The Joint Detection-Estimation framework has been proposed in [1-3] as a generalization of regression methods for the analysis of fMRI data. It enables the detection of brain activation elicited by stimuli along an experimental paradigm. Also at the subject level, it makes the analysis of brain dynamics feasible through the estimation of regional Hemodynamic Response Functions. Up to now, the JDE framework has been developed to discriminate activating voxels from non-activating ones. Here, we extend this paradigm to also account for putative deactivations that may appear for instance in pathologies (epilepsy). To this end, for any brain region we introduce spatially adaptive 3-class mixture models and 3D Potts field to embody the spatial correlation over the hidden states of the voxels. The regularization is spatially adaptive and varies across experimental conditions. We finally illustrate the interest of this novel approach on synthetic and real unsmoothed fMRI data.