Abstract : Registration is a classical problem in computer vision which is essential in many tasks of medical image analysis. The general principle of a registration algorithm is to optimize a criterion measuring the match between two images among a predetermined set of spatial transformations. Choosing the criterion (the similarity measure) is determinant for both the accuracy and the robustness of the algorithm. There exists currently a catalog of similarity measures in which a developer may select the one which is the most appropriate to a given problem, based on his intuition and experience. In order to make this choice more objective, we propose in this thesis to derive similarity measures from probabilistic image acquisition models. Starting with a simple model of functional dependence between the image intensities, we define a new class of similarity measures that closely relate to the concept of correlation ratio. We demonstrate experimentally that these measures are adapted to several multimodal rigid registration problems involving images acquired by Magnetic Resonance (MRI), Computed Tomography, and Scintigraphy. The correlation ratio method is then extended to the multimodal non-rigid case using a few practical adaptations. We finally formalize image registration as a general maximum likelihood estimation problem, which enables us to take into account image dependence models that are more complex than functional models. This approach is applied to rigid registration of 3D Ultrasound images and MRI.