Multi Spectral Feature Extraction to Improve Lung Sound Classification using CNN

The principal method for screening and diagnosing lung disorders is auscultation of respiratory sounds. In the medical field, automated auscultation is a new research field that helps with faster disease diagnosis. The task of automated auscultation is to classify the respiratory sound, which requires a dataset to build a classification model. Deep learning networks are a popular method for producing such classification models, which require a large number of samples to train. Unfortunately, even the largest publicly accessible respiratory sound dataset (ICBHI 2017) of 5.5 hours total length contains only 6898 respiratory cycles, which is still insufficient for training deep learning networks. In this research work, data augmentation and smart padding techniques are utilised to account for the ICBHI dataset’s class imbalance. Further, we have suggested to extract several features from the audio signal, such as MFCC, Chroma STFT, Mel-Spectrogram, Spectral Contrast, and Tonnetz, for better representation of the samples in a class. Further, we have obtained the statistical features such as mean and standard deviation for developing the classification model. We have utilised a deep learning-based CNN architecture for the classification of the learned features. The proposed method achieved a score of 74.08% for a 60-40 train-test split and 75.04% for an 80-20 train-test split, outperforming all state-of-the-art lung sound classification methods.