Lung Sound Diagnosis with Deep Convolutional Neural Network and Two-Stage Pipeline Model

摘要

Lung sounds are very critical in the diagnosis of pulmonary disease clinically. The study of their recognition using computers is considered to be meaningful for doctors. In this paper, we proposed two methods for identifying wheeze, crackle, and normal lung sounds. We first formulate the lung sound identification problem mathematically. And then, we propose a deep convolutional neural network (CNN) model which is consisted of 9 layers (6 conv layers, 3 pooling layers, and 3 fully connected layers). Lung sound segments are extracted to obtain feature bands from log-scaled mel-frequency spectral (LMFS) and are constructed into feature maps byways of bands by frames. The second method is a two-stage pipeline model (TSPM) which is the extension of Gaussian mixture model. Forty-six traditional features are extracted and selected for our TSPM. Testing on our lung sound database recorded from a local hospital, we find that chroma features are the most important to TSPM and the F1 scores of 46 features on three types show an obvious improvement when it is compared with 24 MFCC which are shown to be the optimal features for wheeze recognition in the previous literature. And we finally find that CNN model has better recognition performance than TSPM, because F1 scores about CNN model's testing are 0.8516 for wheeze, 0.8471 for crackle, and 0.8571 for normal sound.