Search

Primary sclerosing cholangitis (PSC) affects the biliary tree and liver. Magnetic resonance imaging (MRI) with cholangiopancreatography (MRCP) is used to diagnose and follow-up patients. Its current use and applications, however, are not standardised and relies on qualitative assessments susceptible to inter-observer variability. The aim of this thesis was to identify standardised and objective imaging metrics derived from quantitative MRI techniques that could potentially be used in future clinical assessment of patients with PSC, longitudinal studies, and interventional drug trials. The quantitative techniques used in this thesis include (i) MRCP+™, a three-dimensional (3D) biliary tree model computed from MRCP data to derive biliary metrics; (ii) T1 and T2* mapping to compute the iron-corrected T1 (cT1) using LiverMultiScan™; and (iii) co-registration of MRCP+ biliary tree model with LiverMultiScan cT1 maps to calculate periductal cT1. Patients with PSC were categorised into either high-risk or low-risk group for disease progression based on published thresholds of revised Mayo Risk Score (MRS), Amsterdam-Oxford Model (AOM), upper limit of normal (ULN) alkaline phosphatase (ALP), transient elastography liver stiffness (LS), enhanced liver fibrosis score (ELF), modified Amsterdam intrahepatic stricture severity score (ISSS), Anali score, presence of extrahepatic disease, and presence of dominant stricture. Patient reported outcomes were collected using generic SF-36 and disease-specific PSC-PRO questionnaires. The intrahepatic sum of the relative severity of dilatations (SumRelSevDilat) was consistently higher in the high-risk compared to the low-risk PSC group (MRS: p<0.0001; AOM: p=0.0017; 2.2x ULN ALP: p=0.0007; 1.5x ULN ALP: p=0.0025; LS: p=0.0022; ELF: p=0.0189; ISSS: p<0.0001; Anali: p=0.0299). It identified the high-risk MRS > 0 group (AUC=0.77, 95% CI 0.66-0.88) numerically better than other non-invasive markers and was an independent predictor of this high-risk group (OR: 29.2, p=0.038). SF-36 bodily pain (rs=-0.26, p=0.025) and PSC-PRO activities of daily living (rs=0.29, p=0.014) showed the strongest correlations with intrahepatic SumRelSevDilat. Two liver cT1 metrics, interquartile range of cT1 (IQR cT1) and mean superpixel variance cT1, were identified as potential markers of liver parenchymal heterogeneity in PSC. IQR cT1 had the best performance to detect the high-risk PSC group stratified by LS (AUC=0.84, 95% CI 0.75-0.93, p<0.0001). Mean superpixel variance cT1 differentiated PSC from every other liver disease (p<0.0001) but did not differentiate between the other liver diseases. Median cT1 correlated with all the domains of SF-36 with the strongest being mental health (rs=-0.37, p=0.001) and social functioning: rs=-0.32, p=0.005). The periductal mean cT1 1.5-3.5mm surrounding the bile ducts were higher than regions further away from the bile duct (p<0.0001) in PSC but not in healthy controls. Patients with LS > 9.6kPa (AUC=0.74, 95% CI 0.60-0.88, p=0.0021) and AOM > 2 (AUC=0.72, 95% CI 0.59-0.85, p=0.0031) had higher values than those with LS ≤ 9.6kPa and AOM ≤ 2, respectively. The mental health (rs=0.38, p=0.002) and role emotional (rs=0.37, p=0.010) components of SF-36 showed the strongest correlations with periductal mean cT1. Temporal changes in mean values of IQR cT1 showed a significant interaction with time and baseline risk categorisation by AOM (p<0.001), LS (p=0.010), ELF (p=0.004), ISSS (p=0.039) and Anali (p<0.001). The diagnostic performances of single baseline measurements were numerically better than temporal changes to detect the high-risk group. Validation against clinical endpoints is warranted before these metrics can be tested in multicentre prospective studies and considered as surrogate endpoints in clinical trials.

Cancer is one of the leading causes of premature deaths across the world today, and there is an urgent need for imaging techniques that can help in early diagnosis and treatment planning for cancer patients. In the last four decades, magnetic resonance imaging (MRI) has emerged as one of the leading modalities for non-invasive imaging of tumours. By using dynamic contrast-enhanced magnetic resonance imaging (DCEMRI), this modality can be used to acquire information about perfusion and vascularity of tumours, which can help in predicting response to treatment. There are many factors that complicate the analysis of DCE-MRI data, and make clinical predictions based on it unreliable. During data acquisition, there are many sources of uncertainties and errors, especially patient motion, which result in the same image position being representative of many different anatomical locations across time. Apart from motion, there are also other inherent uncertainties and noise associated with the measurement of DCE-MRI parameters, which contribute to the model-fitting error observed when trying to apply pharmacokinetic (PK) models to the data. In this thesis, a probabilistic, model-based registration and parameter estimation (MoRPE) framework for motion correction and PK-parameter estimation in DCE-MRI sequences is presented. The MoRPE framework is first compared with conventional motion correction methods on simulated data, and then applied to data from a clinical trial involving twenty colorectal cancer patients. On clinical data, the ability of MoRPE to discriminate between responders and non-responders to combined chemoand radiotherapy is tested, and found to be superior to other methods. The effect of incorporating different arterial input functions within MoRPE is also assessed. Following this, a quantitative analysis of the uncertainties associated with the different PK parameters is performed using a variational Bayes mathematical framework. This analysis provides a quantitative estimate of the extent to which motion correction affects the uncertainties associated with different parameters. Finally, the importance of estimating spatial heterogeneity of PK parameters within tumours is assessed. The efficacy of different measures of spatial heterogeneity, in predicting response to therapy based on the pre-therapy scan alone are compared, and the prognostic value of a new derived PK parameter the 'acceleration constant' is investigated. The integration of uncertainty estimates of different DCE-MRI parameters into the calculation of their heterogeneity measures is also shown to improve the prediction of response to therapy.

The overall clinical motivation of this thesis is to differentiate between the different stages of liver disease stratifying into: no disease, mild disease, and severe fibrosis using Magnetic Resonance Imaging (MRI). As a related aim, we seek to differentiate as much as possible pericellular and nonpericellular fibrosis. This latter is clinically important, but currently no method exists that is able to perform this. Quickly, we realised that these aims push low level image analysis beyond their current bounds and so a great deal of the thesis is dedicated to extending such techniques before they can be applied. To work on the most fundamental low level image analysis concepts and algorithms we choose one of the most recent developments, namely continuous intrinsic dimensionality (ciD), which allows the continuous classification of homogeneous patches from 1D structures to intrinsically 2D structures. We show that the current formalism has several fundamental limitations and we propose a number of developments to improve on these. We re-evaluated feature energy statistics that were originally proposed in ciD, and additionally we examined the confidence one may have in stateof- the-art methods to estimate the orientation of features. We show that new statistical methods are required for feature energy, and that orientation predictability is more important than correctness of the estimation. This evaluation led us to the monogenic signal local orientation. Analysis of feature or texture energy is also a main contribution of this thesis. Within this framework we propose the Riesz-weighted phase congruency model. This is able to detect internal texture structures but it is not capable of delineating boundaries. Nevertheless, it proves an appropriate basis for texture quantification. Finally, we show that in contrast to using the standard established Kovesi approach, the developed texture measure leads to good results on the suboptimal T1w MRI liver image staging images. We show that we are able to differentiate automatically between the separate disease scores and between pericellular and non-pericellular fibrosis.

Medical imaging has seen a rapid development in its clinical use in assessment of treatment outcome, disease monitoring and diagnosis over the last few decades. Yet, the vast amount of available image data limits the practical use of this potentially very valuable source of information for radiologists and physicians. Therefore, the design of computer-aided medical image analysis is of great importance to imaging in clinical practice. This thesis deals with the problem of deformable image registration in the context of lung imaging, and addresses three of the major challenges involved in this challenging application, namely: designing an image similarity for multi-modal scans or scans of locally changing contrast, modelling of complex lung motion, which includes sliding motion, and approximately globally optimal mathematical optimisation to deal with large motion of small anatomical features. The two most important contributions made in this thesis are: the formulation of a multi-dimensional structural image representation, which is independent of modality, robust to intensity distortions and very discriminative for different image features, and a discrete optimisation framework, based on an image-adaptive graph structure, which enables a very efficient optimisation of large dense displacement spaces and deals well with sliding motion. The derived methods are applied to two different clinical applications in pulmonary image analysis: motion correction for breathing-cycle computed tomography (CT) volumes, and deformable multi-modal fusion of CT and magnetic resonance imaging chest scans. The experimental validation demonstrates improved registration accuracy, a high quality of the estimated deformations, and much lower computational complexity, all compared to several state-of-the-art deformable registration techniques.

We describe an online clinical decision support (CDS) system, Lung Cancer Assistant (LCA), which we have developed to aid the clinicians in arriving at informed treatment decisions for lung cancer patients at multidisciplinary team (MDT) meetings. LCA integrates rule-based and probabilistic decision support within a single platform. To our knowledge, this is the first time this has been achieved in the context of CDS in cancer care. Rule-based decision support is achieved by an original ontological guideline rule inference framework that operates on a domain-specific module of Systematized Nomenclature of Medicine-Clinical Terms (SNOMED-CT), containing clinical concepts and guideline rule knowledge elicited from the major national and international guideline publishers. It adopts a conventional argumentation-based decision model, whereby the decision options are listed along with arguments derived by matching the patient records to the guideline rule base. As an additional feature of this framework, when a new patient is entered, LCA displays the most similar patients to the one being viewed. Probabilistic inference is provided by a Bayesian Network (BN) whose structure and parameters have been learned based on the English Lung Cancer Database (LUCADA). This allows LCA to predict the probability of patient survival and lay out how the selection of different treatment plans would affect it. Based on a retrospective patient subset from LUCADA, we present empirical results on the treatment recommendations provided by both functionalities of LCA and discuss their strengths and weaknesses. Finally, we present preliminary work, which may allow utilising the BN to calculate survival odd ratios that could be translated into quantitative degrees of support for the guideline rule-based arguments. An online version of LCA is accessible on http://lca.eng.ox.ac.uk.

The number of women diagnosed with the breast cancer continues to rise year on year. Breast cancer is now the most common type of cancer in the UK, with over 55000 cases reported last year. In most cases, mammography is the first step towards diagnosing breast cancer. However, it continues to have many practical limitations as compared to more sophisticated modalities such as MRI. The relatively low cost of mammography, together with the ever increasing risk of women contracting the disease, has led to many developed countries having a breast screening program. These routine breast screens are taken at different points in time and are called temporal mammograms. Currently, a radiologist tends to qualitatively assess temporal mammograms and look for any abnormalities or suspicious regions that might be of a concern. In this thesis, we develop an automatic shape analysis model that can detect and quantify such changes inside the breast. This will not only help in early diagnosis of the disease, which is key to survival, but will potentially aid prognosis and post treatment care. The core to this thesis is the use of Circular Integral Invariants. We explore its multi-scale properties and use it for image smoothing to reduce image noise and enhance features for segmentation. We implement, modify and enhance a segmentation method which previously has been successfully used to acquire breast regions of interest. We applied such Integral Invariants for shape description, to be used for shape matching as well as for subdividing shapes into sub-regions and quantifying the differences between two such shapes. We combine boundary information with the information from inside a shape, thus eccentrically transforming shapes before describing their structure. We develop a novel false positives reduction method based on Integral Invariants scale space. A second aspect of the thesis is the evaluation of and emphasis on the use of breast density maps against the commonly used intensity maps or x-rays. We find density maps sufficient to use in clinical practice. The methods developed in this thesis aim to help clinicians in making diagnostic decision at the point of case. Our shape analysis model is easy to compute, fast and very general in nature that could be deployed in a wide range of applications, beyond mammography.

Lung cancer is a disease affecting millions of people every year and poses a serious threat to global public health. Accurate lung cancer staging is crucial to choose an appropriate treatment protocol and to determine prognosis, this requires the acquisition of contrast-enhanced diagnostic CT (d-CT) that is usually followed by a PET/CT scan. Information from both d-CT and PET scan is used by the clinician in the staging process; however, these images are not intrinsically aligned because they are acquired on different days and on different scanners. Establishing anatomical correspondence, i.e., aligning the d-CT and the PET images is an inherently difficult task due to the absence of a direct relationship between the intensities of the images. The CT acquired during the PET/CT scan is used for attenuation correction (AC-CT) and is implicitly aligned with the PET image as they are acquired at the same time using a hybrid scanner. Patients are required to maintain shallow breathing for both scans. In contrast to that, the d-CT image is acquired after the injection of a contrast agent, and patients are required to maximally inhale, for better view of the lungs. Differences in the AC-CT and d-CT image volumes are thus due to differences in breathhold positions and image contrast. Nonetheless, both images are from the same modality. In this thesis, we present a new approach that aligns the d-CT with the PET image through an indirect registration process that uses the AC-CT. The deformation field obtained after the registration of the AC-CT to d-CT is used to align the PET image to the d-CT. Conventional image registration techniques deform the entire image using homogeneous regularization without taking into consideration the physical properties of the various anatomical structures. This homogeneous regularization may lead to physiologically and physically implausible deformations. To register the d-CT and AC-CT images, we developed a 3D registration framework based on a fluid transformation model including three physically motivated properties: (i) sliding motion of the lungs against the pleura; (ii) preservation of rigid structures; and (iii) preservation of topology. The sliding motion is modeled using a direction dependent regularization that decouples the tangential and the normal components of the external force term. The rigid shape of the bones is preserved using a spatially varying filter for the deformations. Finally, the topology is maintained using the concept of log-unbiased deformations. To solve the multi-modal registration problem due to the contrast injected for the d-CT, but lack thereof in the AC-CT, we use local cross correlation (LCC) as the similarity measure. To illustrate and validate the proposed registration framework, different intra-patient CT datasets are used, including the NCAT phantom, EMPIRE10 and POPI datasets. Results show that our proposed registration framework provides improved alignment and physically motivated deformations when compared to the classic elastic and fluid registration techniques. The final goal of our work was to demonstrate the clinical utility of our new approach that aligns d-CT and PET/AC-CT images for fusion. We apply our method to ten real patients. Our results show that the PET images have much improved alignment with the d-CT images using our proposed registration technique. Our method was successful in providing a good overlap of the lungs, improved alignment of the tumours and a lower target registration error for landmarks in comparison to the classic fluid registration. The main contribution of this thesis is the development of a comprehensive registration framework that integrates important physical properties into a state-of-the-art transformation model with application to lung imaging in cancer.

Endometriosis is an oestrogen-dependent gynaecological condition defined as the presence of endometrial tissue outside the uterus cavity. The condition is predominantly found in women in their reproductive years, and associated with significant pelvic and abdominal chronic pain and infertility. The disease is believed to affect approximately 33% of women by a recent study. Currently, surgical intervention, often laparoscopic surgery, is the gold standard for diagnosing the disease and it remains an effective and common treatment method for all stages of endometriosis. Magnetic resonance imaging (MRI) of the patient is performed before surgery in order to locate any endometriosis lesions and to determine whether a multidisciplinary surgical team meeting is required. In this dissertation, our goal is to use image processing techniques to aid surgical planning. Specifically, we aim to improve quality of the existing images, and to automatically detect bladder endometriosis lesion in MR images as a form of bladder wall thickening. One of the main problems posed by abdominal MRI is the sparse anisotropic frequency sampling process. As a consequence, the resulting images consist of thick slices and have gaps between those slices. We have devised a method to fuse multi-view MRI consisting of axial/transverse, sagittal and coronal scans, in an attempt to restore an isotropic densely sampled frequency plane of the fused image. In addition, the proposed fusion method is steerable and is able to fuse component images in any orientation. To achieve this, we apply the Riesz transform for image decomposition and reconstruction in the frequency domain, and we propose an adaptive fusion rule to fuse multiple Riesz-components of images in different orientations. The adaptive fusion is parameterised and switches between combining frequency components via the mean and maximum rule, which is effectively a trade-off between smoothing the intrinsically noisy images while retaining the sharp delineation of features. We first validate the method using simulated images, and compare it with another fusion scheme using the discrete wavelet transform. The results show that the proposed method is better in both accuracy and computational time. Improvements of fused clinical images against unfused raw images are also illustrated. For the segmentation of the bladder wall, we investigate the level set approach. While the traditional gradient based feature detection is prone to intensity non-uniformity, we present a novel way to compute phase congruency as a reliable feature representation. In order to avoid the phase wrapping problem with inverse trigonometric functions, we devise a mathematically elegant and efficient way to combine multi-scale image features via geometric algebra. As opposed to the original phase congruency, the proposed method is more robust against noise and hence more suitable for clinical data. To address the practical issues in segmenting the bladder wall, we suggest two coupled level set frameworks to utilise information in two different MRI sequences of the same patients - the T2- and T1-weighted image. The results demonstrate a dramatic decrease in the number of failed segmentations done using a single kind of image. The resulting automated segmentations are finally validated by comparing to manual segmentations done in 2D.

Currently, mammography is the most common imaging technology used in breast screening. Low dose X-rays are passed through the breast to generate images called mammograms. One type of breast abnormality is a cluster of microcalcifications. Usually, in benign cases, microcalcifications result from the death of fat cells or are due to secretion by the lobules. However, in some cases, clusters of microcalcifications are indicative of early breast cancer, partly because of the secretions by cancer cells or the death of such cells. Due to the different attenuation characteristics of normal breast tissue and microcalcifications, the latter ideally appear as bright white spots and this allows detection and analysis for breast cancer classification. Microcalcification detection is one of the primary foci of screening and has led to the development of computer-aided detection (CAD) systems. However, a fundamental limitation of mammography is that it gives a 2D view of the tightly compressed 3D breast. The depths of entities within the breast are lost after this imaging process, even though the breast tissue is spread out as a result of the compression force applied to the breast. The superimposition of tissues can occlude cancers and this has led to the development of digital breast tomosynthesis (DBT). DBT is a three-dimensional imaging involving an X-ray tube moving in an arc around the breast, over a limited angular range, producing multiple images, which further undergo a reconstruction step to form a three-dimensional volume of breast. However, reconstruction remains the subject of research and small microcalcifications are "smeared" in depth by current algorithms, preventing detailed analysis of the geometry of a cluster. By using the geometry of the DBT acquisition system, we derive the "epipolar" trajectory of a microcalcification. As a first application of the epipolars, we develop a clustering algorithm after using the Hough transform to find corresponding points generated from a microcalcification. Noise points can also be isolated. In addition, we show how microcalcification projections can be detected adaptively. Epipolar analysis has also led to a novel detection algorithm for DBT using a Bayesian method, which estimates a maximum a posterior (MAP) labelling in each individual image and subsequently for all projections iteratively. Not only does this algorithm output the binary decision of whether a pixel is a microcalcification, it can predict the approximate depth of the microcalcification in the breast if it is. Based on the epipolar analysis, reconstruction of just a region of interest (ROI) e.g. microcalcification clusters is possible and it is more straightforward than any existing method using reconstruction slices. This potentially enables future classification of breast cancer when more clinical data becomes available.

Medical imaging has revolutionized cancer care and its use has grown massively over the past several decades. Images are increasingly stored in large digital image repositories such as hospital Picture Archiving and Communication System, which will hopefully provide a wealth of information on patient conditions and therapy outcomes as cancer diagnosis and therapy moves from 'one size fits all' to more personalized approaches tailored to each particular patient. However, converting the unstructured avalanche of data at thousands of different hospitals into clinically valuable biomarkers and tools requires that the images of different patients can be compared and efficiently searched. Our research aims to develop novel methods to compare whole-body scans of multiple patients; methods which incorporate 'intelligent' prior knowledge of the internal structure of the human body, as opposed to current methods of image registration which mostly rely on matching the voxel intensities and disregard their anatomical meaning. We develop computational methods for accurate and reliable automated localization of anatomical structures in whole-body images, which will help to automate key steps in cancer diagnosis and radiation treatment planning and save expensive clinicians' time while improving the reliability of their decisions. Conventional approaches to determining spatial correspondences between pairs or sets of images in medical imaging typically rely on image registration methods. There have been considerable advances in registration of multiple images of the same patient taken at different time-points, known as longitudinal studies. However, conventional methods, which rely on optimizing certain integral functions of voxel values over the entire image, are unreliable when applied to aligning whole-body images of different patients. Whole-body Computed Tomography (CT) images contain many different anatomical structures whose physical attributes and consequent appearance can be highly variable between patients. This substantial, but normal, variability is further increased by the presence of pathologies such as tumours and non-cancerous diseases, surgical interventions and degenerative changes due to aging as well as different patterns of contrast agent uptake. Conventional registration methods often get trapped in local minima that abound in such images, resulting in unreliable and inaccurate anatomical correspondences. The methods developed in this thesis tackle the problem of inter-patient registration by incorporating prior anatomical knowledge into parts-based graphical models that accurately and reliably localize arbitrary skeletal and soft-tissue anatomical landmarks in whole-body clinical oncology scans. We optimize parts-based graphical models called Pictorial Structures for accurate and reliable landmark localization in CT images and introduce novel methods that replace standard population models by models personalized to the particular patient. We also propose methods that further improve landmark localization while minimizing, as far as possible, the high costs of ground-truth annotation by expert radiologists. We do this by automatically discovering new landmark correspondences from a database of partially annotated images. The performance of the algorithms developed in my thesis is evaluated on a large database of clinical lung cancer PET/CT scans, showing superior accuracy and reliability of landmark localization compared to conventional methods.

Abnormal vascular structure has been identified as one of the major characteristics of tumours. In this thesis, we carry out quantitative analysis on different tumour vascular structures and research the relationship between vascular structure and its transportation efficiency. We first study segmentation methods to extract the binary vessel representations from microscope images. We found that local phase-hysteresis thresholding is able to segment vessel objects from noisy microscope images. We also study methods to extract the centre lines of segmented vessel objects, a process termed as skeletonization. We modified the conventional thinning method to regularize the extremely asymmetrical structure found in the segmented vessel objects. We found this method is capable to produce vessel skeletons with satisfactory accuracy. We have developed a software for 3D vessel structural analysis. This software is consisted of four major parts: image segmentation, vessel skeletonization, skeleton modification and structure quantification. This software has implemented local phase-hysteresis thresholding and structure regularization-thinning methods. A GUI was introduced to enable users to alter the skeleton structures based on their subjective judgements. Radius and inter branch length quantification can be conducted based on the segmentation and skeletonization results. The accuracy of segmentation, skeletonization and quantification methods have been tested on several synthesized data sets. The change of tumour vascular structure after drug treatment was then investigated. We proposed metrics to quantify tumour vascular geometry and statistically analysed the effect of tested drugs on normalizing tumour vascular structure. finally, we developed a spatio-temporal model to simulate the delivery of oxygen and 3-18 F-fluoro-1-(2-nitro-1-imidazolyl)-2-propanol (Fmiso), which is the hypoxia tracer that gives out PET signal in an Fmiso PET scanning. This model is based on compartmental models, but also considers the spatial diffusion of oxygen and Fmiso. We validated our model on in vitro spheroid data and simulated the oxygen and Fmiso distribution on the segmented vessel images. We contend that the tumour Fmiso distribution (as observed in Fmiso PET imaging) is caused by the abnormal tumour vascular structure which is further aroused from tumour angiogenesis process. We depicted a modelling framework to research the relationships between tumour angiogenesis, vessel structure and Fmiso distribution, which is going to be the focus of our future work.

In this work, we will present several contributions towards automatic recognition of BSL signs from continuous signing video sequences. Specifically, we will address three main points: (i) automatic detection and tracking of the hands using a generative model of the image; (ii) automatic learning of signs from TV broadcasts using the supervisory information available from subtitles; and (iii) generalisation given sign examples from one signer to recognition of signs from different signers. Our source material consists of many hours of video with continuous signing and corresponding subtitles recorded from BBC digital television. This is very challenging material for a number of reasons, including self-occlusions of the signer, self-shadowing, blur due to the speed of motion, and in particular the changing background. Knowledge of the hand position and hand shape is a pre-requisite for automatic sign language recognition. We cast the problem of detecting and tracking the hands as inference in a generative model of the image, and propose a complete model which accounts for the positions and self-occlusions of the arms. Reasonable configurations are obtained by efficiently sampling from a pictorial structure proposal distribution. The results using our method exceed the state-of-the-art for the length and stability of continuous limb tracking. Previous research in sign language recognition has typically required manual training data to be generated for each sign, e.g. a signer performing each sign in controlled conditions - a time-consuming and expensive procedure. We show that for a given signer, a large number of BSL signs can be learned automatically from TV broadcasts using the supervisory information available from subtitles broadcast simultaneously with the signing. We achieve this by modelling the problem as one of multiple instance learning. In this way we are able to extract the sign of interest from hours of signing footage, despite the very weak and "noisy" supervision from the subtitles. Lastly, we show that automatic recognition of signs can be extended to multiple signers. Using automatically extracted examples from a single signer, we train discriminative classifiers and show that these can successfully classify and localise signs in new signers. This demonstrates that the descriptor we extract for each frame (i.e. hand position, hand shape, and hand orientation) generalises between different signers.

In this thesis, we develop a motion correction method to overcome the degradation of image quality introduced by respiratory motion in positron emission tomography (PET), so that diagnostic performance for lung cancer can be improved. Lung cancer is currently the most common cause of cancer death both in the UK and in the world. PET/CT, which is a combination of PET and CT, providing clinicians with both functional and anatomical information, is routinely used as a non-invasive imaging technique to diagnose and stage lung cancer. However, since a PET scan normally takes 15-30 minutes, respiration is inevitable in data acquisition. As a result, thoracic PET images are substantially degraded by respiratory motion, not only by being blurred, but also by being inaccurately attenuation corrected due to the mismatch between PET and CT. If these challenges are not addressed, the diagnosis of lung cancer may be misled. The main contribution of this thesis is to propose a novel process for respiratory motion correction, in which non-attenuation corrected PET images (PET-NAC) are registered to a reference position for motion correction and then multiplied by a voxel-wise attenuation correction factor (ACF) image for attenuation correction. The ACF image is derived from a CT image which matches the reference position, so that no attenuation correction artefacts would occur. In experiments, the motion corrected PET images show significant improvements over the uncorrected images, which represent the acquisitions typical of current clinical practice. The enhanced image quality means that our method has the potential to improve diagnostic performance for lung cancer. We also develop an automatic lesion detection method based on motion corrected images. A small lung lesion is only 2 or 3 voxels in diameter and of marginal contrast. It could easily be missed by human observers. Our method aims to provide radiologists with a map of potential lesions for decision so that diagnostic efficiency can be improved. It utilises both PET and CT images. The CT image provides a lung mask, to which lesion detection is confined, whereas the PET image provides distribution of glucose metabolism, according to which lung lesions are detected. Experimental results show that respiratory motion correction significantly increases the success of lesion detection, especially for small lesions, and most of the lung lesions can be detected by our method. The method can serve as a useful computer-aided image analysing tool to help radiologists read images and find malignant lung lesions. Finally, we explore the possibility of incorporating temporal information into respiratory motion correction. Conventionally, respiratory gated PET images are individually registered to the reference position. Temporal continuity across the respiratory period is not considered. We propose a spatio-temporal registration algorithm, which models temporally smooth deformation in order to improve the registration performance. However, we discover that the improvement introduced by temporal information is relatively small at the cost of a much longer computation time. Spatial registration with regularisation yields similar results but is superior in speed. Therefore, it is preferable for respiratory motion correction.

Secondly, we incorporate knowledge of the physiology, or how we expect the anatomy to change due to treatment. We can represent these changes using the Jacobian of the deformation, which describes the local size and type of change. This is used to regularise the registration, and can be incorporated simultaneously with the iterations of the registration. The final result is an accurate and robust registration result that is clinically useful for finding corresponding features on pre- and post-treatment datasets.

The study of the correlation between the radiological characteristics of the breast and the likelihood of the breast containing, or subsequently developing, a malignant lesion is termed breast density. In this thesis, a technique is developed for measuring the x-rav attenuation properties of breast tissue from an x-ray mammogram. The measure is independent of the characteristics and configuration of the x-ray equipment used to acquire the image. The intention is to provide a tool which facilitates the study of breast density by providing quantitative measurements of tissue characteristics, which it is hoped will ultimately form the basis of future risk estimation and malignant lesion diagnosis techniques. Part I of the thesis discusses the various area and volumetric measurements of breast density presented in the literature, together with the current hypotheses explaining the statistical results observed in epidemiological studies from a histopathological standpoint. The aim is to develop an inter-disciplinary understanding of the area under investigation to facilitate optimal development. A novel physics based model describing the complete process of mammographic image formation is presented in part II of the thesis. The design, build and test methodology executed for each component of the mammographic equipment is discussed through the chapters: the x-rav tube; the image receptor; and the anti-scatter grid. A ray tracing algorithm for computing the details of traversal paths is presented, as is an algorithm for calculating the scattered energy incident upon the image receptor. In each chapter, the underlying physical model is described, followed by the software design and implementation details, and finally the results of the experimental phantom validations studies to verify the implementation and underlying model. Where appropriate, sensitivity analysis is included so that uncertainty bounds may be estimated. Part III of the thesis proposes a novel measure of breast density, and describes how it overcomes the limitations of many existing techniques, in particular the Highnam and Brady 1996 hint model. The results of a clinical trial are included which demonstrate a favourable correlation between the proposed measure and the occurrence of malignancy compared to that observed using other methods.

Previous research on multimedia information filtering has mainly concentrated on key frame identification and video skim generation for browsing purposes, however applications requiring the generation of summaries as the final product for user con- sumption are of equal scientific and commercial interest. Recent advances in computer vision have enabled the extraction of semantic events from an audio-visual signal, so it can be assumed for our purposes that such semantic labels are already available for use. We concentrate instead on developing methods to prioritise these semantic elements for inclusion in a summary which can be personalised to meet a particular user's needs. Our work differentiates itself from that in the literature as it is driven by the results of a knowledge elicitation study with expert summarisers. The experts in our study believe that summaries structured as a narrative are better able to convey the content of the original data to a user. Motivated by the information filtering problem, the primary contribution of this thesis is the design and implementation of a system to summarise sequences of events by automatic modelling of the causal relationships between them. We show, by com- parison against summaries generated by experts and with the introduction of a new coherence metric, that modelling the causal relationships between events increases the coherence and accuracy of summaries. We suggest that this claim is valid, not only in the domain of soccer highlights generation, in which we carry out the bulk of our experiments, but also in any other domain in which causal relationships can be iden- tified between events. This proposal is tested by applying our summarisation system to another, significantly different domain, that of business meeting summarisation, using the soccer training set and a manually generated ontology mapping. We introduce the concept of a context-group of causally related events as a first step towards modelling narrative episodes and present a comparison between a case based reasoning and a two-stage Markov model approach to summarisation. For both methods we show that by including entire context-groups in the summary, rather than single events in isolation, more accurate summaries can be generated. Our approach to personalisation biases a summary according to particular narrative plotlines using different subsets of the training data. Results show that the number of instances of certain event classes can be increased by biasing the training set appropriately. This method gives very similar results to a standard weighting method, while avoiding the need to tailor the weights to a particular application domain.

This work addresses the distortions in Functional Magnetic Resonance Images (FMRI) caused by subject motion. FMRI is a non-invasive technique which shows great promise in providing researchers and clinicians with neurological information both about healthy subjects and clinical patients by mapping functional activation within the brain using Echo Planar Imaging (EPI). If reliable information is to be obtained from these images, motion correction must be carried out in order to remove or suppress the artefacts arising from subject movement. This work begins by using exploratory data techniques to describe these artefacts so that they can be characterised according to their origin and spatio-temporal manifestation. Based on testing of the accuracy and consistency of existing rigid-body motion correction methods on FMRI data, a new registration algorithm Motion Correction using the FMRIB Linear Image Registration Tool (MCFLIRT) has been developed. It is shown that while MCFLIRT is both more accurate and more robust than previous methods, rigid-body registration schemes in general cannot completely remove the distortions associated with motion and so subsequent analysis of the images may still be inaccurate. Furthermore, it is demonstrated that failure to use a sufficiently detailed model of subject motion in FMRI can in fact lead to degradation of the images through the use of existing motion correction algorithms. Based on these findings, alternative schemes including non-rigid registration and adaptive real-time methods are evaluated. Leading on from this investigation, a framework for Temporally-Integrated Geometric EPI Realignment (TIGER), incorporating both spatial and temporal information about the images, is proposed. An implementation based on this novel modality-specific model is developed and tested against existing rigid-body registration methods. Results show that this new approach is able to achieve significantly more accurate results than previous methods. The quality of correction provided by this new approach brings more subtle artefacts in the data to the fore, suggesting a number of avenues of further research in this area. These are outlined in the final chapter of the thesis.

This thesis develops a self-adaptive architecture for image understanding that addresses certain kinds of lack of robustness common in image understanding programs. The architecture provides support for making image understanding programs that can manipulate their own semantics and thereby adjust their structure in response to changes in the environment that might cause static image understanding systems to fail. The general approach taken has been to explore the ideas of self-adaptive software and implement an architectural framework that addresses a class of problems that we term "interpretation problems" common in image understanding. Self-adaptive software is a relatively new idea and this thesis represents one of the first implementations of the general idea. The general idea is that to make programs robust to changing environmen- tal conditions that they should be "aware" of their relationship with the environment and be able to restructure themselves at runtime in order to "track" changes in the environment. The implementation takes the form of a multi-layered reflective interpreter that ma- nipulates and runs simple agents. The interpreter framework utilizes Monte-Carlo sam- pling as a mechanism for estimating most likely solutions, uses Minimum Descriptin Length (MDL) as a central coordinating device, and includes a theorem prover based compiler to restructure the program when necessary. To test the architectural ideas developed in the thesis a test domain of interpreting aerial images was chosen. Much of the research described in the thesis addresses issues in that problem domain. The task of the program is to segment, label, and parse aerial images so as to produce an image description similar to descriptions produced by a human expert. An image corpus is developed that is used as the source of domain knowledge. The first processing stage of the program segments the aerial images into segments similar to those found in the annotated corpus. To accomplish this a new segmentation algorithm that we call semantic segmentation was developed that not only used MDL as a principle to drive the low-level segmentation but also allows higher level semantics to influence the segmentation. In our usage of the algorithm those semantics take the form of labeling and parsing the resulting segmentation. The second stage labels the regions and parses the regions into a parse tree. To do this we develop a 2D statistical parser. Rules of grammar are induced from the corpus and an MDL parser finds approximations to the most probable parse of the regions of the segmented image.

This thesis develops a methodology for the segmentation of anatomical structures within "sparse" MR images. Sparse images were acquired in large numbers prior to the emergence of high-resolution MRI and they form the basis of many long term imaging studies. The term sparse refers to the fact that the volumetric image has very poor spatial resolution in the direction perpendicular to the slice plane. This leads to a significant degradation in image quality and effectively destroys the spatial continuity of the imaged object. Consequently, generic segmentation schemes --- particularly those based on voxel classification --- will yield poor results unless they have been augmented in some manner. Our Segmentation approach is based on a deformable simplex mesh surface, which iteratively interpolates extracted boundary point data. Prior information is mobilised at two levels. Boundary points are found using a matching algorithm based on a database of pre-specified piecewise constant models. These models represent possible idealised intensity profiles for the object boundary. In addition to the boundary model, there is a shape template. The template is generated from a training set of pre-segmented structures, which means that only shapes similar to those in the training set will be recovered. The segmentation proceeds in two phases. The first recovers the normal shape component, determined by the training set, whilst the second deforms smoothly from this constrained solution to produce a more veridical boundary representation. The segmentation scheme is applied to a number of sparse brain images. Qualitative validation --- accomplished by registering the surface extracted from the sparse data to a high resolution scan acquired at the same time-point --- indicates that a good approximation to the underlying boundary is obtainable from such images.

Metastasised breast cancer kills. There is no known cure, there are no known preventative measures, there are no drugs available with proven capacity to abate its effects. Early identification and excision of a malignancy prior to metastasis is the only method currently available for reducing the mortality due to breast disease. Automated analysis of mammograms has been proposed as a tool to aid radiologists detect breast disease earlier and with greater efficiency and success. This thesis addresses some of the major difficulties associated with the automated analysis of mammograms, in particular the difficulties caused by the high-frequency, relatively insignificant curvi-linear structures (CLS) comprising the blood vessels, milk-ducts and fibrous tissues. Previous attempts at automation have been overlooked these structures and the resultant complexity of that oversight has been handled inappropriately. We develop a model-based analysis of the CLS features, from the very anatomy of the breast, through mammography and digitisation to the image intensities. The model immediately dictates an algorithm for extracting a high-level feature description of the CLS features. This high-level feature description allows a systematic treatment of these image features prior to searching for instances of breast disease. We demonstrate a procedure for implementing such prior treatment by 'removing' the CLS features from the images. Furthermore, we develop a model of the expected appearance of mammographic densities in the CLS-removed image, which leads directly to an algorithm for their identification. Unfortunately the model also extracts many regions of the image that are not significant mammographic densities, and this therefore requires a subsequent segmentation stage. Unlike previous attempts which apply neural networks to this task, and therefore incorporate inherent insignificance as a consequence of insufficient data availability describing the significant mammographic densities, we illustrate the application of a new statistical method (novelty analysis) for achieving a statistically significant segmentation of the mammographic densities from the plethora of candidates identified at the previous stage. We demonstrate the ability of the CLS feature description to identify instances of radial-scar in mammograms, and note the suitability of the CLS and density descriptions for assessment of bilateral and temporal asymmetry. Some additional potential applications of these feature descriptions in arenas other than mammogram analysis are also noted.

Turtle mobile robot which is able to interact intelligently with a dynamically changing environment.

An extensive literature survey has revealed that the majority of previous work in infrared image processing has ignored the processes leading to the formation of infrared images. Processing has normally either been restricted to simple lowlevel image enhancement convolutions or has consisted of algorithms copied from computer vision without regard for the inherent differences between infrared and visible images. In this thesis, we address the problem of infrared image formation and derive an irradiance equation for simple infrared scenes. We consider the complications caused by mutual illumination of one or more bodies and indicate how the infrared irradiance equation can also be specified for more complex scenes. The infrared irradiance equation we derive is solved in closed form for some simple geometries for both Lambertian and non-Lambertian surfaces. An infrared imager has been built and is described. Images taken with the imager of a variety of scene geometries show that the experimental results compare favourably with the theoretically derived equations, indicating the validity of the theoretical analysis. We describe how a knowledge of the formation of infrared images can be used to predict the image irradiance pattern of a particular object. We also show how, with a knowledge of the radiance properties and surface geometry of the object, it is possible to detect instances of that object in a scene. Examples are given of successful object detection based on an understanding of the image irradiance. We present a brief history of infrared imagers and a description of the principles on which modern infrared imagers are based. In addition to the survey of the literature published on infrared image processing, a brief summary of some techniques from the computer vision literature and their suitability to infrared image processing is given. A selection of vision techniques are applied to both infrared and visible images to verify conclusions reached in the thesis.