Author ORCID Identifier
Date of Award
Master of Science (MS)
KC KS Santosh
In this paper, we analyze deep visual features from 2D data representation(s) of the respiratory sound to detect evidence of lung abnormalities. The primary motivation behind this is that visual cues are more important in decision-making than raw data (lung sound). Early detection and prompt treatments are essential for any future possible respiratory disorders, and respiratory sound is proven to be one of the biomarkers. In contrast to state-of-the-art approaches, we aim at understanding/analyzing visual features using our Convolutional Neural Networks (CNN) tailored Deep Learning Models, where we consider all possible 2D data such as Spectrogram, Mel-frequency Cepstral Coefficients (MFCC), spectral centroid, and spectral roll-off. In our experiments, using the publicly available respiratory sound database named ICBHI 2017 (5.5 hours of recordings containing 6898 respiratory cycles from 126 subjects), we received the highest performance with the area under the curve of 0.79 from Spectrogram as opposed to 0.48 AUC from the raw data from a pre-trained deep learning model: VGG16. We also used machine learning algorithms using reliable data to improve Our study proved that 2D data representation could help better understand/analyze lung abnormalities as compared to 1D data. Our findings are also contrasted with those of earlier studies. For purposes of generality, we used the MFCC of neutrinos to determine if picture data or raw data produced superior results.
2D Data Representation, Deep Visual Features, Lung Abnormality, Respiratory Sound
University of South Dakota
Alice, Rafia Sharmin, "2D respiratory sound analysis to detect lung abnormalities" (2022). Dissertations and Theses. 105.