The research in this thesis was undertaken because information on the relationship between scan parameters and image quality for the MARS spectral CT was lacking. However, the MARS spectral CT is expected to extend into clinical use in the future, so it is absolutely crucial that we know how the quality of the images that it produces is effected by different can parameters. This will allow us to make further improvements to the machine, and ultimately help clinicians to visualise important information in patients which are not revealed by other imaging modalities. This thesis provides information on how the image quality is affected by different scan parameters on the MARS spectral CT using a Medipix3 silicon quad detector. In particular, it explores how different numbers of projections, exposure time products (mAs), and peak tube voltages (kVp) with different threshold energies (kV) effect the image noise, image resolution and image uniformity, respectively. This provides a set of guidelines for future work using the MARS scanner to obtain images of optimal quality. This thesis also determines that the new image reconstruction software mART developed by Niels de Ruiter, is a suitable replacement for the reconstruction software OctopusCT that is currently being used by the MARS team. Using mART reduces the scan times and dose delivered by the MARS spectral CT.

The extra energy information provided by spectral x-ray imaging using novel photon counting x-ray detectors may allow for improved decomposition of materials compared to conventional and dual-energy imaging. The information content of spectral x-ray images, however, depends on how the photons are grouped together. This thesis deals with the theoretical aspect of optimising material discrimination in spectral x-ray imaging. A novel theoretical model was developed to map the confidence region of material thicknesses to determine the uncertainties in thickness quantification. Given the thickness uncertainties, photon counts per pixel can be optimised for material quantification in the most dose efficient manner. Minimisation of the uncertainties enables the optimisation of energy bins for material discrimination. Using Monte Carlo simulations based on the BEAMnrc package, material decomposition of up to 3 materials was performed on projection images, which led to the validation of the theoretical model. With the inclusion of scattered radiation, the theoretical optima of bin border energies were accurate to within 2 keV. For the simulated photon counts, excellent agreement was achieved between the theoretical and the BEAMnrc models regarding the signal-to-noise ratio in a decomposed image, particularly for the decomposition of two materials. Finally, this thesis examined the implementation of the Medipix detector. The equalisation of pixel sensitivity variations and the processing of photon counting projection images were studied. Measurements using the Medipix detector demonstrated promising results in the charge summing and the spectroscopic modes of acquisition, even though the spectroscopic performance of the detector was relatively limited due to electronic issues known to degrade the equalisation process. To conclude, the theoretical model is sufficient in providing guidelines for scanning parameters in spectral x-ray imaging and may be applied on spectral projection measurements using e.g. the redesigned MedipixRX detector with improved spectroscopic performance, when it becomes available.

This thesis is about the development of the MARS camera: a standalone portable digital x-ray camera with spectral sensitivity. It is built for use in the MARS Spectral system from the Medipix2 and Medipix3 imaging chips. Photon counting detectors and Spectral CT are introduced, and Medipix is identified as a powerful new imaging device. The goals and strategy for the MARS camera are discussed. The Medipix chip physical, electronic and functional aspects, and experience gained, are described. The camera hardware, firmware and supporting PC software are presented. Reports of experimental work on the process of equalisation from noise, and of tests of charge summing mode, conclude the main body of the thesis. The camera has been actively used since late 2009 in pre-clinical research. A list of publications that derive from the use of the camera and the MARS Spectral scanner demonstrates the practical benefits already obtained from this work. Two of the publications are first-author, eight are co-authored, and a further four acknowledge use of the MARS camera as part of the MARS scanner. The work has been presented at three MARS group meetings, two departmental conferences, and at an internal Medipix3 collaboration meeting hosted by ESRF in Grenoble.

Spectral computed tomography (CT) is a novel medical imaging technique that involves simultaneously counting photons at several energy levels of the x-ray spectrum to obtain a single multi-variate dataset. Visualisation of such data poses significant challenges due its extremely large size and the need for interactive performance for scientific and medical end-users. This thesis explores the properties of spectral CT datasets and presents two algorithms for GPU-accelerated real-time rendering from compressed spectral CT data formats. In addition, we describe an optimised implementation of a volume raycasting algorithm on modern GPU hardware, tailored to the visualisation of spectral CT data.

Spectral x-ray computed tomography (CT) is an important nascent imaging modality with several exciting potential applications. The research presented in this thesis separates into two primary areas with the common underlying theme of spectral CT; the first area is Compton scatter estimation and the second is interior tomography. First, the research is framed and outputs are identified. Background on the concepts used in the thesis is offered, including x-ray imaging and computed tomography, CT scanner architecture, spectral imaging, interior tomography and x-ray scatter. The mathematical background of techniques for image reconstruction from x-ray transmission measurements are presented. Many of the tools used to perform the research, both hardware and software, are described. An algorithm is developed for estimating the intensity of Compton scattered photons within a spectral CT scan, and a major approximation used by the algorithm is analysed. One proposed interior reconstruction algorithm is briefly evaluated; while this is not directly linked to spectral CT, it is related to the work on a novel hybrid spectral interior micro-CT architecture. Conclusions are summarised and suggestions for future work are offered. Scatter is known to cause artefacts in CT reconstructions, and several methods exist to correct data that has been corrupted by scatter. Compton scatter affects the energy of photons, therefore spectral CT measurements offer the potential to correct for this phenomenon more accurately than conventional measurements. A Compton scatter algorithm is developed and is found to match very well to Monte Carlo validation simulations, with the constraints that the object be at the micro-CT scale and that electron-binding effects are omitted. Development of the algorithm uses an approximation of the post-scatter attenuation to simplify the estimation problem and enable implementation. The consequences of this approximation are analysed, and the error introduced is found to be less than 5% in most biomedical micro-CT situations. Interior tomography refers to the incomplete data situation caused by the truncation of some or all CT projections, and is an active research area. A recently proposed interior reconstruction algorithm is evaluated with regard to its sensitivity to input error, and is found to have mediocre performance in this respect. Published results are not found to be reproducible, suggesting some omission from the published algorithm. A novel hybrid spectral interior architecture is described, along with an iterative reconstruction algorithm for hybrid data sets. The system combines a full field of view conventional imaging chain and an interior field of view spectral imaging chain to enable spectral measurement of a region of interest, and addresses some important limitations of spectral x-ray detectors; promising results are shown. Spectral reconstructions from interior data are shown to have sufficient information to distinguish two k-edge contrast agents (iodine and gadolinium) not only within the interior field of view but also beyond it. The architecture is further explored in the context of radiation exposure reduction, including testing of an analytical hybrid reconstruction algorithm.

The goal of this research is to contribute to the development of MARS spectral CT and to demonstrate the feasibility of molecular imaging using the technology. MARS is a newly developed micro CT scanner, incorporating the latest spectroscopic Medipix photon counting detector. I show that the scanner can identify both drug markers and stenosis of atherosclerosis labelled with non-toxic nanoparticles. I also show that spectral computed tomography using Medipix x-ray detectors can give quantitative measurements of concentrations of gold nanoparticles in phantoms, mice and excised atheroma. The characterisation of the Medipix2 assemblies with Si and CdTe x-ray sensors using poly-energetic x-ray sources has been performed. I measure the inhomogeneities within the sensors; individual pixel sensitivity response; and their saturation effects at higher photon fluxes. The effects of charge sharing on the performance of Medipix2 have been assessed, showing that it compromises energy resolution much more than spatial resolution. I have commissioned several MARS scanners incorporating several different Medipix2 and Medipix3 cameras. After the characterization of x-ray detectors and the geometrical assessment of MARS-CT, spectral CT data has been acquired, using x-ray energies that are appropriate for human imaging. The outcome shows that MARS scanner has the ability to discriminate among low atomic number materials, and from various concentrations of heavy atoms. This new imaging modality, used with functionalized gold nanoparticles, gives a new tool to assess plaque vulnerability. I demonstrated this by using gold nanoparticles, attached to antibodies, which targeted to thrombotic events in excised plaque. Likewise, the imaging modality can be used to track drugs labelled with any heavy atoms to assess how much drug gets into a target organ. Thus the methodology could be used to accelerate development of new drug treatments for cancers and inflammatory diseases.

Spectral computed tomography (CT) has proven an important development in biomedical imaging, yet there are several limitations to this nascent technology. Near-term implementation of spectral CT imaging can be enhanced using a hybrid architecture that integrates a narrow-beam spectral 'interior' imaging chain integrated with a traditional wide-beam 'global' imaging chain. The first study demonstrates the feasibility of hybrid spectral micro-CT architecture with a first-of-its-kind system implementation and preliminary results showing improved contrast resolution and spatial resolution. The second study seeks to characterize the hybrid spectral micro-CT scan protocol for reduction of radiation exposure. In the third study, the spectral 'interior' imaging chain was optimized for K-edge imaging of high-z elemental contrast agents. In the final study, an open-source, low-cost solution for managing digital content in an academic setting was demonstrated. The results of these studies confirm the merits of a hybrid architecture and warrant further consideration in future pre-clinical and clinical spectral micro-CT and CT scanner design and protocols. / Ph. D.

Computed Tomography (CT) is a non-invasive x-ray imaging method capable of reconstructing highly detailed cross-sectional interior maps of an object. CT is used in a range of medical applications such as detection of skeletal fractures, organ trauma and artery calcification. Reconstructing CT images requires the use of a forward operator, which is essentially a simulation of the scanning process. Photon-Counting CT is a rapidly developing alternative to conventional CT that promises higher spatial resolution, more accurate material separation and more robust reconstructions. A major difficulty in Photon-Counting CT is to model cross-talk between detectors. One way is to incorporate a wide point-spread function into the forward operator. Although this method works, it drastically slows down the reconstruction process.  In this thesis, we accelerate image reconstruction tasks for photon-counting CT by approximating the cross-talk component of the forward operator with a deep neural network, resulting in a learned forward operator. The learned operator reduces reconstruction error by an order of magnitude at the cost of a 20% increase in computation time, compared to ignoring cross-talk altogether. Furthermore, it generalises well to both unseen data and unseen detector settings. Our results indicate that a learned forward operator is a suitable way of approximating the forward operator in photon-counting CT. / Datortomografi (CT) är en icke-invasiv röntgenmetod som kan skapa högupplösta tvärsnittsbilder av objekt. CT används i en stor mängd tillämpningar, exempelvis vid detektion av frakturer, mjukvävnadstrauma och åderförkalkning. När man rekonstuerar tvärsnitt i CT krävs en simuleringsmodell som kallas framåtoperatorn. Fotonräknande CT är ett alternativ till konventionell CT som utlovar högre upplösning, mer precis uppdelning av material och högre robusthet i rekonstruktionerna. I fotonräknande CT är det viktigt att ta hänsyn till överhörning mellan detektorerna. Ett sätt är att inkorporera en punktspridningsfunktion i framåtoperatorn, vilket dessvärre saktar ned rekonstruktionsprocessen drastiskt.  I detta examensarbete approximerar vi överhörningseffekten mellan detektorer med ett djupt neuralt nätverk, med syfte att accelerera rekonstruktionsprocessen för fotonräknande spektral CT. Den inlärda framåtoperatorn reducerar rekonstruktionsfelet med en faktor tio på bekostnad av en 20-procentig ökning i beräkningstid, jämfört med en framåtoperator som inte modellerar överhörning. Vi visar att den inlärda framåtoperatorn generaliserar väl till data som den inte är tränad på, men även detektorinställningar den inte är van vid. Våra resultat tyder på att den inlärda framåtoperatorn är en lämplig approximationsmetod för framåtoperatorn i fotonräknande CT.

Depuis quelques années les détecteurs à pixels hybrides ont ouvert la voie au développement de la tomographie à rayon X spectrale ou tomodensitométrie (TDM) spectrale. La TDM spectrale permet d’extraire plus d’information concernant la structure interne de l’objet par rapport à la TDM d’absorption classique. Un de ses objectifs dans l’imagerie médicale est d’identifier et quantifier des composants d’intérêt dans un objet, tels que des marqueurs biologique appelés agents de contraste (iode, baryum, etc.). La majeure partie de l’état de l’art procède en deux étapes : - la "pré-reconstruction" qui consiste à séparer les composants dans l’espace des projections puis reconstruire, - la "post-reconstruction", qui reconstruit l’objet puis sépare les composants.On s’intéresse dans ce travail de thèse à une approche qui consiste à séparer et reconstruire simultanément les composants de l’objet. L’état de l’art des méthodes de reconstruction et séparation simultanées de données de TDM spectrale reste à ce jour peu fourni et les approches de reconstruction existantes sont limitées dans leurs performances et ne tiennent souvent pas compte de la complexité du modèle d’acquisition.L’objectif principal de ce travail de thèse est de proposer des approches de reconstruction et séparation tenant compte de la complexité du modèle afin d’améliorer la qualité des images reconstruites. Le problème à résoudre est un problème inverse, mal-posé, non-convexe et de très grande dimension. Pour le résoudre, nous proposons un algorithme proximal à métrique variable. Des résultats prometteurs sont obtenus sur des données réelles et montrent des avantages en terme de qualité de reconstruction. / In recent years, hybrid pixel detectors have paved the way for the development of spectral X ray tomography or spectral tomography (CT). Spectral CT provides more information about the internal structure of the object compared to conventional absorption CT. One of its objectives in medical imaging is to obtain images of components of interest in an object, such as biological markers called contrast agents (iodine, barium, etc.).The state of the art of simultaneous reconstruction and separation of spectral CT data methods remains to this day limited. Existing reconstruction approaches are limited in their performance and often do not take into account the complexity of the acquisition model.The main objective of this thesis work is to propose better quality reconstruction approaches that take into account the complexity of the model in order to improve the quality of the reconstructed images. Our contribution considers the non-linear polychromatic model of the X-ray beam and combines it with an earlier model on the components of the object to be reconstructed. The problem thus obtained is an inverse, non-convex and misplaced problem of very large dimensions.To solve it, we propose a proximal algorithmwith variable metrics. Promising results are shown on real data. They show that the proposed approach allows good separation and reconstruction despite the presence of noise (Gaussian or Poisson). Compared to existing approaches, the proposed approach has advantages over the speed of convergence.

Ring artifacts are a common problem with the use of photon-counting detectors and commercial deployment rests on being able to compensate for them. Deep learning has been proposed as a candidate for tackling the inefficiency or high cost of traditional techniques. In that spirit, we propose a new approach to ring artifact reduction, namely one that employs Residual Networks in sinogram domain. We train them on data simulated via a realistic photon-counting CT model based on numerical phantoms of real scans acquired by the KiTS19 Challenge dataset. By exploring various architectures we find that shallow ResNets achieve a significant artifact reduction by staying more true to the ground truth in terms of not introducing new artifacts. All networks introduce a smoothing effect which is attributed to the use of MSE as a loss function. An alternative training scheme using patches instead of whole sinograms is tested and it shows a slightly improved model stability. Lastly, we demonstrate via a performance metric study that common metrics are not suitable for quantifying the performance in this problem, save for a potential new approach in the virtual mono-energetic domain. / Ringartefakter är ett vanligt problem vid användning av fotonräknande detektorer och kommersiell introduktion kräver att man kan kompensera för dem. Djupinlärning har föreslagits som en kandidat för att hantera ineffektiviteten eller de höga kostnaderna för traditionella tekniker. I den andan föreslår vi ett nytt tillvägagångssätt för att reducera ringartefakter, nämligen en som använder sig av residualnätverk i sinogramdomänen. Vi tränar dem på data simulerad via en realistisk fotonräkning CT modell baserad på numeriska fantomer av verkliga skanningar från datamängen KiTS19 Challenge. Genom att utforska olika arkitekturer finner vi att grunda ResNet uppnår en betydande minskning av artefakter genom bevara en större likhet med den sanna bilden när det gäller att inte introducera nya artefakter. Alla nätverk introducerar en utsmetningseffekt som tillskrivs användningen av MSE som en förlustfunktion. Ett alternativt träningsschema med utsnitt istället för hela sinogram testas och det visar en något förbättrad modellstabilitet. Slutligen visar vi genom en prestandamåttstudie att vanliga prestandamått inte är lämpliga för att kvantifiera prestandan i detta problem med undantag för ett potentiellt nytt tillvägagångssätt i den virtuella monoenergetiska domänen.