In this paper a novel framework for brain classiflcation is pro- posed in the context of mental health research. A learning by example method is introduced by combining local measurements with non linear Support Vector Machine. Instead of considering a voxel-by-voxel compar- ison between patients and controls, we focus on landmark points which are characterized by local region descriptors, namely Scale Invariance Feature Transform (SIFT). Then, matching is obtained by introducing the local kernel for which the samples are represented by unordered set of features. Moreover, a new weighting approach is proposed to take into account the discriminative relevance of the detected groups of features. Experiments have been performed including a set of 54 patients with schizophrenia and 54 normal controls on which region of interest (ROI) have been manually traced by experts. Preliminary results on Dorso- lateral PreFrontal Cortex (DLPFC) region are promising since up to 75% of successful classiflcation rate has been obtained with this tech- nique and the performance has improved up to 85% when the subjects have been stratifled by sex.