L'objectif de cette thèse est de proposer des méthodes fiables, efficaces et rapides pour minimiser des critères convexes apparaissant dans la résolution de problèmes inverses en imagerie. Ainsi, nous nous intéresserons à des problèmes de restauration/reconstruction lorsque les données sont dégradées par un opérateur linéaire et un bruit qui peut être non additif. La fiabilité de la méthode sera assurée par l'utilisation d'algorithmes proximaux dont la convergence est garantie lorsqu'il s'agit de minimiser des critères convexes. La quête d'efficacité impliquera le choix d'un critère adapté aux caractéristiques du bruit, à l'opérateur linéaire et au type d'image à reconstruire. En particulier, nous utiliserons des termes de régularisation basés sur la variation totale et/ou favorisant la parcimonie des coefficients du signal recherché dans une trame. L'utilisation de trames nous amènera à considérer deux approches : une formulation du critère à l'analyse et une formulation du critère à la synthèse. De plus, nous étendrons les algorithmes proximaux et leurs preuves de convergence aux cas de problèmes inverses multicomposantes. La recherche de la rapidité de traitement se traduira par l'utilisation d'algorithmes proximaux parallélisables. Les résultats théoriques obtenus seront illustrés sur différents types de problèmes inverses de grandes tailles comme la restauration d'images mais aussi la stéréoscopie, l'imagerie multispectrale, la décomposition en composantes de texture et de géométrie. Une application attirera plus particulièrement notre attention ; il s'agit de la reconstruction de l'activité dynamique en Tomographie par Emission de Positrons (TEP) qui constitue un problème inverse difficile mettant en jeu un opérateur de projection et un bruit de Poisson dégradant fortement les données observées. Pour optimiser la qualité de reconstruction, nous exploiterons les caractéristiques spatio-temporelles de l'activité dans les tissus / The objective of this work is to propose reliable, efficient and fast methods for minimizing convex criteria, that are found in inverse problems for imagery. We focus on restoration/reconstruction problems when data is degraded with both a linear operator and noise, where the latter is not assumed to be necessarily additive.The methods reliability is ensured through the use of proximal algorithms, the convergence of which is guaranteed when a convex criterion is considered. Efficiency is sought through the choice of criteria adapted to the noise characteristics, the linear operators and the image specificities. Of particular interest are regularization terms based on total variation and/or sparsity of signal frame coefficients. As a consequence of the use of frames, two approaches are investigated, depending on whether the analysis or the synthesis formulation is chosen. Fast processing requirements lead us to consider proximal algorithms with a parallel structure. Theoretical results are illustrated on several large size inverse problems arising in image restoration, stereoscopy, multi-spectral imagery and decomposition into texture and geometry components. We focus on a particular application, namely Positron Emission Tomography (PET), which is particularly difficult because of the presence of a projection operator combined with Poisson noise, leading to highly corrupted data. To optimize the quality of the reconstruction, we make use of the spatio-temporal characteristics of brain tissue activity

La modélisation de l’électrophysiologie in silico a été un sujet de recherche important ces dernières décennies. Afin de pouvoir utiliser ces progrès importants dans les applications cliniques, il faut mettre en place des modèles macroscopiques qui peuvent être utilisés pour la planification et le guidage des procédures cliniques.L’objectif de cette thèse est de construire de tels modèles macroscopiques spécifiques à chaque patient pour le diagnostic et la prévision, dans le but d’améliorer la planification et le guidage de l’ablation par radio-fréquence (ARF) des patients souffrant de tachycardie ventriculaire (TV) après infarctus. Dans ce travail, nous avons proposé un cadre pour la personnalisation d’un modèle cardiaque 3D, le modèle de Mitchell-Schaeffer (MS), et nous avons évalué sa puissance prédictive dans plusieurs configurations de stimulation. Ceci a été réalisé sur des données ex vivo de cœurs porcins à l’aide d’images médicales et de données cartographiques optiques de l’épicarde. Ce cadre a ensuite été appliqué à un ensemble de données cliniques provenant d’imagerie hybride XMR et d’une procédure de cartographie électrophysiologique sur un patient souffrant d’insuffisance cardiaque.Ensuite, le modèle 3D MS a également été adapté pour simuler le comportement macroscopique structural de la fibrose près des cicatrices. La simulation d’une étude in silico de stimulation de TV en utilisant le modèle adapté personnalisé MS a été réalisée pour quantifier le risque de TV en termes de cartes d’inductibilité, de réentrées des modèles et de cartes de points de sortie. Une approche de modélisation pour l’ablation par RF fondée sur l’état de l’art a été proposée. Enfin, l’étude in silico de stimulation de TV a été appliquée aux données in vivo personnalisées des patients, qui ont suivi ce protocole. Ceci a permis une validation de la prévision in silico de TV post-infarctus par comparaison avec la TV clinique induite. Ler ôle de l’hétérogénéité spatiale des propriétés des tissus cardiaques estimés dans la genèse de TV ischémique a été évalué, ainsi que les caractéristiques des points de sortie, qui sont les candidats potentiels à l’ablation par RF. / Modelling cardiac electrophysiology for arrhythmias in silico has been an important research topic for the last decades. In order to translate this important progress into clinical applications, there is a requirement to make macroscopic models that can be used for the planning and guidance of clinical procedures. The objective of this thesis was to construct such macroscopic EP models specifict o each patient for study and prediction, in order to improve the planning and guidance of radio frequency ablation (RFA) the rapieson patients suffering from post infarction Ventricular Tachycardia (VT). In this work, we proposed a framework for the personalisation of a 3D cardiac EP model, the Mitchell-Schaeffer (MS) model, an devaluated its volumetric predictive power under various pacing scenarios.This was performed on ex vivo large porcine healthy heart susing Diffusion Tensor MRI (DT-MRI) and dense optical mapping data of the epicardium. This framework was then also applied to a clinical dataset derived from a hybrid XMR imaging and sparse electroanatomical mapping on a patient with heart failure. Next, the 3DMS model was also adapted to simulate the macroscopic structural behaviour of fibrosis near the scars. The simulation of an in silico VT stimulation study using the personalised adapted MS model was then performed, to quantify VT risk in terms of inducibility maps, re-entry patterns and exit point maps. A rule-based modelling approach for RF ablation lesions based on state of the art studies was proposed. Lastly, the in silico VT stimulation study was applied to in vivo personalised data of patients who underwent a clinical VT stimulation study. A validation of the in silico post-infarct VT prediction was performed against the clinically induced VT. Therole of spatial heterogeneity of the estimated patient’s cardiac tissue properties in the genesis of ischemic VT was learnt, along with their characteristics for entry/exit points, which are the potential candidates for RF ablation.

La recherche sur les méthodes bayésiennes non-paramétriques connaît un essor considérable depuis les vingt dernières années notamment depuis le développement d'algorithmes de simulation permettant leur mise en pratique. Il est donc nécessaire de comprendre, d'un point de vue théorique, le comportement de ces méthodes. Cette thèse présente différentes contributions à l'analyse des propriétés fréquentistes des méthodes bayésiennes non-paramétriques. Si se placer dans un cadre asymptotique peut paraître restrictif de prime abord, cela permet néanmoins d'appréhender le fonctionnement des procédures bayésiennes dans des modèles extrêmement complexes. Cela permet notamment de détecter les aspects de l'a priori particulièrement influents sur l’inférence. De nombreux résultats généraux ont été obtenus dans ce cadre, cependant au fur et à mesure que les modèles deviennent de plus en plus complexes, de plus en plus réalistes, ces derniers s'écartent des hypothèses classiques et ne sont plus couverts par la théorie existante. Outre l'intérêt intrinsèque de l'étude d'un modèle spécifique ne satisfaisant pas les hypothèses classiques, cela permet aussi de mieux comprendre les mécanismes qui gouvernent le fonctionnement des méthodes bayésiennes non-paramétriques. / Research on Bayesian nonparametric methods has received a growing interest for the past twenty years, especially since the development of powerful simulation algorithms which makes the implementation of complex Bayesian methods possible. From that point it is necessary to understand from a theoretical point of view the behaviour of Bayesian nonparametric methods. This thesis presents various contributions to the study of frequentist properties of Bayesian nonparametric procedures. Although studying these methods from an asymptotic angle may seems restrictive, it allows to grasp the operation of the Bayesian machinery in extremely complex models. Furthermore, this approach is particularly useful to detect the characteristics of the prior that are strongly influential in the inference. Many general results have been proposed in the literature in this setting, however the more complex and realistic the models the further they get from the usual assumptions. Thus many models that are of great interest in practice are not covered by the general theory. If the study of a model that does not fall under the general theory has an interest on its owns, it also allows for a better understanding of the behaviour of Bayesian nonparametric methods in a general setting.

<p> In this thesis, both direct and inverse elastic scattering problems are considered. For a given incident wave, the direct problem is to determine the displacement of wave field from the known structure, which could be an obstacle or a surface in this thesis; The inverse problem is to determine the structure from the measurement of displacement on an artificial boundary.</p><p>In the second chapter, we consider the scattering of an elastic plane wave by a rigid obstacle, which is immersed in a homogeneous and isotropic elastic medium in two dimensions. Based on a Dirichlet-to-Neumann (DtN) operator, an exact transparent boundary condition is introduced and the scattering problem is formulated as a boundary value problem of the elastic wave equation in a bounded domain. By developing a new duality argument, an a posteriori error estimate is derived for the discrete problem by using the finite element method with the truncated DtN operator. The a posteriori error estimate consists of the finite element approximation error and the truncation error of the DtN operator which decays exponentially with respect to the truncation parameter. An adaptive finite element algorithm is proposed to solve the elastic obstacle scattering problem, where the truncation parameter is determined through the truncation error and the mesh elements for local refinements are chosen through the finite element discretization error.<br></p><p>In chapter 3, we extend the argument developed in chapter 2 to elastic surface grating problem, where the surface is assumed to be periodic and elastic rigid; Then, we treat the obstacle scattering in three dimensional space; The direct problem is shown to have a unique weak solution by examining its variational formulation. The domain derivative is studied and a frequency continuation method is developed for the inverse problem. Finally, in chapter 4, a rigorous mathematical model and an efficient computational method are proposed to solve the inverse elastic surface scattering problem which arises from the near-field imaging of periodic structures. The surface is assumed to be a small and smooth perturbation of an elastically rigid plane. By placing a rectangle slab of a homogeneous and isotropic elastic medium with larger mass density above the surface, more propagating wave modes can be utilized from the far-field data which contributes to the reconstruction resolution. Requiring only a single illumination, the method begins with the far-to-near field data conversion and utilized the transformed field expansion to derive an analytic solution for the direct problem, which leads to an explicit inversion formula for the inverse problem; Moreover, a nonlinear correction scheme is developed to improve the accuracy of the reconstruction; Numerical examples are presented to demonstrate the effectiveness of the proposed methods for solving the questions mentioned above.<br></p>

[pt] O problema de estimação pode ser entendido como um caso particular dos problemas inversos. Dadas observações da resposta de um sistema para certas causas, deseja-se estimar certas características do sistema. Essas características, em um sistema dinâmico, geralmente são representadas por parâmetros. Assim, para uma representação físico-matemática do sistema, dada uma excitação e observando a resposta, é possível obter uma estimação dos parâmetros. A estimação paramétrica é de grande importância e utilizada em diversas situações, desde experimentalistas, ao observar fenômenos no laboratório, até quem estuda o comportamento de setores sociais por amostras populacionais. A parte inicial desta dissertação apresenta uma breve introdução ao problema inverso do marco da estatística Bayesiana. Neste marco trata-se a estimação paramétrica como resultado da resolução de um problema inverso. Duas técnicas de estimação s ao deduzidas a partir da inferência estatística Bayesiana. A primeira delas, mínimos quadrados, coleta todos os dados e logo faz a estimação. A segunda, filtro de Kalman (e filtro de Kalman extendido), melhora o estado do conhecimento dos parâmetros a serem estimados a cada nova observação. Para a abordagem destas técnicas de estimação, de modo de poder compará-las, é apresentada a resolução analítica de um sistema harmônico de um e dois graus de liberdade. Por último, é apresentada uma modelagem de uma bancada experimental, em escala de laboratório, que emula uma coluna de perfura ção acoplada a um motor. Esta bancada foi desenvolvida para estudos de dinâmica torcional, na dissertação de mestrado de Bruno C. Cayres A., de modo que aqui só é de interesse a caracterização da mesma. As técnicas de estimação paramétrica são usadas de forma teórica, simulando os dados a partir de soluções analíticas para diferentes parâmetros da modelagem do motor e da coluna. Também usa-se medições feitas na bancada para estimar os parâmetros da modelagem, obtendo assim um conhecimento melhorado dos parâmetros envolvidos no sistema coluna-motor. / [en] The estimation problem can be understood as a particular case of an inverse problem. Given observations of the response of a system, due to certain causes, one wants to estimate certain characteristics of the problem. These features, in a dynamic system, are usually represented by parameters. Thus, for a mathematical representation of the physical system, given an excitation and given the observing response, it is possible to give an estimation of the parameters. The parameter estimation is of great importance and used in countless situations, such as experimental obseration of a phenomena in the laboratory or even by those who study the behaviors social sectors by population samples. The initial part of this dissertation presents a brief introduction to the inverse problem the framework of the Bayesian statistics. In this context, the parametric estimation is a result of the resolution of an inverse problem. Two estimation techniques are derived from the Bayesian statistical inference. The first of these, least squares, collects all the data and then makes the estimation. The second, Kalman filter (and extended filter Kalman), improves the state of knowledge of the parameters to be estimated, with each new observation. To address these estimation techniques, in order to be able to compare them, presents the analytical resolution of a harmonious system of one and two degrees of freedom. Finally, it is presented a model for an experimental setup, in laboratory scale, which emulates a drillstring coupled to a motor. This experimental setup was developed to study the dynamic torsional and by the author of the dissertation of Bruno C. Cayres A., the mode that is of interest here only the characterization of it. These techniques are used for parameter estimation in theoretical way, simulating data from the analytical solutions, for different parameters involved in the column-motor modeling. Also, we use measurements obtained from the experimental setup to estimate the parameters of the column-motor model. Thereby, we obtain an improved knowledge of the parameters involved in the column-motor.

L'objectif de ce travail est de développer de nouvelles méthodes statistiques de reconstruction d'image spatiale (3D) et spatio-temporelle (3D+t) en Tomographie par Émission de Positons (TEP). Le but est de proposer des méthodes efficaces, capables de reconstruire des images dans un contexte de faibles doses injectées tout en préservant la qualité de l'interprétation. Ainsi, nous avons abordé la reconstruction sous la forme d'un problème inverse spatial et spatio-temporel (à observations ponctuelles) dans un cadre bayésien non paramétrique. La modélisation bayésienne fournit un cadre pour la régularisation du problème inverse mal posé au travers de l'introduction d'une information dite a priori. De plus, elle caractérise les grandeurs à estimer par leur distribution a posteriori, ce qui rend accessible la distribution de l'incertitude associée à la reconstruction. L'approche non paramétrique quant à elle pourvoit la modélisation d'une grande robustesse et d'une grande flexibilité. Notre méthodologie consiste à considérer l'image comme une densité de probabilité dans (pour une reconstruction en k dimensions) et à chercher la solution parmi l'ensemble des densités de probabilité de . La grande dimensionalité des données à manipuler conduit à des estimateurs n'ayant pas de forme explicite. Cela implique l'utilisation de techniques d'approximation pour l'inférence. La plupart de ces techniques sont basées sur les méthodes de Monte-Carlo par chaînes de Markov (MCMC). Dans l'approche bayésienne non paramétrique, nous sommes confrontés à la difficulté majeure de générer aléatoirement des objets de dimension infinie sur un calculateur. Nous avons donc développé une nouvelle méthode d'échantillonnage qui allie à la fois bonnes capacités de mélange et possibilité d'être parallélisé afin de traiter de gros volumes de données. L'approche adoptée nous a permis d'obtenir des reconstructions spatiales 3D sans nécessiter de voxellisation de l'espace, et des reconstructions spatio-temporelles 4D sans discrétisation en amont ni dans l'espace ni dans le temps. De plus, on peut quantifier l'erreur associée à l'estimation statistique au travers des intervalles de crédibilité. / The aim of this work is to develop new statistical methods for spatial (3D) and space-time (3D+t) Positron Emission Tomography (PET) reconstruction. The objective is to propose efficient reconstruction methods in a context of low injected doses while maintaining the quality of the interpretation. We tackle the reconstruction problem as a spatial or a space-time inverse problem for point observations in a \Bayesian nonparametric framework. The Bayesian modeling allows to regularize the ill-posed inverse problem via the introduction of a prior information. Furthermore, by characterizing the unknowns with their posterior distributions, the Bayesian context allows to handle the uncertainty associated to the reconstruction process. Being nonparametric offers a framework for robustness and flexibility to perform the modeling. In the proposed methodology, we view the image to reconstruct as a probability density in(for reconstruction in k dimensions) and seek the solution in the space of whole probability densities in . However, due to the size of the data, posterior estimators are intractable and approximation techniques are needed for posterior inference. Most of these techniques are based on Markov Chain Monte-Carlo methods (MCMC). In the Bayesian nonparametric approach, a major difficulty raises in randomly sampling infinite dimensional objects in a computer. We have developed a new sampling method which combines both good mixing properties and the possibility to be implemented on a parallel computer in order to deal with large data sets. Thanks to the taken approach, we obtain 3D spatial reconstructions without any ad hoc space voxellization and 4D space-time reconstructions without any discretization, neither in space nor in time. Furthermore, one can quantify the error associated to the statistical estimation using the credibility intervals.

La neuromodulation offre une possibilité de traitement pour des pathologies pharmoco-resistantes. Dans ce domaine, l'émergence de matrices d'électrodes à l'échelle microscopique ouvre la voie à des interfaces neurales sélectives. Cependant, leur fonctionnalité est réduite par le manque d'information sur l'anatomie fonctionnelle du nerf ciblé. L'objectif global de ce projet de thèse est d'explorer les possibilités d'imager un nerf de manière non-invasive par tomographie d'impédance électrique (EIT). Modalité d'imagerie des tissus mous, l'EIT déduit des cartes de conductivité à partir de mesures sur la frontière du domaine étudié. Une plateforme expérimentale a été mise en place et a permis de valider les développements des méthodes numériques effectués pour la prédiction des données et l'estimation des paramètres. Des tests in vivo ont été réalisés dans le contexte de la stimulation du nerf vague et du nerf sciatique. Des spécifications pour de futures expériences ont été déduites, avec l'utilisation d'électrodes plus robustes comprenant un plus grand nombre de contacts par section. / Neuromodulation offers treatments for drug resistant pathologies. In this field, the emergence of micro-scale multi-electrode arrays paves the way for selective neural interfaces. But they suffer from the lack of information on the nerve functional anatomy. The global aim of this PhD project is to explore the possibilities of imaging the inside of a nerve in a non-invasive way through electrical impedance tomography (EIT). As a soft-field imaging modality, EIT infers conductivity maps from boundary measurements. An experimental platform was built and allowed the validation of numerical methods developed for data prediction and parameter estimation. In vivo tests were performed in the context of vagus and sciatic nerve stimulation. Specifications were deduced for future experiments, with more reliable electrodes, embedding a higher number of contacts per cross-section.

Submitted by Maria Cristina (library@lncc.br) on 2017-08-14T18:34:35Z No. of bitstreams: 1 Welerson_Dissertação.pdf: 2360212 bytes, checksum: 58a44b5d4fff215888d80a0a417700de (MD5) / Approved for entry into archive by Maria Cristina (library@lncc.br) on 2017-08-14T18:34:46Z (GMT) No. of bitstreams: 1 Welerson_Dissertação.pdf: 2360212 bytes, checksum: 58a44b5d4fff215888d80a0a417700de (MD5) / Made available in DSpace on 2017-08-14T18:34:57Z (GMT). No. of bitstreams: 1 Welerson_Dissertação.pdf: 2360212 bytes, checksum: 58a44b5d4fff215888d80a0a417700de (MD5) Previous issue date: 2016-11-04 / In this work the optimal control problem with respect to a set of pointwise sources is studied. In particular, the control is given by a finite linear combination of Dirac mass and the state is solution to the associated elliptic boundary value problem. The basic idea consists in minimizing a functional which measures the distance between the state and a target function, with respect to the number, intensities and locations of pointwise loads. The sensitivity of the cost functional with respect to a number of pointwise sources in the set of admissible solutions is derived in its explicit form with help of auxiliaries boundary value problems. The obtained result is then used to devise a non-iterative second order reconstruction algorithm, independent of any initial guess and without introducing regularization techniques. Finally, the devised reconstruction algorithm is applied for numerically solving a set of control and inverse reconstruction problems. / Neste trabalho o problema de controle ótimo com respeito a um conjunto de fontes puntuais é estudado. Em particular, o controle é dado por uma combinação linear finita de massas de Dirac e o estado é solução de um problema de valor de contorno elíptico. Objetiva-se, portanto, minimizar um funcional, que mede a distância entre o estado e uma função alvo, com respeito ao número, intensidades e localizações das cargas puntuais. A sensibilidade do funcional de custo, em relação a um certo número de fontes puntuais no conjunto de soluções admissíveis, é analisada na sua forma explícita com o auxílio de problemas de valor de contorno auxiliares. O resultado obtido é então utilizado para conceber um algoritmo de reconstrução de segunda ordem não iterativo, independente de qualquer chute inicial e sem a introdução de técnicas de regularização. Finalmente, o algoritmo de reconstrução elaborado é aplicado para resolver numericamente um conjunto de problemas de controle e de problemas inversos de reconstrução de fontes.

Neste estudo, o novo algoritmo hibrido autoconfigurado PSOS (Particle Swarm Optimization - Simplex) para avaliação da integridade estrutural a partir de respostas dinâmicas é apresentado. A formulação da função objetivo para o problema de minimização definido emprega funções de resposta em freqüência e/ou dados modais do sistema. Uma nova estratégia para o controle dos parâmetros do algoritmo Particle Swarm Optimization (PSO), baseada no uso do método de Nelder - Mead é desenvolvida; conseqüentemente, a convergência do PSO fica independente dos parâmetros heurísticos e sua estabilidade e precisão são melhoradas. O método híbrido proposto teve melhor desempenho, nas diversas funções teste analisadas, quando comparado com os algoritmos simulated annealing, algoritmos genéticos e o PSO. São apresentados diversos problemas de detecção de dano, levando em conta os efeitos do ruído e da falta de dados experimentais. Em todos os casos, a posição e extensão do dano foram determinadas com sucesso. Finalmente, usando o PSOS, os parâmetros de um oscilador não linear (oscilador de Duffing) foram identificados. / In this study, a new auto configured Particle Swarm Optimization - Simplex algorithm for damage detection has been proposed. The formulation of the objective function for the minimization problem is based on the frequency response functions (FRFs) and the modal parameters of the system. A novel strategy for the control of the Particle Swarm Optimization (PSO) parameters based on the Nelder-Mead algorithm (Simplex method) is presented; consequently, the convergence of the PSOS becomes independent of the heuristic constants and its stability and accuracy are enhanced. The formulated hybrid method performs better in different benchmark functions than the Simulated Annealing (SA), the Genetic Algorithm (GA) and the basic PSO. Several damage identification problems, taking into consideration the effects of noisy and incomplete data, were studied. In these cases, the damage location and extent were determined successfully. Finally, using the PSOS, a non-linear oscillator (Duffing oscillator) was identified with good results.

La méthode de diffusion inverse est la plus efficace dans la théorie des systèmes intégrables. Introduite dans les années soixantes, d'importants résultats ont été obtenus pour les problèmes de dimension 1+1 et notamment sur l'interaction de solitons. Depuis quelques années, l'intérêt est porté sur des problèmes de dimensions supérieures comme les équations de Davey-Sterwartson, une généralisation de l'équation intégrable de Schrödinger cubique non linéaire en dimension 1+1. Des études numériques en limite semi-classique de l'équation de Davey-Stewartson II (DSII) défocalisant, font apparaître des points communs avec le cas réduit unidimensionnel, par exemple sur l'existence d'ondes de choc dispersives : des conditions initiales lisses mènent à une région d'oscillations rapides et modulées dans le voisinage des chocs des solutions des équations non dispersives dotées des mêmes conditions initiales.Cette thèse donne les premières étapes pour l'étude analytique de ce problème basée sur la méthode de la transformée de diffusion inverse. Les deux types de méthodes, directe et inverse, pour l'équation de DSII permettent de réécrire le problème sous la forme des équations D-bar. On considère la transformée spectrale directe pour l'équation DSII avec des conditions initiales lisses en limite semi-classique. La transformée spectrale directe mène à un système de Dirac elliptique singulièrement perturbé en deux dimensions. On introduit une méthode de type BKW pour ce problème et on montre qu'il est bien défini pour des paramètres spectraux k ∈ ℂ dont les modules sont suffisamment grands en controllant la solution d'une équation eikonale non linéaire. Aussi cette méthode donne des résultats numériques précis pour de tels k en limite semi-classique. Ces résultats reposent sur la solution numérique du système de Dirac singulièrement perturbé et la solution numérique du problème eikonal.On résout le problème eikonal de manière explicite pout tout k dans le cas d'un potentiel particulier. Ces calculs donnent une explication sur le fait que l'on ne puisse pas appliquer la méthode BKW pour des valeurs de |k| plus petites. On présente une nouvelle méthode numérique pour calculer la solution du problème eikonal avec des valeurs de |k| suffisamment grandes.Les calculs numériques de la transformée spectrale directe offrent une manière d'analyser le système de Dirac singulièrement perturbé pour des valeurs de |k| si petites qu'il n'y a pas de solution globale au problème eikonal. On donne une analyse semi-classique rigoureuse sur la solution pour des potentiels radiaux en k = 0, ce qui donne une expression asymptotique du coefficient de réflexion pour k = 0 et suggère une structure annulaire pour la solution, ce qui peut être utilisé quand |k| ≠ 0 est petit. L'étude numérique suggère aussi que pour certains potentiels, le coefficient de réflexion converge simplement, quand ε ↓ 0, vers une fonction limite définie pour des valeurs de k pour lesquelles le problème eikonal n'a pas de solution globale. On propose que les singularités de la fonction eikonale jouent un rôle aussi similaire que les points tournants de la théorie unidimensionelle. / Inverse scattering is the most powerful tool in theory of integrable systems. Starting in the late sixties resounding great progress was made in (1+1) dimensional problems with many break-through results as on soliton interactions. Naturally the attention in recent years turns towards higher dimensional problems as the Davey-Stewartson equations, an integrable generalisation of the (1+1)-dimensionalcubic nonlinear Schrödinger equation. The defocusing Davey-Stewartson II equation, in its semi-classical limit has been shown in numerical experiments to exhibit behavior that qualitatively resembles that of its one-dimensional reduction, namely the generation of a dispersive shock wave: smooth initial data develop a zone rapid modulated oscillations in the vicinity of shocks of solutions for the corresponding dispersionless equations for the same initial data. The present thesis provides a first step to study this problem analytically using the inverse scattering transform method. Both the direct and inverse scattering transform for DSII can be expressed as D-bar equations. We consider the direct spectral transform for the defocusing Davey-Stewartson II equation for smooth initial data in the semi-classical limit. The direct spectral transform involves a singularly perturbed elliptic Dirac system in two dimensions. We introduce a WKB-type method for this problem and prove that it is well defined for sufficiently large modulus of the spectral parameter k ∈ ℂ by controlling the solution of an associated nonlinear eikonal problem. Further, we give numerical evidence that the method is accurate for such k in the semiclassical limit. Producing this evidence requires both the numerical solution of the singularly perturbed Dirac system and the numerical solution of the eikonal problem. We present a new method for the numerical solution of the eikonal problem valid for sufficiently large |k|. For a particular potential we are able to solve the eikonal problem in a closed form for all k, acalculation that yields some insight into the failure of the WKB method for smaller values of |k|. The numerical calculations of the direct spectral transform indicate how to study the singularly perturbed Dirac system for values of |k| so small that there is no global solution of the eikonal problem. We provide a rigorous semiclassical analysis of the solution for real radial potentials at k=0, which yields an asymptotic formula for the reflection coefficient at k = 0 and suggests an annular structure for the solution that may be exploited when |k| ≠ 0 is small. The numerics also suggest that for some potentials the reflection coefficient converges point-wise as ε ↓ 0 to a limiting function that is supported in the domain of k-values on which the eikonal problem does not have a global solution. We suggest that singularities of the eikonal function play a role similar to that of turning points in the one-dimensional theory.