SENet,a typical soft attention mechanism,a cardiac sound pattern recognition method based on complex-valued SENet is proposed in this dissertation.The complex-valued STFT spectrum is input to the built network,consisting of complex-valued convolution module and complex-valued SENet,to extract the deep features,and classification results are output through the fully connected layer.Compared with other existing cardiac sound recognition methods,the proposed method achieves the highest recognition indexes.Combined with the cardiopulmonary sound separation method proposed in this dissertation,the proposed cardiac sound recognition method is applied to cardiopulmonary sound mixture data,and superior recognition results are also obtained.For real application scenarios,solutions for two successive problems in the electronic stethoscope,cardiopulmonary sound separation and recognition,were put forward exploiting the advantages of complex-valued neural network in the comprehensive utilization of amplitude and phase information.The proposed separation method based on complex-valued neural network and cyclostationarity can provide purified cardiac sounds and lung sounds.Complex-valued SENet based cardiac sound recognition after separation can give accurate cardiac sound recognition results.The study in this dissertation may provide a stable and reliable methodology for the study of automatic diagnosis of cardiopulmonary diseases by electronic stethoscope in practical scenarios.