We present a comparison between a voxel based approach and a region based technique for detecting brain activation signals in sequences of functional Magnetic Resonance Images (fMRI). The region based approach uses an automatic parcellation of the brain that can incorporate anatomical constraints. A standard univariate voxel based detection method (Statistical Parametric Mapping [5]) is used and the results are compared to those obtained when performing detection of signals extracted from the parcels. Results on a fMRI experimental protocol are presented and show a greater sensitivity using the parcelling technique. This result remains true when the data are analyzed at several resolutions.