When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.
Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.
Image registration is a crucial step in many medical image processing pipelines. The process aligns images of the same tissue taken at different times or with different imaging modalities. The first focus of this thesis is on the registration of ultrasound (US) images, which are low-cost, portable, safe, real-time, and commonly employed in several image-guided operations. Image registration of intraoperative US with preoperative images is required in image-guided surgeries. Computed Tomography (CT) scans and Magnetic Resonance Imaging (MRI) generally visualize the bones and soft tissues with better spatial details than US. Therefore, surgeons and interventionalists prefer them to US for the preoperative planning. These preoperative images should be registered to the intraoperative US images in image-guided interventions, which is a challenging task and an open area of research. Beyond image-guided interventions, image registration is a critical step in several other medical image analysis pipelines. The second focus of this work is on inter-contrast CT and MRI registrations. MRI is the primary modality for diagnosing neurodegenerative diseases such as Alzheimer’s Disease. MRI comes with various contrasts, and the fusion of these contrasts taken at different times or from many subjects can give clinicians valuable information. However, MRI has a longer waiting time and less availability than CT. Thus, designing inter-modal image registration techniques to align MRI data with CT scans is essential in medical image analysis. Novel methods to tackle this problem are proposed in this thesis. The traditional image registration methods, which solve an optimization problem iteratively, can be time-inefficient for analyzing large datasets. Image registration using Deep Learning (DL) can accelerate the process but usually require training data. In this thesis, several novel methods for performing inter-contrast image registration are proposed in Chapters 3 to 5. These methods span both energy- and DL-based techniques with DL-based methods being more computationally efficient. We conclude the thesis in Chapter 6 by providing possible future research directions.