L'imagerie ultrasonore est une modalité d'acquisition privilégiée en imagerie médicale en raison de son innocuité, sa simplicité d'utilisation et son coût modéré d'utilisation. Néanmoins, la résolution limitée et le faible contraste limitent son utilisation dans certaines d'applications. C'est dans ce contexte que différentes techniques de post-traitement visant à améliorer la qualité de telles images sont proposées dans ce manuscrit. Dans un premier temps, nous proposons d'aborder le problème conjoint de la déconvolution et de la segmentation d'images ultrasonores en exploitant l'interaction entre ces deux problèmes. Le problème, énoncé dans un cadre bayésien, est résolu à l'aide d'un algorithme MCMC en raison de la complexité de la loi a posteriori des paramètres d'intérêt. Dans un second temps, nous proposons une nouvelle méthode rapide de super-résolution fondée sur la résolution analytique d'un problème de minimisation l2-l2. Il convient de remarquer que les deux approches proposées peuvent être appliquées aussi bien à des images ultrasonores qu'à des images naturelles ou constantes par morceaux. Enfin, nous proposons une méthode de déconvolution aveugle basée sur un modèle paramétrique de la réponse impulsionelle de l'instrument ou du noyau de flou. / In the field of medical image analysis, ultrasound is a core imaging modality employed due to its real time and easy-to-use nature, its non-ionizing and low cost characteristics. Ultrasound imaging is used in numerous clinical applications, such as fetus monitoring, diagnosis of cardiac diseases, flow estimation, etc. Classical applications in ultrasound imaging involve tissue characterization, tissue motion estimation or image quality enhancement (contrast, resolution, signal to noise ratio). However, one of the major problems with ultrasound images, is the presence of noise, having the form of a granular pattern, called speckle. The speckle noise in ultrasound images leads to the relative poor image qualities compared with other medical image modalities, which limits the applications of medical ultrasound imaging. In order to better understand and analyze ultrasound images, several device-based techniques have been developed during last 20 years. The object of this PhD thesis is to propose new image processing methods allowing us to improve ultrasound image quality using postprocessing techniques. First, we propose a Bayesian method for joint deconvolution and segmentation of ultrasound images based on their tight relationship. The problem is formulated as an inverse problem that is solved within a Bayesian framework. Due to the intractability of the posterior distribution associated with the proposed Bayesian model, we investigate a Markov chain Monte Carlo (MCMC) technique which generates samples distributed according to the posterior and use these samples to build estimators of the ultrasound image. In a second step, we propose a fast single image super-resolution framework using a new analytical solution to the l2-l2 problems (i.e., $\ell_2$-norm regularized quadratic problems), which is applicable for both medical ultrasound images and piecewise/ natural images. In a third step, blind deconvolution of ultrasound images is studied by considering the following two strategies: i) A Gaussian prior for the PSF is proposed in a Bayesian framework. ii) An alternating optimization method is explored for blind deconvolution of ultrasound.

abstract: Adaptive processing and classification of electrocardiogram (ECG) signals are important in eliminating the strenuous process of manually annotating ECG recordings for clinical use. Such algorithms require robust models whose parameters can adequately describe the ECG signals. Although different dynamic statistical models describing ECG signals currently exist, they depend considerably on a priori information and user-specified model parameters. Also, ECG beat morphologies, which vary greatly across patients and disease states, cannot be uniquely characterized by a single model. In this work, sequential Bayesian based methods are used to appropriately model and adaptively select the corresponding model parameters of ECG signals. An adaptive framework based on a sequential Bayesian tracking method is proposed to adaptively select the cardiac parameters that minimize the estimation error, thus precluding the need for pre-processing. Simulations using real ECG data from the online Physionet database demonstrate the improvement in performance of the proposed algorithm in accurately estimating critical heart disease parameters. In addition, two new approaches to ECG modeling are presented using the interacting multiple model and the sequential Markov chain Monte Carlo technique with adaptive model selection. Both these methods can adaptively choose between different models for various ECG beat morphologies without requiring prior ECG information, as demonstrated by using real ECG signals. A supervised Bayesian maximum-likelihood (ML) based classifier uses the estimated model parameters to classify different types of cardiac arrhythmias. However, the non-availability of sufficient amounts of representative training data and the large inter-patient variability pose a challenge to the existing supervised learning algorithms, resulting in a poor classification performance. In addition, recently developed unsupervised learning methods require a priori knowledge on the number of diseases to cluster the ECG data, which often evolves over time. In order to address these issues, an adaptive learning ECG classification method that uses Dirichlet process Gaussian mixture models is proposed. This approach does not place any restriction on the number of disease classes, nor does it require any training data. This algorithm is adapted to be patient-specific by labeling or identifying the generated mixtures using the Bayesian ML method, assuming the availability of labeled training data. / Dissertation/Thesis / Ph.D. Electrical Engineering 2012

abstract: Electrical neural activity detection and tracking have many applications in medical research and brain computer interface technologies. In this thesis, we focus on the development of advanced signal processing algorithms to track neural activity and on the mapping of these algorithms onto hardware to enable real-time tracking. At the heart of these algorithms is particle filtering (PF), a sequential Monte Carlo technique used to estimate the unknown parameters of dynamic systems. First, we analyze the bottlenecks in existing PF algorithms, and we propose a new parallel PF (PPF) algorithm based on the independent Metropolis-Hastings (IMH) algorithm. We show that the proposed PPF-IMH algorithm improves the root mean-squared error (RMSE) estimation performance, and we demonstrate that a parallel implementation of the algorithm results in significant reduction in inter-processor communication. We apply our implementation on a Xilinx Virtex-5 field programmable gate array (FPGA) platform to demonstrate that, for a one-dimensional problem, the PPF-IMH architecture with four processing elements and 1,000 particles can process input samples at 170 kHz by using less than 5% FPGA resources. We also apply the proposed PPF-IMH to waveform-agile sensing to achieve real-time tracking of dynamic targets with high RMSE tracking performance. We next integrate the PPF-IMH algorithm to track the dynamic parameters in neural sensing when the number of neural dipole sources is known. We analyze the computational complexity of a PF based method and propose the use of multiple particle filtering (MPF) to reduce the complexity. We demonstrate the improved performance of MPF using numerical simulations with both synthetic and real data. We also propose an FPGA implementation of the MPF algorithm and show that the implementation supports real-time tracking. For the more realistic scenario of automatically estimating an unknown number of time-varying neural dipole sources, we propose a new approach based on the probability hypothesis density filtering (PHDF) algorithm. The PHDF is implemented using particle filtering (PF-PHDF), and it is applied in a closed-loop to first estimate the number of dipole sources and then their corresponding amplitude, location and orientation parameters. We demonstrate the improved tracking performance of the proposed PF-PHDF algorithm and map it onto a Xilinx Virtex-5 FPGA platform to show its real-time implementation potential. Finally, we propose the use of sensor scheduling and compressive sensing techniques to reduce the number of active sensors, and thus overall power consumption, of electroencephalography (EEG) systems. We propose an efficient sensor scheduling algorithm which adaptively configures EEG sensors at each measurement time interval to reduce the number of sensors needed for accurate tracking. We combine the sensor scheduling method with PF-PHDF and implement the system on an FPGA platform to achieve real-time tracking. We also investigate the sparsity of EEG signals and integrate compressive sensing with PF to estimate neural activity. Simulation results show that both sensor scheduling and compressive sensing based methods achieve comparable tracking performance with significantly reduced number of sensors. / Dissertation/Thesis / Ph.D. Electrical Engineering 2013

La croissance des plantes en interaction avec l'environnement peut être décrite par des modèles mathématiques. Ceux-ci présentent des perspectives prometteuses pour un nombre considérable d'applications telles que la prévision des rendements ou l'expérimentation virtuelle dans le contexte de la sélection variétale. Dans cette thèse, nous nous intéressons aux différentes solutions capables d'améliorer les capacités prédictives des modèles de croissance de plantes, en particulier grâce à des méthodes statistiques avancées. Notre contribution se résume en quatre parties.Tout d'abord, nous proposons un nouveau modèle de culture (Log-Normal Allocation and Senescence ; LNAS). Entièrement construit dans un cadre probabiliste, il décrit seulement les processus écophysiologiques essentiels au bilan de la biomasse végétale afin de contourner les problèmes d'identification et d'accentuer l'évaluation des incertitudes. Ensuite, nous étudions en détail le paramétrage du modèle. Dans le cadre Bayésien, nous mettons en œuvre des méthodes Monte-Carlo Séquentielles (SMC) et des méthodes de Monte-Carlo par Chaînes de Markov (MCMC) afin de répondre aux difficultés soulevées lors du paramétrage des modèles de croissance de plantes, caractérisés par des équations dynamiques non-linéaires, des données rares et un nombre important de paramètres. Dans les cas où la distribution a priori est peu informative, voire non-informative, nous proposons une version itérative des méthodes SMC et MCMC, approche équivalente à une variante stochastique d'un algorithme de type Espérance-Maximisation, dans le but de valoriser les données d'observation tout en préservant la robustesse des méthodes Bayésiennes. En troisième lieu, nous soumettons une méthode d'assimilation des données en trois étapes pour résoudre le problème de prévision du modèle. Une première étape d'analyse de sensibilité permet d'identifier les paramètres les plus influents afin d'élaborer une version plus robuste de modèle par la méthode de sélection de modèles à l'aide de critères appropriés. Ces paramètres sélectionnés sont par la suite estimés en portant une attention particulière à l'évaluation des incertitudes. La distribution a posteriori ainsi obtenue est considérée comme information a priori pour l'étape de prévision, dans laquelle une méthode du type SMC telle que le filtrage par noyau de convolution (CPF) est employée afin d'effectuer l'assimilation de données. Dans cette étape, les estimations des états cachés et des paramètres sont mis à jour dans l'objectif d'améliorer la précision de la prévision et de réduire l'incertitude associée. Finalement, d'un point de vue applicatif, la méthodologie proposée est mise en œuvre et évaluée avec deux modèles de croissance de plantes, le modèle LNAS pour la betterave sucrière et le modèle STICS pour le blé d'hiver. Quelques pistes d'utilisation de la méthode pour l'amélioration du design expérimental sont également étudiées, dans le but d'améliorer la qualité de la prévision. Les applications aux données expérimentales réelles montrent des performances prédictives encourageantes, ce qui ouvre la voie à des outils d'aide à la décision en agriculture. / Plant growth models aim to describe plant development and functional processes in interaction with the environment. They offer promising perspectives for many applications, such as yield prediction for decision support or virtual experimentation inthe context of breeding. This PhD focuses on the solutions to enhance plant growth model predictive capacity with an emphasis on advanced statistical methods. Our contributions can be summarized in four parts. Firstly, from a model design perspective, the Log-Normal Allocation and Senescence (LNAS) crop model is proposed. It describes only the essential ecophysiological processes for biomass budget in a probabilistic framework, so as to avoid identification problems and to accentuate uncertainty assessment in model prediction. Secondly, a thorough research is conducted regarding model parameterization. In a Bayesian framework, both Sequential Monte Carlo (SMC) methods and Markov chain Monte Carlo (MCMC) based methods are investigated to address the parameterization issues in the context of plant growth models, which are frequently characterized by nonlinear dynamics, scarce data and a large number of parameters. Particularly, whenthe prior distribution is non-informative, with the objective to put more emphasis on the observation data while preserving the robustness of Bayesian methods, an iterative version of the SMC and MCMC methods is introduced. It can be regarded as a stochastic variant of an EM type algorithm. Thirdly, a three-step data assimilation approach is proposed to address model prediction issues. The most influential parameters are first identified by global sensitivity analysis and chosen by model selection. Subsequently, the model calibration is performed with special attention paid to the uncertainty assessment. The posterior distribution obtained from this estimation step is consequently considered as prior information for the prediction step, in which a SMC-based on-line estimation method such as Convolution Particle Filtering (CPF) is employed to perform data assimilation. Both state and parameter estimates are updated with the purpose of improving theprediction accuracy and reducing the associated uncertainty. Finally, from an application point of view, the proposed methodology is implemented and evaluated with two crop models, the LNAS model for sugar beet and the STICS model for winter wheat. Some indications are also given on the experimental design to optimize the quality of predictions. The applications to real case scenarios show encouraging predictive performances and open the way to potential tools for yield prediction in agriculture.

Rapid stress release at the source of an earthquake produces seismic waves. Observations of the particle motions from such waves are used in source inversion to characterise the dynamic behaviour of the source and to help in understanding the driving processes. Earthquakes either occur naturally, such as in volcanic eruptions and natural geothermal fields, or are linked to anthropogenic activities including hydrofracture of gas and oil reservoirs, mining events and extraction of geothermal fluids. Source inversion is very sensitive to uncertainties in both the model and the data, especially for low magnitude, namely microseismic, events. Many of the uncertainties can be poorly quantified, and are often not included in source inversion. This thesis proposes a Bayesian framework enabling a complete inclusion of uncertainties in the resultant probability distribution using Bayesian marginalisation. This approach is developed for polarity and amplitude ratio data, although it is possible to use any data type, provided the noise model can be estimated. The resultant posterior probability distributions are easily visualised on different plots for orientation and source-type. Several different algorithms can be used to search the source space, including Monte Carlo random sampling and Markov chain Monte Carlo sampling. Relative information between co-located events may be used as an extension to the framework, improving the constraint on the source. The double-couple source is the commonly assumed source model for many earthquakes, corresponding to slip on a fault plane. Two methods for estimating the posterior model probability of the double-couple source type are explored, one using the Bayesian evidence, the other using trans-dimensional Markov chain Monte Carlo sampling. Results from both methods are consistent with each other, producing good estimates of the probability given sufficient samples. These provide estimates of the probability of the source being a double-couple source or not, which is very useful when trying to understand the processes causing the earthquake. Uncertainty on the polarity estimation is often hard to characterise, so an alternative approach for determining the polarity and its associated uncertainty is proposed. This uses a Bayesian estimate of the polarity probability and includes both the background noise and the arrival time pick uncertainty, resulting in a more quantitative estimate of the polarity uncertainty. Moreover, this automated approach can easily be included in automatic event detection and location workflows. The inversion approach is discussed in detail and then applied to both synthetic events generated using a finite-difference code, and to real events acquired from a temporary seismometer network deployed around the Askja and Krafla Volcanoes, Iceland.

Mendelian randomization is an epidemiological method for using genetic variationto estimate the causal effect of the change in a modifiable phenotype onan outcome from observational data. A genetic variant satisfying the assumptionsof an instrumental variable for the phenotype of interest can be usedto divide a population into subgroups which differ systematically only in thephenotype. This gives a causal estimate which is asymptotically free of biasfrom confounding and reverse causation. However, the variance of the causalestimate is large compared to traditional regression methods, requiring largeamounts of data and necessitating methods for efficient data synthesis. Additionally,if the association between the genetic variant and the phenotype is notstrong, then the causal estimates will be biased due to the “weak instrument”in finite samples in the direction of the observational association. This biasmay convince a researcher that an observed association is causal. If the causalparameter estimated is an odds ratio, then the parameter of association willdiffer depending on whether viewed as a population-averaged causal effect ora personal causal effect conditional on covariates. We introduce a Bayesian framework for instrumental variable analysis, whichis less susceptible to weak instrument bias than traditional two-stage methods,has correct coverage with weak instruments, and is able to efficiently combinegene–phenotype–outcome data from multiple heterogeneous sources. Methodsfor imputing missing genetic data are developed, allowing multiple genetic variantsto be used without reduction in sample size. We focus on the question ofa binary outcome, illustrating how the collapsing of the odds ratio over heterogeneousstrata in the population means that the two-stage and the Bayesianmethods estimate a population-averaged marginal causal effect similar to thatestimated by a randomized trial, but which typically differs from the conditionaleffect estimated by standard regression methods. We show how thesemethods can be adjusted to give an estimate closer to the conditional effect. We apply the methods and techniques discussed to data on the causal effect ofC-reactive protein on fibrinogen and coronary heart disease, concluding withan overall estimate of causal association based on the totality of available datafrom 42 studies.

Direct marketing problems have been intensively reviewed in the marketing literature recently, such as purchase frequency and time, sales profit, and brand choices. However, modeling the customer response, which is an important issue in direct marketing research, remains a significant challenge. This thesis is an empirical study of predicting customer response to direct marketing and applies a Bayesian approach, including the Bayesian Binary Regression (BBR) and the Hierarchical Bayes (HB). Other classical methods, such as Logistic Regression and Latent Class Analysis (LCA), have been conducted for the purpose of comparison. The results of comparing the performance of all these techniques suggest that the Bayesian methods are more appropriate in predicting direct marketing customer responses. Specifically, when customers are analyzed as a whole group, the Bayesian Binary Regression (BBR) has greater predictive accuracy than Logistic Regression. When we consider customer heterogeneity, the Hierarchical Bayes (HB) models, which use demographic and geographic variables for clustering, do not match the performance of Latent Class Analysis (LCA). Further analyses indicate that when latent variables are used for clustering, the Hierarchical Bayes (HB) approach has the highest predictive accuracy.

Identifying a peptide on the basis of a scan from a mass spectrometer is an important yet highly challenging problem. To identify peptides, we present a Bayesian approach which uses prior information about the average relative abundances of bond cleavages and the prior probability of any particular amino acid sequence. The scoring function proposed is composed of two overall distance measures, which measure how close an observed spectrum is to a theoretical scan for a peptide. Our use of our scoring function, which approximates a likelihood, has connections to the generalization presented by Bissiri and co-workers of the Bayesian framework. A Markov chain Monte Carlo algorithm is employed to simulate candidate choices from the posterior distribution of the peptide sequence. The true peptide is estimated as the peptide with the largest posterior density.

Annotated textual corpora are an essential language resource, facilitating manual search and discovery as well as supporting supervised Natural Language Processing (NLP) techniques designed to accomplishing a variety of useful tasks. However, manual annotation of large textual corpora can be cost-prohibitive, especially for rare and under-resourced languages. For this reason, developers of annotated corpora often attempt to reduce annotation cost by offering annotators various forms of machine assistance intended to increase annotator speed and accuracy. This thesis contributes to the field of annotated corpus development by providing tools and methodologies for empirically evaluating the effectiveness of machine assistance techniques. This allows developers of annotated corpora to improve annotator efficiency by choosing to employ only machine assistance techniques that make a measurable, positive difference. We validate our tools and methodologies using a concrete example. First we present CCASH, a platform for machine-assisted online linguistic annotation capable of recording detailed annotator performance statistics. We employ CCASH to collect data detailing the performance of annotators engaged in syriac morphological analysis in the presence of two machine assistance techniques: pre-annotation and correction propagation. We conduct a preliminary analysis of the data using the traditional approach of comparing mean data values. We then demonstrate a Bayesian analysis of the data that yields deeper insights into our data. Pre-annotation is shown to increase annotator accuracy when pre-annotations are at least 60% accurate, and annotator speed when pre-annotations are at least 80% accurate. Correction propagation's effect on accuracy is minor. The Bayesian analysis indicates that correction propagation has a positive effect on annotator speed after accounting for the effects of the particular visual mechanism we employed to implement it.

A child's parents are the primary source of knowledge and learned behaviors for developing children, and the benefits or repercussions of certain parental practices can be long lasting. Although parenting practices affect behavioral outcomes for children, families tend to be diverse in their circumstances and needs. Research attempting to ascertain cause and effect relationships between parental influences and child behavior can be difficult due to the complex nature of family dynamics and the intricacies of real life. Structural equation modeling (SEM) is an appropriate method for this research as it is able to account for the complicated nature of child-parent relationships. Both Frequentist and Bayesian methods are used to estimate the effect of latent parental behavior variables on child aggression and anxiety in order to allow for comparison and contrast between the two statistical paradigms in the context of structural equation modeling. Estimates produced from both methods prove to be comparable, but subtle differences do exist in those coefficients and in the conclusions to which a researcher would arrive. Although model estimates between the two paradigms generally agree, they diverge in the model selection process. The mother's behaviors are estimated to be the most influential on child aggression, while the influence of the father, socio-economic status, parental involvement, and the relationship quality of the couple also prove to be significant in predicting child aggression.