Year: 2008
Type: dissertation
Abstract: The availability of numerous complementary imaging modalities allows us to obtain a detailed picture of the body and its functioning. To aid diagnostics and surgical planning, all available information can be presented by visually aligning images from different modalities using image registration. This dissertation investigates strategies to improve the performance of image registration algorithms that use intensity-based similarity metrics. Nonrigid warp estimation using intensity-based registration can be very time consuming. We develop a novel framework based on importance sampling and stochastic approximation techniques to accelerate nonrigid registration methods while preserving their accuracy. Registration results for simulated brain MRI data and human lung CT data demonstrate the efficacy of the proposed framework. Functional MRI (fMRI) is used to non-invasively detect brain-activation by acquiring a series of brain images, called a time-series, while the subject performs tasks designed to stimulate parts of the brain. Consequently, these studies are plagued by subject head motion. Mutual information (MI) based slice-to-volume (SV) registration algorithms used to estimate time-series motion are less accurate for end-slices (i.e., slices near the top of the head scans), where a loss in image complexity yields noisy MI estimates. We present a strategy, dubbed SV-JP, to improve SV registration accuracy for time-series end-slices by using joint pdf priors derived from successfully registered high complexity slices near the middle of the head scans to bolster noisy MI estimates. Although fMRI time-series registration can estimate head motion, this motion also spawns extraneous intensity fluctuations called spin saturation artifacts. These artifacts hamper brain-activation detection. We describe spin saturation using mathematical expressions and develop a weighted-average spin saturation (WASS) correction scheme. An algorithm to identify time-series voxels affected by spin saturation and to implement WASS correction is outlined. The performance of registration methods is dependant on the tuning parameters used to implement their similarity metrics. To facilitate finding optimal tuning parameters, we develop a computationally efficient linear approximation of the (co)variance of MI-based registration estimates. However, empirically, our approximation was satisfactory only for a simple mono-modality registration example and broke down for realistic multi-modality registration where the MI metric becomes strongly nonlinear.
Author: Roshni Bhagalia
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2024421978