Date of Award
Spring 5-15-2023
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
Nuclear-medicine imaging (PET/SPECT) provides a mechanism to study in vivo physiological processes in the human body and, thus, plays an important role in cancer diagnosis, prognosis, and treatment. Recently, quantitative imaging (QI), i.e., the procedure to extract numerical or statistical features from medical images, is showing substantial promise in multiple clinical applications. However, clinical advancement of quantitative PET/SPECT requires methods for improved quantification from images and techniques to objectively evaluate these methods with patient data. The goal of this dissertation is to develop methodologies to fulfill these needs.Reliable quantification of features (metabolic tumor volume, regional tracer uptake) from PET/SPECT images often requires accurate segmentation of regions of interest. However, segmentation in PET/SPECT is challenging, a major reason being partial volume effects (PVEs). The PVEs arise from the limited spatial resolution of imaging systems and reconstruction of images over finite-sized voxel grids. The latter results in tissue-fraction effects (TFEs), i.e., a voxel can contain a mixture of different tissue types. Existing segmentation methods typically yielded limited performance due to their sensitivity to PVEs and, in particular, inability to model the TFEs. To address this challenge, we develop a deep-learning-based tissue-fraction estimation method (E-Seg-Net) that accounts for both the sources of PVEs when performing segmentation. We demonstrate the efficacy of the proposed E-Seg-Net to improve segmentation and quantification from PET/SPECT images of patients with lung cancer/Parkinson disease.Clinical translation of quantitative PET/SPECT methods also requires assessing their performance on clinically relevant tasks. However, image-segmentation methods have typically been evaluated using metrics that are not explicitly designed to correlate with clinical-task performance. We show that evaluating PET segmentation methods based on such task-agnostic metrics can lead to interpretations that are discordant with evaluation on specified clinical tasks. This finding emphasizes the need for task-specific evaluation of QI methods.One approach for conducting such evaluation is through in silico imaging trials. However, in these trials, it is important that the synthetic images must be clinically realistic. Thus, mechanisms that can quantitatively evaluate this clinical realism are much needed. Towards this goal, we develop and demonstrate the utility of a web-based software to facilitate the conducting of this evaluation with expert human observers. Further, we present a theoretical formalism for the use of an ideal observer to quantitatively evaluate the similarity in distributions between real and synthetic images. A second and highly desirable approach is to conduct task-specific evaluation of QI methods directly with patient data. However, this is often hindered by the lack of available gold standards. To address this challenge, we develop a no-gold-standard evaluation (NGSE) technique that objectively evaluates QI methods in the absence of gold standards. We demonstrate the ability of the proposed NGSE technique to objectively evaluate PET segmentation methods even without any knowledge of the true quantitative values.
Language
English (en)
Chair
Abhinav K. Jha
Committee Members
Pamela K. Woodard, Clifford G. Robinson, Yuan-Chuan Tai, Joyce C. Mhlanga,