Computer Science ETDs

Publication Date

Fall 12-17-2022

Abstract

A staggering amount of data has been collected and analyzed since the beginning of the COVID-19 pandemic. Some of that data, however, particularly computed tomography (CT) scans of lungs, are difficult to analyze computationally. Segmentation of COVID-19 lesions provides researchers with insights into where lesions form, their volume, and in the case of time course data, how lesions grow as disease progresses. Spatial information, which has received little attention in the literature, has the potential to provide valuable insight to within-host viral and immune dynamics. Unfortunately, state-of-the-art supervised deep learning methods require labeled data, which can be prohibitively expensive to obtain. In this work, LENS (Lung lEsioN Segmentation), a novel image processing-based method, is proposed to segment COVID-19-positive lung CT scans without the need for labeled data. This method is novel both because it modifies the lung CT-specific algorithms it employs and because it combines several lung CT-specific algorithms to segment lesions within the lung into a complete, fully-automated pipeline. LENS first segments the lung from the CT, then segments and removes the bronchial tree from the lung. Removing the bronchial tree reduced percent error from 2262.4% to 42%, a 54-fold reduction. LENS demonstrates that low-cost, unsupervised segmentation of COVID-19 lesions can be used successfully in research applications.

Language

English

Keywords

lung CT, segmentation, image processing, COVID-19

Document Type

Thesis

Degree Name

Computer Science

Level of Degree

Masters

Department Name

Department of Computer Science

First Committee Member (Chair)

Melanie Moses

Second Committee Member

Judy Cannon

Third Committee Member

Shuang Luan

Download

Share

COinS