Date of Award
7-18-2024
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
Three-dimensional deformable image registration (DIR) is an important technique in medical imaging, facilitating the alignment of images acquired from different sources or at different times for clinical diagnosis and treatment planning. In clinical proton therapy dose planning, dual-energy computed tomography (DECT) has notably improved the accuracy of estimating the stopping-power ratio map, reducing proton beam range uncertainty. To address the inter-scan motion artifacts inherent in sequential DECT acquisition, we propose a novel motion-compensation scheme to incorporate a three-dimensional DIR method into the joint statistical iterative DECT reconstruction algorithm. We successfully demonstrate that inter-scan motion corrections can be integrated into the DECT statistical iterative reconstructions process, enabling accurate imaging of radiological quantities on conventional single-energy CT scanners, without significant loss of either computational efficiency or accuracy. We enhance DIR methods using deep-learning-based approaches, with a specific emphasis on improving registration accuracy and error estimation critical for clinical applications. We introduce a 3D deep-learning-based multi-modality image registration correction network (RCN) designed to automatically assess DIR accuracy using dense target registration error (TRE) estimation. The RCN refines the initial result using the dense TRE map, improving registration accuracy significantly. Furthermore, we introduce an unsupervised inverse-consistent diffeomorphic registration network termed IConDiffNet, which achieves rapid and accurate DIR on a 3D image registration task for pairs of images. Compared to state-of-the-art deep-learning (DL)-based diffeomorphic DIR methods, IConDiffNet demonstrates better performance on a large-scale brain MRI image dataset containing 375 subjects with better Dice scores, lower Hausdorff distance metrics, and reduced total energy expenditure during deformation. Visualization results highlight IConDiffNet's ability to produce intricate diffeomorphic transformations that effectively align structures.
Language
English (en)
Chair
Joseph O’Sullivan