Document Type
Thesis
Date of Award
2022
Degree Name
Master of Science (MS)
Department
Computer Science
First Advisor
KC Santosh
Abstract
Despite the general success of employing artificial intelligence (AI) to help radiologists perform computer-aided patient diagnosis, building machine learning models with limited datasets at different sites is not trivial. In addition, medical imaging is no exception. In medical imaging informatics, precise detection of lung disease helps clinicians treat patients effectively while averting possible fatalities. To effectively build machine learning models, we propose to study federated learning mechanisms, so we can use/learn datasets from different sources/regions/sites. . Individual sites may jointly train a global model using this approach, referred to as federated learning. In other words, without explicitly sharing datasets, federated learning (FL) combines training results from various sites to produce a global model, making sure we respect patients’ confidentiality across all sites. Additionally, the additional supervision gained from partner sites' results enhances the global performance, as our primary goal is to determine how the FL mechanism offers an improvement of machine learning models (robust, accurate, and unbiased in decision-making). To validate this goal, using 324 lung sound audio recordings (Source: King Abdullah University Hospital, https://data.mendeley.com/) in the form of Mel-spectrograms, a Convolutional Neural Network, which we call Federated-Net model, was used to classify them between healthy and unhealthy. In our experiments, we achieved a validation accuracy of 71.88% and a test accuracy of 59.96%. With this, this study also provided challenges to federated learning adoption and potential future benefits.
Subject Categories
Computer Sciences
Keywords
Despite the general success of employing artificial intelligence (AI)
Number of Pages
78
Publisher
University of South Dakota
Recommended Citation
Farjana, Afia, "Federated Learning for Lung Sound Analysis" (2022). Dissertations and Theses. 304.
https://red.library.usd.edu/diss-thesis/304
