Feature Extraction and Classification of Respiratory Sound and Lung Diseases

Abstract

Bacteria, viruses, and fungi can cause respiratory infections. It is usually possible to detect respiratory diseases early by listening to the lung sounds with a stethoscope. In reality, lung sound analysis is a time-consuming and difficult task that depends on medical skills and recognition experience. Recent advances in automatic respiratory sound recognition and classification have attracted more attention. The outbreak of COVID-19 throughout the world and the high patient numbers have placed a great deal of pressure on medical professionals. A smart algorithm is therefore a necessity to provide a faster and more accurate detection of lung infections by automatically processing the sounds of the lungs. This paper proposes two new lung sound feature extraction, maximum entropy Gabor filter bank (MAGFB), and maximum entropy Mel filter bank (MAMFB). The classification is performed by a deep neural convolution network (DCNN) by using 50% of data for training the classifier. The filter banks have been substituted, instead of the convolutional layers. Experiments were conducted on the ICBHI 2017 Challenge dataset (with eight classes). The proposed method has a better performance compared to famous methods such as MFCC and Wavelet transform. Particularly, the performance of the second method is significant. For ICBHI 2017 challenge dataset, the overall accuracy of MFCC, Wavelet, MAGFB and MAMFB were 87%, 86%,90% and 93%, respectively.