Cerebral structure segmentation from 3D MRI data is an important task for several medical applications. Brain segmentation methods can focus on specific structures such as the cortical surface, or intend to detect the principal parts of the brain. Independently of their final purpose, they are primarily based on the classification of the intracranial volume into classes corresponding to the main cerebral tissues: cerebrospinal uid (CSF), grey matter (GM), and white matter (WM). These classes present complex geometrical properties. However, they can be discriminated thanks to their distinct signal in modalities such as T1 or T2 MRI; moreover, the cerebral tissues present invariant and specific topological properties. Based on these assumptions, some topology-driven brain tissue classi cation techniques have been proposed. The method described in this short paper belongs to the same family of techniques, since its purpose is the classification of the brain into four classes: sulcal CSF, GM, WM, and ventricular CSF. These classes are modelled (with some simplifying hypotheses) as hierarchically included spheres. Starting from a pre-segmentation based on this model, the four classes then evolve under photometric constraints. This process can be formalised as a discrete multi-class deformable model.