Ocular pathology is one of the main health problems worldwide. The number of people with retinopathy symptoms has increased considerably in recent years. Early adequate treatment has demonstrated to be effective to avoid the loss of the vision. The analysis of fundus images is a non intrusive option for periodical retinal screening. Different models designed for the analysis of retinal images are based on supervised methods, which require of hand labelled images and processing time as part of the training stage. On the other hand most of the methods have been designed under the basis of specific characteristics of the retinal images (e.g. field of view, resolution). This compromises its performance to a reduce group of retinal image with similar features. For these reasons an unsupervised model for the analysis of retinal image is required, a model that can work without human supervision or interaction. And that is able to perform on retinal images with different characteristics. In this research, we have worked on the development of this type of model. The system locates the eye structures (e.g. optic disc and blood vessels) as first step. Later, these structures are masked out from the retinal image in order to create a clear field to perform the lesion detection. We have selected the Graph Cut technique as a base to design the retinal structures segmentation methods. This selection allows incorporating prior knowledge to constraint the searching for the optimal segmentation. Different link weight assignments were formulated in order to attend the specific needs of the retinal structures (e.g. shape). This research project has put to work together the fields of image processing and ophthalmology to create a novel system that contribute significantly to the state of the art in medical image analysis. This new knowledge provides a new alternative to address the analysis of medical images and opens a new panorama for researchers exploring this research area.

Dans cette thèse, nous nous intéressons à un récent algorithme de segmentation d’images basé sur un processus de régularisation de graphes. L’objectif d’un tel algorithme est de calculer une fonction indicatrice de la segmentation qui satisfait un critère de régularité ainsi qu’un critère d’attache aux données. La particularité de cette approche est de représenter les images à l’aide de graphes de similarité. Ceux-ci permettent d’établir des relations entre des pixels non-adjacents, et ainsi de procéder à un traitement non-local des images. Afin d’en améliorer la précision, nous combinons cet algorithme à une seconde approche non-locale : des caractéristiques de textures. Un nouveau terme d’attache aux données est dans un premier temps développé. Inspiré des travaux de Chan et Vese, celui-ci permet d’évaluer l’homogénéité d’un ensemble de caractéristiques de textures. Dans un second temps, nous déléguons le calcul de l’attache aux données à un classificateur supervisé. Entrainé à reconnaitre certaines classes de textures, ce classificateur permet d’identifier les caractéristiques les plus pertinentes, et ainsi de fournir une modélisation plus aboutie du problème. Cette seconde approche permet par ailleurs une segmentation multiclasse. Ces deux méthodes ont été appliquées à la segmentation d’images texturées 2D et 3D. / In this thesis, we improve a recent image segmentation algorithm based on a graph regularization process. The goal of this method is to compute an indicator function that satisfies a regularity and a fidelity criteria. Its particularity is to represent images with similarity graphs. This data structure allows relations to be established between similar pixels, leading to non-local processing of the data. In order to improve this approach, combine it with another non-local one: the texture features. Two solutions are developped, both based on Haralick features. In the first one, we propose a new fidelity term which is based on the work of Chan and Vese and is able to evaluate the homogeneity of texture features. In the second method, we propose to replace the fidelity criteria by the output of a supervised classifier. Trained to recognize several textures, the classifier is able to produce a better modelization of the problem by identifying the most relevant texture features. This method is also extended to multiclass segmentation problems. Both are applied to 2D and 3D textured images.