Jan Rune Rühaak

• This PhD thesis presents a generic approach to more efficient image registration in a classical optimization-based framework. The calculation of the objective function derivatives is targeted as a whole, thereby fusing the computations for image transformation, interpolation and image similarity to one explicit, matrix-free expression. The employed concept reduces the auxiliary space requirements for derivative computation from linear to constant and greatly facilitates parallel execution. We study its usage both for affine-linear and deformable image registration algorithms designed for monomodal and multimodal problems. Additionally, implementations on alternative platforms such as graphics processing units or embedded digital signal processors are discussed. The practical value of the considered approach is illustrated on real-world applications from various areas of medical image computing. In particular, pulmonary image registration algorithms are studied for two use cases: computer assistance for lung cancer screening and oncological follow-up assessment, and quantitative analysis for chronic obstructive pulmonary disease. For both applications, efficient solutions are presented that combine state-of-the-art accuracy with clinically acceptable runtime and memory footprint. Our algorithms are extensively evaluated on three publicly available, expert-annotated data collections. The thesis is completed by a study of three further medical applications of image registration: alignment of positron emission tomography (PET) with computed tomography (CT) scans for joint radiological reading, registration of magnetic resonance and CT images in the context of deep brain stimulation, and real-time tracking of liver vessels in long ultrasound sequences for motion compensation. In all three applications, the proposed matrix-free calculation schemes contribute to moderate resource demands, which allow for unproblematic execution in a clinical environment.