Detecting Respiratory Diseases from Recorded Lung Sounds by 2D CNN

Respiratory disease is among the leading causes of deaths around the world. A large amount of population is being affected regularly with some kinds of lung function disorders which eventually lead to respiratory diseases. Prevention and early detection are essential steps in managing respiratory diseases. To decrease the fatality, an efficient detection model is needed. In this paper, 2D convolutional neural network (CNN) is used to detect respiratory diseases from the recorded lung sounds at early stages. The proposed method can detect respiratory diseases like bronchiectasis, pneumonia, bronchiolitis, chronic obstructive pulmonary disease, upper respiratory tract infection, and healthy by using Mel-frequency cepstral co-efficients (MFCC). In the proposed scheme, a data frame is recorded and after extracting the statistical features from the audio clips, the data is loaded in the data frame where further classification is done using 2D CNN. The model is based on 2D CNN architecture where the number of layers is reduced to a certain extent to achieve more accuracy. The proposed model has only 13 CNN layers where each convolution layer is being associated with a pooling layer of max-pooling 2D type. The final convolution layer has a global-average pooling 2D layer. The proposed method obtained an accuracy of over 92.39%.