One of the most important topics in computer vision is pattern recognition and classification in images. Any classification process requires from a feature extraction process and a learning technique that categorizes each data sample. However, sometimes, it is not enough to have just a classification since we could need to introduce high-level knowledge constraints to obtain a meaningful classification. Deformable models are one of the possible tools to achieve this goal. This PhD thesis describes several new techniques to be used in this scenario regarding deformable models and classification theory. The definition of deformable models guided using a external potential derived from a generative model is proposed. This approach is called generative snakes. To illustrate this process parametric snakes in a texture based context are used. The extension of the former work to geodesic deformable models is done
by reformulating the geometric deformation process, leading to the Stop and Go formulation. A new tool for mixing labelled and unlabelled data for semi-supervised and particularization problems is developed and validated. This new technique allows supervised and unsupervised processes to compete for each data sample, defining the supervised clustering competition scheme. These techniques are motivated by and applied to medical image analysis, in particular to Intravascular Ultrasound (IVUS) tissue segmentation and characterization. This work also studies the tissue characterization problem in IVUS images and defines a new framework for automatic plaque recognition.