Title

AI-Driven 3D Segmentation of COVID-19 Infection from Lung CT

Document Type

Dissertation

Abstract

Coronavirus disease 2019 (COVID-19) has been declared by the World Health Organization (WHO) as a global pandemic in 2020 that led to over 6.55 million deaths up to date [51]. The disease is caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV- 2) targeting the body’s lungs. The diagnosis of the disease is carried out following a conventional test known as reverse-transcription polymerase chain reaction (RT-PCR). Unfortunately, RT-PCR test results are often associated with delays. That being said, instant and accurate diagnosis is crucial to effectively implement due procedures in order to tend to diagnosed patients as well as close contact cases, while simultaneously ensuring that health authorities are able to put in place all preventative measures to reduce mortality rates and impede disease spread. In order to automate and speed up COVID-19 diagnosis, an Artificial Intelligence (AI) based approach based on lung Computed Tomography (CT) scans is proposed. This will: 1) provide quick and accurate patient isolation decision in parallel with RT-PCR test, 2) allow for infection severity quantification 3) yield more efficiency and circumvent the issue of radiologists scarcity/availability required for CT analysis. In this thesis, we investigate multiple known deep learning approaches for segmenting COVID-19 infection from 3D lung CT scans using labeled images from COVID-19 Lung CT Lesion Segmentation Challenge - 2020. We analyze the U-Net-based baseline method, in addition to exploring Convolution Neural Network (CNN) based methods, Transformer based methods, multi-class learning, self-supervised pre-training, and self- configuring methods for 3D lesion segmentation. We propose to use nnU-NET for 3D semantic segmentation of COVID-19 infection from lung CT scans. nnU-Net, in particular, streamlines and automates the crucial choices needed to create an effective segmentation pipeline. The proposed approach achieves superior performance over the challenge baseline and the rest of the methods in identifying COVID-19 infection with sensitivity, specificity, and precision of 0.8166, 0.7426, 0.7603 respectively. Furthermore, it ranked 2nd place when evaluated in the COVID-19 Lung CT Lesion Segmentation Challenge - 2020 for the post-challenge leader-board with a Dice score of 0.7525 ± 0.19 by the date of thesis submission.

First Page

i

Last Page

54

Publication Date

12-30-2022

Comments

Thesis submitted to the Deanship of Graduate and Postdoctoral Studies

In partial fulfillment of the requirements for the M.Sc degree in Machine Learning

Advisors: Dr. Mohammad Yaqub, Dr. Huan Xiong

Online access for MBZUAI patrons

Link to Full Text

Share

COinS