In this article we describe a novel method for regularized regression and apply it to the prediction of a behavioural variable from brain activation images. In the context of neuroimaging, regression or classification techniques are often plagued with the curse of dimensionality, due to the extremely high number of voxels and the limited number of activation maps. A commonly-used solution is the regularization of the weights used in the parametric prediction function. It entails the difficult issue of introducing an adapted amount of regularization in the model; this question can be addressed in a Bayesian framework, but model specification needs a careful design to balance adaptiveness and sparsity. Thus, we introduce an adaptive multi-class regularization to deal with this cluster-based structure of the data. Based on a hierarchical model and estimated in a Variational Bayes framework, our algorithm is robust to overfit and more adaptive than other regularization methods. Results on simulated data and preliminary results on real data show the accuracy of the method in the context of brain activation images.