A novel method for classification of Magnetic Resonance brain images is presented in this paper. We construct a computational framework for discriminative image feature subspaces. Magnetic Resonance Images of patients in Alzheimer's Disease and normal brain MR images are classified with Support Vector Machines. The framework for the novel method bases on the extraction of gabor features from 2D-Magnetic Resonance images in different scales and orientations. Experiments show that Gabor wavelets can significantly improve classification performance with respect to other popular approaches reported recently in the literature. Combination of gabor features in 3 scales and 8 orientations give 100% classification performance. Experimental results with promising improvements and comparison to related studies are provided.