Model-based image segmentation is a popular approach for the segmentation of anatomical structures from medical images because it includes prior knowledge about the shape and appearance of structures of interest. This paper focuses on the formulation of a novel appearance prior that can cope with large variability between subjects, for instance due to the presence of pathologies. Instead of relying on Principal Component Analysis such as in Statistical Appearance Models, our approach relies on a multimodal intensity profi le atlas from which a point may be assigned to several profi le modes consisting of a mean pro le and its covariance matrix. These profi le modes are first estimated without any intra-subject registration through a boosted EM classi cation based on spectral clustering. Then, they are projected on a reference mesh whose role is to store the appearance information in a common geometric representation. We show that this prior leads to better performance than the classical monomodal Principal Component Analysis approach while relying on fewer pro file modes.