Research And Application Of Lightweight Cardiopulmonary Sound Separation Network

Auscultation is the easiest and most convenient method for screening cardiovascular and respiratory diseases,and this method has the advantages of high efficiency and noninvasiveness.Heart and lung sound signals that characterize the health of the heart and lungs can be obtained through a stethoscope.In real life,the sound signal directly heard by a doctor using a stethoscope is often a mixed signal of heart and lung sounds,which is difficult to use for refined disease analysis.Deep learning has been proven to be useful for the separation of cardiopulmonary sound signals.The cardiopulmonary sound separation method based on deep learning has better separation effect,more network layers and larger network scale.However,they also have higher requirements on the computing resources of computer equipment.The cardiopulmonary sound separation method based on the non-negative matrix factorization model requires less computing resources,but the separation effect is not as good as the separation method based on deep learning.In view of the above problems,this thesis studies a lightweight heart and lung sound separation network,designs and develops an intelligent digital auscultation system to improve the efficiency of heart and lung sound separation.The main research contents of this thesis are as follows:（1）Research the application of knowledge distillation in lightweight cardiopulmonary sound separation network.Aiming at the problems of a large number of parameters,a large amount of calculation,and a slow running speed in the deep learning-based heart and lung sound separation network,this thesis uses the knowledge distillation method to compress the heart and lung sound separation network.This method uses the "student-teacher" framework,first trains a teacher network with better effect,and uses the prediction result of the teacher network as a "pseudo-label" to supervise the training of the student network with a smaller network size,so as to transfer the knowledge in the teacher network to student network.The experimental results show that due to the simple network structure and other reasons,the separation effect of the distilled student network can not surpass the teacher network,but compared with the separately trained student network,the signal-to-noise ratio of the separated heart sound signal and lung sound signal of the distilled student network is 0.3d B higher than that of the separately trained student network（2）Research the application of multi-teacher knowledge distillation in lightweight cardiopulmonary sound separation network.In order to improve the problem of insufficient knowledge polymorphism in the traditional single-teacher knowledge distillation method,this thesis applies the multi-teacher knowledge distillation method to the heart and lung sound separation network.Teacher network A has a wider structure and can learn more hidden information;teacher network B has deeper layers and better generalization ability.Using the strategy of two teachers’ cooperative teaching knowledge,it provides more abundant information for the student network and improves the separation accuracy of the student network.Experiments show that the signal-to-noise ratio of heart sound signal and lung sound signal separated by the student network under the supervision of multiple teachers is 0.5d B higher than that of the student network trained alone.（3）Design and implement an intelligent digital auscultation system.In order to apply the lightweight cardiopulmonary sound separation network to the clinic,this thesis builds an intelligent digital auscultation system suitable for non-professional medical conditions such as outside the hospital,and builds an algorithm server platform based on the Pyside2 library.The experimental test of the system shows that the system has certain practicality.It can meet the needs of out-of-hospital intelligent auscultation under certain conditions.