Robust segmentation and statistical shape modelling : application to cardiac imaging
File(s)
Author(s)
Abi Nahed, Julien
Type
Thesis
Abstract
Cardiac segmentation involves accurate delineation of cardiac anatomy and it is required for deriving both local and global anatomical and functional indices. Consistent, robust and accurate segmentation has become a major bottleneck in standard clinical workflows. Whilst manual segmentation still remains a common practice in many clinical settings, it is getting increasingly impractical due to the large amount of data involved. Thus far, fully automatic techniques still remain difficult due to inconsistent image quality and high degree of morphological variations between different subjects. However, segmentation with active shape and appearance models are amongst popular techniques that are gaining significant clinical interest. These methods exploit a priori knowledge about the geometry and appearance of the anatomical structures, making them appealing for practical applications. For these techniques to be clinical relevant, several limitations need to be carefully addressed. For example, model construction requires establishing valid correspondence between shapes, and the effect of outliers on performance needs to be tackled. The purpose of this thesis is to present a practical and robust segmentation and modelling framework based on active shape models combined with the random walker algorithm and reinforcement learning. For shape modelling, a strategy based on harmonic embedding and the minimum description length is proposed. We also present a general framework for explicit handling of outliers in active shape models by employing deterministic annealing and softassign. The random walker algorithm is used for robust feature extraction based on discrete harmonic functions on graphs. Finally, an interactive framework based on reinforcement learning is developed so as to minimise user interaction when adjusting pre-segmentation results from the proposed model-based approach. Throughout this thesis, the methods proposed have been extensively validated with in vivo data and detailed statistical analysis has been provided for evaluating the accuracy and robustness of different aspects of the proposed techniques.
Version
Open Access
Date Awarded
2010
Copyright Statement
Attribution NoDerivatives 4.0 International Licence (CC BY-ND)
Advisor
Yang, Guang-Zhong
Jolly, Marie-Pierre
Sponsor
Imperial College London; Siemens Corporate Research, Princeton, USA.
Publisher Department
Deaprtment of Computing.
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)