Ahmadvand, Payam - Machine Learning Driven Active Surfaces for 3D Segmentation of Tumour Lesions in PET Images...

View the thesis

This thesis has been submitted to the Library for purposes of graduation, but needs to be audited for technical details related to publication in order to be approved for inclusion in the Library collection.
Term: 
Summer 2017
Degree: 
M.Sc.
Degree type: 
Thesis
Department: 
School of Computing Science
Faculty: 
Applied Sciences
Senior supervisor: 
Ghassan Hamarneh
Thesis title: 
Machine Learning Driven Active Surfaces for 3D Segmentation of Tumour Lesions in PET Images
Given Names: 
Payam
Surname: 
Ahmadvand
Abstract: 
One of the key challenges facing wider adoption of positron emission tomography (PET) as an imaging biomarker of disease is the development of reproducible quantitative image interpretation tools. Quantifying changes in tumor tissue, due to disease progression or treatment regimen, often requires accurate and reproducible delineation of lesions. Lesion segmentation is necessary for measuring tumor proliferation/shrinkage and radiotracer-uptake to quantify tumor metabolism. In this thesis, we develop an active surface model for segmenting lesions from PET images. We first implemented a non-convex level set active surface method with likelihood terms trained on manually-collected seeds points. We evaluated this approach on the following datasets: Images of phantoms collected by our collaborators at UBC; Quantitative Imaging Network (QIN) phantom images; and images of real patients from the QIN Head and Neck challenge. Secondly, to avoid user interaction, we developed an improved version of our method by training a machine learning system on anatomically and physiologically meaningful imaging cues to distinguish normal organ activity from tumorous lesion activity. Then, the inferred lesion likelihoods are used to guide a convex active surface segmentation model. The result is a lesion segmentation method that does not require user-initialization, manual seeding, or parameter-tweaking and, thus, guaranteeing reproducible results. We tested this enhanced method on data from the Cancer Imaging Archive. Our method not only produces more accurate segmentation than state-of-the-art segmentation results, but also does not need any user interaction.
Keywords: 
Machine Learning, Segmentation, Active Contour Model, Functional Imaging, Positron Emission Tomography (PET), Head and Neck Cancer
Total pages: 
88