Classification of Auscultation Sounds into Objective Spirometry Findings using MVMD and 3D CNN

Millions of people suffer from respiratory illness globally. Early diagnosis of respiratory diseases is hindered because of the lack of cost-effective and simple methods. Spirometry is the pulmonary function test used for diagnosis of obstructive diseases like asthma, chronic obstructive pulmonary disease (COPD) and restrictive diseases like interstitial lung disease (ILD), etc. This test requires repeated manoeuvre, is expensive and is done in laboratory which are not available in resource poor areas. Auscultation is an easy and cost-effective method which can play a vital role in early diagnosis of respiratory diseases. In this paper, a technique is proposed which could classify auscultation sounds into normal, obstructive and restrictive disease category similar to the findings of spirometry. The proposed work uses combination of multivariate variational mode decomposition and dynamic time warping for enhancing multi-channel signal. Further, pre-trained 3D ResNet18 neural network model is used for classification into three classes. Encouraging results are achieved with accuracy of 94.57%, sensitivity of 100% and specificity of 94.11%.