Convolutional neural networks based approach for classification of lung sounds

Over the past few decades, lung sounds detection has been a significant focus of research in the bioinformatics field. Lung sounds convey essential information about respiratory diseases, and to diagnose cases with pulmonary disorders, practitioners use traditional statoscopes to auscultate lung sounds. This technique, however, has many restrictions. For example, if the physician is not adequately trained, this may result in an incorrect diagnosis. Furthermore, lung sounds signals are non-stationary, complicating the analysis and recognition process. From these points, research on identifying the ability of CNN to classify lung sounds can aid in overcoming these constraints. In this research, we applied three different CNN approaches for classifying lung sounds and used the ICBHI 2017 database, the most popular and largest available database. In the first approach, we extracted spectrogram images using Short-time Fourier transform and used weights from ImageNet, then fed spectrograms into VGG16 model. In the second and third approaches, we performed three different data augmentation techniques (Noise, Pitch Shifting, and Dynamic change compression) on all classes, then extracted MFCC features then fed features into Dense CNN; in the third approach, data augmentation was only made on the types that have fewer samples. The validation accuracy obtained 83%,91%, and 93 %, respectively, for the three mentioned approaches.