Research And Implementation Of Lung Sound Classification Algorithm Using Digital Stethoscope

In recent years,with the acceleration of urban industrialization and the declining of air quality,respiratory diseases have become an important hidden danger to human health.Professional medical clinicians are required to diagnosis of respiratory diseases with the use of a stethoscope for auscultation.To ensure that respiratory diseases can be diagnosed as early as possible,and to reduce the workload of medical staff,intelligent lung sound auscultation research based on machine learning has received increasing attention,especially in poorly medically equipped remote areas.Therefore,this dissertation focus on the development of intelligent lung sound auscultation system,LungSys,based on deep neural learning technology.LungSys collects the user’s lung sound data using a commercial digital stethoscope.Then,we use our application software to classify the data into"normal","wheeze","crackle",and "crackle plus wheeze" classes.To improve the accuracy of the classification results,we implemented pre-processing,denoising,and feature extraction techniques to process the lung sound signals.We developed a bi-ResNet network model based on convolutional neural networks,which learn two different features of the same sample.Based on the ICBHI 2017 dataset,we have verified our algorithms in terms of accuracy,confusion matrix,official score,sensitivity,and specificity.We have also completed the development of the Android software application design.Our software provides real-time analysis of the signals and displayed the classification result through a graphical interface.Finally,we tested LungSys with five volunteers with reasonable accuracy.In the test,We found that the accuracy of the results for each experimental subject is between 67%-75%.The time required for all samples to run the network model is less than 10s.