Among the numerous 3D medical image segmentation methods proposed in the literature, very few have intended to provide topologically satisfying results, a fortiori for multiple object segmentation. In this paper, we present a method devoted to parallel segmentation of the main classes of cerebral tissues from 3D magnetic resonance imaging data. This method is based on a multi-class discrete deformable model strategy, starting from a topologically correct model, and guiding its evolution in a topology-preserving fashion. Validations on a commonly used cerebral image database provide promising results and justify the further development of a general methodological framework based on the concepts exposed in this preliminary work.