Topology preserving nonlinear image registration

D. Yoan L. Mekontchou Yomba

doi:10.17918/xswr-bk19

Image Registration is an algorithmic optimization process primarily aimed at estimating the most optimal transformation from one image's coordinate plane onto another. Nonlinear image registration has been the subject of numerous research initiatives. This widespread focus on the topic has resulted in a variety of different registration methods, intrinsic novel findings, as well as application areas. Although many potential use cases exist, the main three typical applications for 2D and 3D registration methods include object tracking in a series of scans, interpatient registrations, and multi-modal registration. Substantial research efforts have also begun to observe the registration process in both constrained and unconstrained settings as well as applied focused work on the regularization and homeomorphic behavior of vector fields for more accurate registration results. In this thesis, we compare the registration accuracy of our log unbiased multi-resolution registration model against the model provided by Igor et al. [16]. We also compare the effects of various objective functions on the ultimate accuracy of our multiresolution registration framework. Additionally, we delineate the short-comings of the log unbiased model [16] and mathematically propose a method to further accommodate these shortcomings as well further enforce topology preservation and smoothness in the displacement fields in hopes of making the model more robust specifically in a continuous setting. Our multi-resolution method utilizes information theory to quantify the magnitude of deformations, generates intuitively correct deformation maps, and correctly represents non-smooth transformations. To demonstrate the efficacy of the proposed framework, we generalize the well known large-deformation viscous fluid model. We are able to show that our proposed multi-resolution scheme generates more accurate transformation maps compared to those generated using the plain viscous fluid model with fluid regularization. We examine the power of this model to detect real changes specifically as we compare its use with different matching functionals. The results presented in this work show that the unbiased methods have higher reproducibility and accuracy than conventional registration models. Numerical results are presented in chapter 4, 5, and 6.

Topology preserving nonlinear image registration

Files and links (1)

Abstract

Metrics

Details

Topology preserving nonlinear image registration

Files and links (1)

Abstract

Metrics

Details

Drexel University Social media