Dissertation Defense of Darko Zikic
: Contributions to Medical Image Registration: Theoretical Insights and New Methodologies
Image registration, also known as image alignment, spatial
normalization, or motion estimation, is the process of computing the
spatial transformation between corresponding structures of objects
depicted in two or more images. In medical scenarios, the spatial
transformation between corresponding points can have various causes,
such as natural patient movement, different patient positioning, or
anatomical changes. Sometimes, the apparent misalignment is caused by
the task of comparing images of different subjects. Thus, image
registration is often the foundation for further analysis and
applications, such as fusion of different modalities for diagnosis and
Image registration presents an active field of research, among other
reasons also due to the complex nature of the associated optimization
problem. Among the registration methods, the so called deformable
registration problem, which treats high-dimensional transformation
models, is particularly challenging.
In the first part of this work, we present an overview of existing
intensity-based registration methods, with focus on deformable
approaches. Our aim is to point out the relations between the different
approaches, and to highlight the common elements.
In the second part of this thesis, we present several of our
contributions to the field of registration of medical images, most of
them affecting deformable registration methods:
1) We propose a simple and efficient preconditioning scheme for
improvement of the convergence speed of gradient-based methods for
arbitrary image-based difference measures in deformable registration.
The proposed scheme is especially useful for deformable multi-modal
registration employing statistical difference measures, since in these
cases, the range of applicable efficient optimization schemes is
strongly limited, due to the structure and size of the problem.
2) Alignment of angiographic 3D scans to 2D projections is an important
issue for navigation during interventions. For the common single-view
setting, in which only one 2D projection is available, we introduce a
method for deformable 2D-3D registration of vascular structures. Prior
to our work, methods employing a rigid transformation model presented
the state of the art for this problem.
3) For linear registration, which is often performed prior to deformable
registration, we discuss the use of Markov Random Field (MRF) modeling
and discrete optimization. The approach is based on the approximation of
the original energy, such that efficient discrete optimization becomes
applicable. The resulting framework allows to transfer the advances from
the active field of discrete optimization to linear registration.
4) For generation of statistical deformation models, we analyze the
deformations resulting from deformable registration and find that they
contain significant linear components. Based on this observation, we
propose the use of minimal deformations, from which the linear
transformation components are removed. We demonstrate that this step is
non-optional for creation of accurate models representing the major
deformations within a population.
: Prof. Cremers, TUM
: Prof. Navab, TUM
: Prof. Rueckert, Imperial College London
: Tuesday, December 20, 2011
: MI 02.09.023
Technische Universität München
Fakultät für Informatik
85748 Garching b. München