Alzheimer’s disease (AD) is a neurodegenerative disease that current treatments are only symptomatic and restricted to reduce the progression rate of mental deterioration (Hong-Qi et al., 2012). AD may affect people in different ways, but the most common first symptom is the inability to remember new information (Alzheimer’s Association, 2012). The transitional stage in which the patient’s cognitive functions are not considered normal, but the patient does not meet the criteria for the dementia is called Mild Cognitive Impairment (MCI). MCI is associated with high risk for development of AD, with conversion rates between 10% and 15% per year (Petersen et al., 2001). There is no reliable method for early diagnosis of AD. Diffusion Tensor Imaging (DTI) is a highly promising imaging technique, whose development may provide much earlier evidences of the disease than neuropsychological symptoms (Várkuti, 2011).Following the hypothesis that DTI measures can potentially be used in the early diagnosis of dementia, this study aims to discriminate AD, MCI patients and healthy volunteers through a multilevel DTI approach. We propose this approach to evaluate whether fractional anisotropy (voxel level), axonal tractography (connection level) and graph theory (network level) are good discriminators of the disease in study.AD, MCI patients and healthy volunteers (15 from each), matching age and education level, were recruited. T1 and DTI were acquired at Institute D’or (Brazil) on a Philips 3T scanner. T1 was acquired by gradient recalled echo, with repetition time = 7.16msec, echo time = 3.41msec, flip angle = 8, matrix = 480x480 with resolution 0.5mm x 0.5mm, 340 slices of 0.5mm. DTI was acquired using spin echo, with repetition time = 5620msec, echo time = 65msec, flip angle = 90, matrix = 96x96 with resolution 2.5mm x 2.5mm, 60 slices of 2.5mm, 32 directions and b-value=1000sec/mm2. The images were preprocessed by realignment, coregistration and normalization by SPM, segmentation and reconstruction by DSI Studio. Fractional Anisotropy (FA) was used for voxel level analysis. Tractography (Kreher et al., 2008) was performed at DSI Studio, considering the whole brain as seed, with FA=0.1, maximum angle = 60, step size = 1.25mm, length constraint from 25 to 100mm. The connectivity matrix, with each Brodmann area (BA) as node, was used for connection level analysis. Degree, Clustering coefficient, Efficiency, Betweenness centrality and Vulnerability of each connectivity matrix was extracted for network level analysis. Leave-One-Out Cross-Validation was performed, including feature seletion (Fisher Score - FS), training and classification (Support Vector Machine).At voxel level, the two biggest clusters are inside both parahippocampal gyrus (Figure 2) and hippocampal cingulum (Figure 3). At hippocampal cingulum, accuracy reached 93% between AD and controls, and 87% between AD and MCI. At parahippocampal gyrus, accuracy reached 90% between AD and controls, and 90% between AD and MCI. At connection level (BAs as nodes), top 1 edge obtained the highest classification, reaching 83% for AD versus controls and 70% for AD versus MCI. Top 1 discriminating edge was the connection between BA 10 and 32 (Figure 4). At network level, top 4 feature set achieved the highest accuracy, reaching 80% for AD versus controls and 70% for both MCI versus AD. Top 4 features includes BA 34 (entorhinal cortex), BA 20 (inferior temporal gyrus), BA 42 (posterior transversal temporal area) and BA 46 (dorsolateral prefrontal cortex) (Figure 5).Based on the results, we can indicate that a stage of specific FA alterations inside hippocampal cingulum and parahippocampal gyrus is good a biomarker for AD. These findings in parahippocampal gyrus are in accordance to indications from previous studies (Dyrba et al., 2013). Parahippocampal gyrus is involved in visual memory (Epstein and Kanwisher, 1998). Hippocampal cingulum is the hippocampal formation’s portion of cingulum (Nir et al., 2013). These results indicate that our approach has strong potential for clinical diagnosis of AD and MCI.