University of Twente Student Theses


Neural implicit representations for diffeomorphic medical image registration

Zwienenberg, Jesse (2021) Neural implicit representations for diffeomorphic medical image registration.

Full text not available from this repository.

Full Text Status:Access to this publication is restricted
Embargo date:1 January 2025
Abstract:Image registration is the task of finding a transformation that aligns given images. Typically, this is solved using discrete image registration algorithms. These discrete algorithms are inherently limited by the resolution of the images. Rather than representing (3D-) images explicitly using a discrete set of pixels (or voxels), they can also be represented implicitly by a neural network. Not only the images can be implicitly represented, but also the transformation itself. One of the advantages of the implicit representations in registration models is that the resulting transformation is defined on any coordinate in the continuous image domain, rather than only on a discrete set of coordinates. This means there is no need for interpolation anywhere, and the image derivatives can easily be computed analytically rather than through finite differences. Also, both the transformation and the images can be sampled at arbitrarily high resolutions. In this thesis, we propose two models for medical image registration in a continuous setting, a small deformation model and a diffeomorphic model. For both models, the image-pair and the transformation are continuous differentiable functions parametrized by neural networks. The models are evaluated on a dataset of 3D CT scans with manually identified landmarks.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:31 mathematics, 54 computer science
Programme:Applied Mathematics MSc (60348)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page