The brain metabolic pattern of vascular dementia (VaD) remains poorly characterized. Univariate voxel-based analysis ignores the functional correlations among structures and may lack sensitivity and specificity. Here, we applied a novel voxel-based multivariate technique to a large ((18)F)2-fluoro-2-deoxy-D-glucose positron emission tomography data set. The sample consisted of 153 subjects, one-third each being probable subcortical VaD, probable Alzheimer disease (AD) (matched for Mini-Mental-State examination (MMSE) and age), and normal controls (NCs). We first applied principal component (PC) analysis and removed PCs significantly correlated to age. The remainders were used as feature vectors in a canonical variate analysis to generate canonical variates (CVs), that is, linear combinations of PC-scores. The first two CVs efficiently separated the groups. CV(1) separated VaD from AD with 100% accuracy, whereas CV(2) separated NC from demented subjects with 72% sensitivity and 96% specificity. Images depicting CV(1) and CV(2) showed that lower metabolism differentiating VaD from AD mainly concerned the deep gray nuclei, cerebellum, primary cortices, middle temporal gyrus, and anterior cingulate gyrus, whereas lower metabolism in AD versus VaD concerned mainly the hippocampal region and orbitofrontal, posterior cingulate, and posterior parietal cortices. The hypometabolic pattern common to VaD and AD mainly concerned the posterior parietal, precuneus, posterior cingulate, prefrontal, and anterior hippocampal regions, and linearly correlated with the MMSE. This study shows the potential of voxel-based multivariate methods to highlight independent functional networks in dementing diseases. By maximizing the separation between groups, this method extracted a metabolic pattern that efficiently differentiated VaD and AD.