Respiratory Sounds Feature Learning with Deep Convolutional Neural Networks

In this paper, we develop a computer-based solution for automatic analysis of respiratory sounds captured using the stethoscope, which has many potential applications including telemedicine and self-screening. Three types of respiratory sounds (e.g. wheezes, crackles, and normal sounds) are captured from 60 patients by a custom-built prototype device. Then we propose a deep Convolutional Neural Networks (CNN) model consisting of 6 convolutional layers, 3 max pooling layers and 3 fully connected layers and optimize its structure. The model is used for automatically learning features of respiratory sounds and identifying them. Through time-frequency transformation, Log-scaled Mel-Frequency Spectral (LMFS) features of 60 bands are extracted frame by frame from the dataset and divided into segments in the size of 23 consecutive frames as inputs of the model. Finally, we test the model by 12 new subjects' dataset and compare it with mean performance of 5 respiratory physicians in both precision and recall. The testing result shows that our CNN model achieves the same of level of identifying accuracy as the respiratory physicians. To the best of our knowledge, this is the first study to apply CNN method to assess medical fields about respiratory sounds.