Gammatonegram based triple classification of lung sounds using deep convolutional neural network with transfer learning

Abstract

Respiratory diseases are the leading cause of death worldwide; their timely diagnosis is essential. The primary tool for diagnosis of respiratory disorder is auscultation using a stethoscope. This conventional technique is subjective and relies on doctor’s experience. An efficient clinical support system by converting subjective listening process of auscultation into computerized proficient auscultation is the need of time. In practical environment auscultation sounds are overlapped by different noises therefore in order to classify them, we need an efficient de-noising technique followed by a classification model. In this paper, a novel pre-processing technique has been proposed that de-noise respiratory sounds using variational mode decomposition(VMD) technique and fed these sound signals to gammatone filter bank to generate time frequency distribution in the form of Gammatonegram images. These Gammatonegram images are classified using different deep convolutional neural network architecture based transfer learning models. Large data set for respiratory sounds are unavailable, CNN model over-fit if the size of data set is small, Therefore transfer learning like AlexNet, GoogLeNet, ResNet-50 and Inceptionv3 have been used for lung sound classification. The proposed method can classify lung sounds into three classes with accuracy, precision, sensitivity and specificity of 98.8%, 97.7%,100% and 97.6%.