Purpose: This dissertation focuses on establishment of pre-clinical methods facilitating the use of PET imaging for selective sub-volume dose escalation. Specifically the problems addressed are 1.) The difficulties associated with comparing multiple PET images, 2.) The need for further validation of novel PET tracers before their implementation in dose escalation schema and 3.) The lack of concrete pre-clinical data supporting the use of PET images for guidance of selective sub-volume dose escalations. Methods and materials: In order to compare multiple PET images the confounding effects of mispositioning and anatomical change between imaging sessions needed to be alleviated. To mitigate the effects of these sources of error, deformable image registration was employed. A deformable registration algorithm was selected and the registration error was evaluated via the introduction of external fiducials to the tumor. Once a method for image registration was established, a procedure for validating the use of novel PET tracers with FDG was developed. Nude mice were used to perform in-vivo comparisons of the spatial distributions of two PET tracers, FDG and FLT. The spatial distributions were also compared across two separate tumor lines to determine the effects of tumor morphology on spatial distribution. Finally, the research establishes a method for acquiring pre-clinical data supporting the use of PET for image-guidance in selective dose escalation. Nude mice were imaged using only FDG PET/CT and the resulting images were used to plan PET-guided dose escalations to a 5 mm sub-volume within the tumor that contained the highest PET tracer uptake. These plans were then delivered using the Small Animal Radiation Research Platform (SARRP) and the efficacy of the PET-guided plans was observed. Results and Conclusions: The analysis of deformable registration algorithms revealed that the BRAINSFit B-spline deformable registration algorithm available in SLICER3D was capable of registering small animal PET/CT data sets in less than 5 minutes with an average registration error of .3 mm. The methods used in chapter 3 allowed for the comparison of the spatial distributions of multiple PET tracers imaged at different times. A comparison of FDG and FLT showed that both are positively correlated but that tumor morphology does significantly affect the correlation between the two tracers. An overlap analysis of the high intensity PET regions of FDG and FLT showed that FLT offers additional spatial information to that seen with FDG. In chapter 4 the SARRP allowed for the delivery of planned PET-guided selective dose escalations to a pre-clinical tumor model. This will facilitate future research validating the use of PET for clinical selective dose escalation.

Deformable image registration (DIR) is an essential tool in medical image processing. It provides a means to combine image datasets, allowing for intra-subject, inter-subject, multi-modality, and multi-instance analysis, as well as motion detection and compensation. One of the most popular DIR algorithms models the displacement vector field (DVF) as B-splines, a sum of piecewise polynomials with coefficients that enable local shape control. B-splines have many advantageous properties in the context of DIR, but they often struggle to adequately model steep local gradients and discontinuities. This dissertation addresses that limitation by proposing the replacement of conventional B-splines with a generalized formulation known as a Non-Uniform Rational B-Splines (NURBS). Beginning with the 1D fitting, heuristic rules are developed to determine the values of the additional free parameters introduced by NURBS. These rules are subsequently modified and extended to the 2D and 3D fitting of anonymized and publicly available patient DVFs. Based on the lessons learned from these increasingly complex test cases, a 2D DIR scheme is developed and tested on slices from a thoracic computed tomography (CT) scan. Finally, an automatic, non-uniform scheme is presented, and its registration performance is compared to the conventional uniform methods.

Different medical imaging modalities provide complementary anatomical and functional information. One increasingly important use of such information is in the clinical management of cardiovascular disease. Multi-modality data is helping improve diagnosis accuracy, and individualize treatment. The Clinical Research Imaging Centre at the University of Edinburgh, has been involved in a number of cardiovascular clinical trials using longitudinal computed tomography (CT) and multi-parametric magnetic resonance (MR) imaging. The critical image processing technique that combines the information from all these different datasets is known as image registration, which is the topic of this thesis. Image registration, especially multi-modality and multi-parametric registration, remains a challenging field in medical image analysis. The new registration methods described in this work were all developed in response to genuine challenges in on-going clinical studies. These methods have been evaluated using data from these studies. In order to gain an insight into the building blocks of image registration methods, the thesis begins with a comprehensive literature review of state-of-the-art algorithms. This is followed by a description of the first registration method I developed to help track inflammation in aortic abdominal aneurysms. It registers multi-modality and multi-parametric images, with new contrast agents. The registration framework uses a semi-automatically generated region of interest around the aorta. The aorta is aligned based on a combination of the centres of the regions of interest and intensity matching. The method achieved sub-voxel accuracy. The second clinical study involved cardiac data. The first framework failed to register many of these datasets, because the cardiac data suffers from a common artefact of magnetic resonance images, namely intensity inhomogeneity. Thus I developed a new preprocessing technique that is able to correct the artefacts in the functional data using data from the anatomical scans. The registration framework, with this preprocessing step and new particle swarm optimizer, achieved significantly improved registration results on the cardiac data, and was validated quantitatively using neuro images from a clinical study of neonates. Although on average the new framework achieved accurate results, when processing data corrupted by severe artefacts and noise, premature convergence of the optimizer is still a common problem. To overcome this, I invented a new optimization method, that achieves more robust convergence by encoding prior knowledge of registration. The registration results from this new registration-oriented optimizer are more accurate than other general-purpose particle swarm optimization methods commonly applied to registration problems. In summary, this thesis describes a series of novel developments to an image registration framework, aimed to improve accuracy, robustness and speed. The resulting registration framework was applied to, and validated by, different types of images taken from several ongoing clinical trials. In the future, this framework could be extended to include more diverse transformation models, aided by new machine learning techniques. It may also be applied to the registration of other types and modalities of imaging data.

Computational fluid dynamics (CFD) has become an attractive tool in understanding the characteristic of air flow in the human lungs. Inter-subject variations make subject-specific simulations essential for understanding structure-function relationship, assessing lung function and improving drug delivery. However, currently the subject-specific CFD analysis remains challenging due, in large part to, two issues: construction of realistic deforming airway geometry and imposition of physiological boundary conditions. To address these two issues, we develop subject-specific, dynamic lung models by utilizing two or multiple volume multi-detector row computed tomography (MDCT) data sets and image registrations in this thesis. A mass-preserving nonrigid image registration algorithm is first proposed to match a pair of three-dimensional (3D) MDCT data sets with large deformations. A novel similarity criterion, the sum of squared tissue volume difference (SSTVD), is introduced to account for changes in intensity with lung inflation. We then demonstrate the ability to develop dynamic lung models by using a pair of lung volumes to account for deformations of airway geometries and subject-specific boundary conditions. The deformation of the airway geometry is derived by the registration-derived deformation field and subject-specific boundary condition is estimated from regional ventilation in a 3D and one-dimensional (1D) coupled multi-scale framework. Improved dynamic lung models are then proposed from three lung volumes by utilizing nonlinear interpolations. The improved lung models account for nonlinear geometry motions and time-varying boundary conditions during breathing. The capability of the proposed dynamic lung model is expected to move the CFD-based interrogation of lung function to the next plateau.

Image super-resolution aims to produce a higher resolution representation of a scene from an ensemble of low-resolution images that may be warped, aliased, blurred and degraded by noise. There are a variety of methods for performing super-resolution described in the literature, and in general they consist of three major steps: image registration, fusion and deblurring. This thesis proposes a novel method of performing the first two of these steps. The ultimate aim of image super-resolution is to produce a higher-quality image that is visually clearer, sharper and contains more detail than the individual input images. Machine algorithms can not assess images qualitatively and typically use a quantitative error criterion, often least-squares. This thesis aims to optimise leastsquares directly using a fast method, in particular one that can be implemented using linear filters; hence, a closed-form solution is required. The concepts of optimal interpolation and resampling are derived and demonstrated in practice. Optimal filters optimised on one image are shown to perform nearoptimally on other images, suggesting that common image features, such as stepedges, can be used to optimise a near-optimal filter without requiring the knowledge of the ground-truth output. This leads to the construction of a pulse model, which is used to derive filters for resampling non-uniformly sampled images that result from the fusion of registered input images. An experimental comparison shows that a 10th order pulse model-based filter outperforms a number of methods common in the literature. The use of optimal interpolation for image registration linearises an otherwise nonlinear problem, resulting in a direct solution. Experimental analysis is used to show that optimal interpolation-based registration outperforms a number of existing methods, both iterative and direct, at a range of noise levels and for both heavily aliased images and images with a limited degree of aliasing. The proposed method offers flexibility in terms of the size of the region of support, offering a good trade-off in terms of computational complexity and accuracy of registration. Together, optimal interpolation-based registration and fusion are shown to perform fast, direct and effective super-resolution.

Image super-resolution aims to produce a higher resolution representation of a scene from an ensemble of low-resolution images that may be warped, aliased, blurred and degraded by noise. There are a variety of methods for performing super-resolution described in the literature, and in general they consist of three major steps: image registration, fusion and deblurring. This thesis proposes a novel method of performing the first two of these steps. The ultimate aim of image super-resolution is to produce a higher-quality image that is visually clearer, sharper and contains more detail than the individual input images. Machine algorithms can not assess images qualitatively and typically use a quantitative error criterion, often least-squares. This thesis aims to optimise leastsquares directly using a fast method, in particular one that can be implemented using linear filters; hence, a closed-form solution is required. The concepts of optimal interpolation and resampling are derived and demonstrated in practice. Optimal filters optimised on one image are shown to perform nearoptimally on other images, suggesting that common image features, such as stepedges, can be used to optimise a near-optimal filter without requiring the knowledge of the ground-truth output. This leads to the construction of a pulse model, which is used to derive filters for resampling non-uniformly sampled images that result from the fusion of registered input images. An experimental comparison shows that a 10th order pulse model-based filter outperforms a number of methods common in the literature. The use of optimal interpolation for image registration linearises an otherwise nonlinear problem, resulting in a direct solution. Experimental analysis is used to show that optimal interpolation-based registration outperforms a number of existing methods, both iterative and direct, at a range of noise levels and for both heavily aliased images and images with a limited degree of aliasing. The proposed method offers flexibility in terms of the size of the region of support, offering a good trade-off in terms of computational complexity and accuracy of registration. Together, optimal interpolation-based registration and fusion are shown to perform fast, direct and effective super-resolution.

Image super-resolution aims to produce a higher resolution representation of a scene from an ensemble of low-resolution images that may be warped, aliased, blurred and degraded by noise. There are a variety of methods for performing super-resolution described in the literature, and in general they consist of three major steps: image registration, fusion and deblurring. This thesis proposes a novel method of performing the first two of these steps. The ultimate aim of image super-resolution is to produce a higher-quality image that is visually clearer, sharper and contains more detail than the individual input images. Machine algorithms can not assess images qualitatively and typically use a quantitative error criterion, often least-squares. This thesis aims to optimise leastsquares directly using a fast method, in particular one that can be implemented using linear filters; hence, a closed-form solution is required. The concepts of optimal interpolation and resampling are derived and demonstrated in practice. Optimal filters optimised on one image are shown to perform nearoptimally on other images, suggesting that common image features, such as stepedges, can be used to optimise a near-optimal filter without requiring the knowledge of the ground-truth output. This leads to the construction of a pulse model, which is used to derive filters for resampling non-uniformly sampled images that result from the fusion of registered input images. An experimental comparison shows that a 10th order pulse model-based filter outperforms a number of methods common in the literature. The use of optimal interpolation for image registration linearises an otherwise nonlinear problem, resulting in a direct solution. Experimental analysis is used to show that optimal interpolation-based registration outperforms a number of existing methods, both iterative and direct, at a range of noise levels and for both heavily aliased images and images with a limited degree of aliasing. The proposed method offers flexibility in terms of the size of the region of support, offering a good trade-off in terms of computational complexity and accuracy of registration. Together, optimal interpolation-based registration and fusion are shown to perform fast, direct and effective super-resolution.

La microscopie confocale classique permet d'obtenir des images à haute réso- lution de cellules en culture ou dans un tissu biologique excisé. Cette technologie peut être adaptée aux applications in vivo grâce à l'utilisation de fibres optiques et d'optiques miniaturisées. A terme, la microscopie confocale fibrée devrait permettre aux médecins et biologistes de réaliser des biopsies optiques; c'est à dire un exa- men histologique, en temps réel, des tissus biologiques à l'intérieur d'un organisme vivant et directement au contact de la zone d'intérêt. Le but premier de cette thèse est de dépasser les limites matérielles de ces in- struments d'imagerie en développant des outils de recalage d'images spécifiques et innovants. En particulier, le propos de ce manuscrit est cadré par l'objectif de pro- poser, au travers d'outils de création de mosaïques d'images, des biopsies optiques à grand champ aux médecins. Cette application est considérée, dans cette thèse, comme un système, ou un circuit, qui prendrait en entrée un flot de données brutes et délivrerait en sortie des mosaïques d'images à grand champ. Nous détaillons les éléments critiques de ce système, en particulier la reconstruction d'images en temps réel, le recalage linéaire d'images et le recalage non linéaire, avant de présenter la structure du système complet. Les données brutes produites par la microscopie confocale fibrée sont difficiles à interpréter parce qu'elle sont modulées par la structure en nid d'abeille du réseau de fibres optiques et parce qu'elle sont entachées d'artefacts géométriques. Dans ce contexte, nous montrons qu'une reconstruction en temps réel des images peut être utilisée en pré-traitement afin de produire des séquences vidéos directement interprétables. Comme la microscopie confocale fibrée est une imagerie qui se fait au contact des tissus, le mouvement relatif du tissu par rapport à la sonde optique implique qu'il est parfois difficile d'obtenir de manière robuste certaines mesures quantitatives d'intérêt. Nous avons donc attaqué le problème du recalage linéaire, efficace et robuste de paires d'images. Nous montrons que des outils ré- cents provenant du domaine du contrôle robotique par la vision peuvent surpasser les solutions standards utilisées en analyse d'images biomédicales. L'adéquation de ces outils au problème du recalage linéaire d'images nous a amenés à revisiter le problème du recalage non-linéaire. En interprétant le recalage non-linéaire comme un problème d'optimisation sur un groupe de Lie, nous développons un algorithme rapide de recalage difféomorphe non-paramétrique d'images. En plus d'être dif- féomorphe, notre algorithme produit des résultats qui sont similaires à ceux de l'algorithme des démons de Thirion mais qui sont plus lisses et plus proche de la vérité. Finalement, nous obtenons une boîte à outils de reconstruction et de recalage d'images que nous utilisons pour proposer un algorithme robuste de création de mosaïques d'images qui permette de calculer un alignement globalement cohérent à partir de résultats locaux, de compenser les distorsions liées au mouvement et de retrouver les déformations non-rigides. Par ailleurs, notre algorithme de mosaïques d'images a récemment été incorporé dans un essai clinique multicentrique. Cet essai illustre l'intérêt clinique de nos outils dans le cadre spécifique de la surveillance de l'oesophage de Barrett.

This dissertation provides the foundations of computer-based automated phenotyping methods for analyzing 3D images of mouse embryos. A group-wise registration technique was utilized and optimized and computerized methods were employed for analysis of 3D MRI images of mouse embryos. The assumption that embryo anatomy is highly conserved among genetically identical specimens was verified. The group-wise registration approach was used to align a group of embryos from the 129S1/SvImJ (129Sv) strain as well as a group of C57BL/6J (C57) embryos. Finally, we shed some light on some of the morphological differences between the 129Sv and C57 strains using automated techniques.

This dissertation provides the foundations of computer-based automated phenotyping methods for analyzing 3D images of mouse embryos. A group-wise registration technique was utilized and optimized and computerized methods were employed for analysis of 3D MRI images of mouse embryos. The assumption that embryo anatomy is highly conserved among genetically identical specimens was verified. The group-wise registration approach was used to align a group of embryos from the 129S1/SvImJ (129Sv) strain as well as a group of C57BL/6J (C57) embryos. Finally, we shed some light on some of the morphological differences between the 129Sv and C57 strains using automated techniques.