Brain Magnetic Resonance Image AnalysisSegmentation and Registration

Authors Organisations
Type

Student thesis: Doctoral ThesisDoctor of Philosophy

Original languageEnglish
Awarding Institution
Supervisors/Advisors
Award date2017
Links
Show download statistics
View graph of relations

Abstract

Automatic and robust 3D brain magnetic resonance (MR) image analysis can assist in disease diagnosis, surgical planning and patient follow-up. As the most crucial techniques for brain image analysis, brain MR image segmentation benefits the pathology detection, volumetric morphometry, surface reconstruction and 3D visualisation; brain image registration helps with multi-modality image fusion, longitudinal analysis and population modelling. A variety of brain image segmentation and registration methods have been proposed. However, there is still room for improvement in terms of algorithm automation, accuracy and efficiency and the characteristics of existing methods for specific applications still can be more thoroughly investigated. Our work aims to obtain a good understanding in the underlying problems of brain image analysis,
including brain segmentation, brain image registration and primary tissue classification, and propose possible improvements for the most popular and/or the most promising techniques. For brain segmentation, an improved Brain Extraction Tool (BET) method is proposed, which overcomes the weaknesses of the original method by enhancing the vertex displacement and embedding an independent brain surface reconstruction step during the iterative process of surface evolution. This strategy effectively deals with the surface self-intersection problem and results in faster algorithm convergence and better brain segmentation. For brain image registration, we propose a salient
edge guided demons method, which uses salient edges detected in 3D scale-space rather than the whole image grid to drive the registration process. This method obtains statistically equal registration performance compared with the demons method using the whole image as demon points, while the execution time is dramatically reduced. For brain tissue classification, we compare the approaches using three main image based features: intensity, local prior and multi-atlas prior. The modelling of these features is described and the effectiveness of these features in brain tissue classification is investigated. This study provides a general guide on what image based features can be used for effective brain tissue classification.