PhD Oral Exam - Soorena Salari, Computer Science

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.

Abstract

Image registration is a fundamental task in computer vision and image processing that aligns two images so that their corresponding features can be spatially matched. Accurate medical image registration is crucial in radiological diagnosis, surgical planning, image-guided interventions, radiotherapy, and large-scale neuroimaging studies, where it directly affects patient safety, treatment outcomes, and the reliability of image-based disease analysis. For example, in surgery and radiotherapy, tissue deformation caused by gravity, physiological motion, resection, or patient movement can shift surgical targets and vital organs from their planned locations. Registration algorithms can help track these deformations by aligning pre-operative, intra-operative, or follow-up scans. However, automatic registration algorithms may still produce inaccurate or unreliable results, while quality control still largely relies on subjective visual inspection. Therefore, effective automatic techniques to assess and visualize registration quality are highly valuable but are under-explored.

This thesis proposes several deep learning-based methods for the automatic quantification of medical image registration quality. In particular, it develops and validates new methodologies and software tools to estimate the accuracy, uncertainty, and reliability of medical image registration, with an emphasis on inter-modal and inter-contrast image registration (e.g., MRI vs. ultrasound, T1w MRI vs. T2w MRI). Chapters 2 and 3 address the need for reliable anatomical correspondences by introducing self-supervised brain landmark detection and discovery frameworks that can detect landmarks across MRI contrasts with minimal manual annotation. Chapter 4 focuses on direct registration quality estimation by proposing uncertainty-aware models that predict MRI-ultrasound registration error locally, including both patch-wise and voxel-wise error estimation. Finally, Chapter 5 addresses the robustness and interpretability of registration by introducing a landmark-driven motion tracking framework that uses explicit anatomical correspondences and foundation-model features to improve alignment under challenging motion and misalignment conditions. Together, these contributions advance registration quality control beyond subjective visual inspection toward quantitative, spatially resolved, and interpretable assessment.