Use this URL to cite or link to this record in EThOS:
Title: Robust methods for medical image registration with application in clinical diagnosis
Author: Santos Ribeiro, Andre Filipe
ISNI:       0000 0004 7229 014X
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Access from Institution:
Automated analysis of medical imaging data allows both researchers and clinicians to develop more accurate and faster diagnoses. Image registration plays an essential role in both longitudinal studies and group analysis, allowing the combination of different imaging modalities, and automatic parcellation of regions of interest. Despite its wide use, image registration is still challenging with many issues such as artefacts, scarcity of correspondences, multi-modality, and computational complexity. Additionally, due to the lack of highly annotated datasets, the validation of image registration, and specifically non-linear registration, is also problematic. In this thesis several of these issues are addressed by introducing: a framework to validate non-linear registration methods; a robust and fast algorithm for non-linear registration; and validating the proposed methods in a conventional analysis. Current techniques used for non-linear image registration validation are explored, and it is shown that techniques based on label overlap are both not theoretically valid while also having poor accuracy. This analysis further leads to the development of a multiscale metric to minimize these problems. Also, a method based on the Demons Framework is proposed to improve the convergence speed of non-linear registration algorithms, and further extended to be robust in the presence of intensity inhomogeneities and contrast variations. The proposed methods are validated in a synthetic simulation platform with a known ground truth, compared with a manually traced region-of-interest, and tested in a voxel-based morphometry analysis of real data. It is shown that the proposed methods outperform other leading registration methods in both the synthetic simulation study and the manually traced data, and present reliable results in the voxel-based morphometry analysis. Furthermore, the impact of different registration algorithms is explored through the voxel-based morphometry study, and shown to affect the final results and their interpretation.
Supervisor: Nutt, David ; McGonigle, John Sponsor: Edmond J. Safra Philanthropic Foundation
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral