Modern cognitive experiments in functional Mag- netic Resonance Imaging (fMRI) often aim at understanding the temporal dynamics of the brain response in regions acti- vated by a given stimulus. The study of the variability of the hemodynamic response function (HRF) and its characteristics can provide some answers. In this context, we aim at improving the accuracy of the HRF estimation. To do so, we relied on a Joint-Detection-Estimation (JDE) framework that enables robust detection of brain activity as well as HRF estimation, in a Bayesian setting [2]. So far, the hemodynamic results provided by the JDE formalism have depended on a prior parcellation of the data performed before JDE inference. In this study, we propose a new approach to relax this prior knowledge: using consensus clustering techniques based on random parcellations of the data, we combine hemodynamics results provided by different parcellations, so as to robustify the HRF estimation.