Document Type
Thesis
Date of Award
2022
Degree Name
Master of Science (MS)
Department
Computer Science
First Advisor
KC Santosh
Abstract
Artificial Intelligence (AI) has contributed a lot since the beginning. Healthcare is no exception. Detecting anomaly/abnormality in (bio)medical image is crucial. In this thesis, we aim at detecting/screening pulmonary abnormalities due to Covid-19 in chest X-rays using deep features. We study CheXNet, DenseNet169, ResNet50 and VggNet16 to analyze CXRs to detect the evidence of Covid-19 in this research. CheXNet was primarily designed for radiologist-level pneumonia detection in Chest X-rays (CXRs). We created a benchmark dataset size of 4,716 CXRs (2,358 Covid-19 positive cases and 2,358 non-Covid cases (Healthy and Pneumonia cases)) and with k(=5) fold cross-validation technique, using the DenseNet, we achieved the best performance: accuracy of 0.99, AUC of 0.99, specificity of 0.99 and sensitivity of 0.98. Similar results were achieved from other deep features extracted from CheXNet, ResNet50, and VggNet16. Our results are compared with other state-of-the-art results. Further, since deep features independently are not themselves explainable, we developed XAI algorithm that aimed at explaining the decisions of the DNN models via pathology localization/visualization in chest X-rays. For this, we generated corresponding heatmap from chest X-rays that helped explain deep features representing the evidence of the pulmonary abnormalities due to Covid-19.
Subject Categories
Computer Sciences
Keywords
Artificial Intelligence, Chest X-rays, CheXNet, Deep Learning, DenseNet, Healthcare
Number of Pages
62
Publisher
University of South Dakota
Recommended Citation
Ghosh, Supriti, "DEEP FEATURES TO ANALYZE PULMONARY ABNORMALITIES IN CHEST X-RAYS DUE TO COVID-19" (2022). Dissertations and Theses. 40.
https://red.library.usd.edu/diss-thesis/40