At present, there is still no consensus on the choice of the reference area in positron emission tomography (PET) studies of Alzheimer's disease (AD). In this study, PET scans with fluorodeoxyglucose-F18 were carried out in the following groups of subjects: 47 patients with probable AD, 8 patients with mild cognitive impairment, and 15 age-similar healthy subjects. Scans normalized to the cerebral global mean (CGM), cerebellum (CBL), and the primary sensorimotor cortex (SMC). We evaluated the effect of the different count normalization procedures on the accuracy of (18)F-FDG PET to detect AD-specific metabolic abnormalities (voxel-based group comparison) and to differentiate between patients and healthy subjects (ROI-based discriminant analysis) with regard to the degree of clinical deterioration. Metabolic reductions in groups of very mildly, mildly and moderate-to-severely affected patients appeared, respectively, 2.2, 2.6, and 2.7 times greater in spatial extent when tracer uptake was normalized to SMC rather than to CGM. The overall accuracy of discrimination was 94%, 91%, and 80% after normalization to SMC, CBL, and CGM, respectively. In general, normalization to SMC was somewhat superior to cerebellar normalization, allowing the detection of more pronounced metabolic deficits and the more accurate discrimination of patients from non-patients. Normalization to CGM should be used with great caution not only in advanced stages of dementia, but also in very mild AD cases.