PDF version of this entire document

Thesis Organisation

The structure of the thesis is as follows:

Chapters 2 and 3 provide detailed descriptions of existing methods of NRR and constructing appearance models, respectively. They also explore the link between the two.

Chapter 4 discusses previous work on constructing shape models using a minimum description length (MDL) approach.

Chapter 5 builds on Chapter 4. It briefly outlines an approach which combines NRR and appearance modelling, by using a model-based MDL-like objective function that drives registration. Results are shown for experiments with 1-D images, and from an additional set of experiments where 2-D images are used.

Chapter 6 exploits the duality of models and registration quality, defining two quantitative measures of model/registration quality which can be used to assess the results of NRR without the need for ground truth.

Chapter 7 describes a series of validation experiments which investigate the effects of deliberately perturbing the registration of an initially registered set of images. The results obtained using the measures introduced in Chapter 6 are compared to those obtained using a `gold standard' method of assessment, based on measuring the overlap of manually-annotated ground truth label images.

Chapter 8 presents a practical application of the new method of evaluating NRR. Three different NRR algorithms are applied to the registration of sets of 2-D MR brain images, demonstrating the superiority of a fully groupwise registration algorithm over a repeated pairwise approach.

Chapter 9 lists several possible extensions and several ways forward. It also describes the existing extension of the method to 3-D, and discusses its limitations.

Chapter 10 draws conclusions and provides a summary of the contributions of this work.