Research On Lung Sound Detection Method Based On Convolutional Neural Network

Abnormal detection of lung sounds collected by an electronic stethoscope,determining whether it contains rales,can greatly improve the efficiency of early screening of respiratory diseases.However,the existing technologies for the detection of abnormal lung sounds still have the following challenges:(1)The effective sample data is small,and the positive and negative samples are not balanced.The current electronic stethoscope has not yet been popularized,and the cost of data marking is high(requiring professional senior doctors to mark),and the amount of normal lung sound data is much larger than abnormal lung sounds(Including rales)data volume;(2)Lung sounds often contain a lot of noise,such as the friction sound between the stethoscope and the skin and clothes,heart sounds,and environmental noises in which speech sounds take a majority part,etc.In order to solve above technical problems,the main work of this article includes:(1)Design a set of lung sound data collection paradigm and evaluation method,perform band-pass filter preprocessing on the collected lung sound to suppress noise interference,and use the mel spectrogram to extract the original features of lung sound.(2)Study the influence of multiple sample enhancement methods on the generalization ability of different detection models,including time-shifting,changing the pitch of lung sounds,and adding white noise according to a certain signal-to-noise ratio.In particular,a data enhancement method in which random speech is added according to the signal-to-noise ratio is proposed,which effectively enhances the robustness of the detection model against speech interference.(3)On the basis of sample enhancement,a subdivision discriminant feature extraction and integration strategy based on VGG convolutional neural network is proposed.First,use two VGG models to extract the feature vectors of crackle sounds and wheeze sounds,and then use 4 fully connected layers to fuse the two sets of features to distinguish between normal lung sounds and rales.It avoids the inclination of feature learning to wheeze sounds with sufficient annotation information,thereby improving the sensitivity of the detection model to crackle sounds.And the Focal Loss function is used to alleviate the imbalance of positive and negative samples,and this function can reduce the weight of easy-to-classify samples during the training process,so that the model can focus more on difficult-to-classify samples.In this paper,we use the Mel spectrogram generated by lung sound as the feature,and use a variety of models to realize the abnormal detection of lung sound.The models include Support Vector Machine(SVM),VGGNet,Google Net,Res Net and the fine classification discriminant model based on VGG convolutional neural network designed in this paper.Experiments prove that the experimental results of fine-tuned VGGNet are better than SVM,Google Net and Res Net.The VGG-based subdivision discriminant feature integration network designed in this paper is slightly better than the fine-tuned VGGNet.