To accurately predict the development of Alzheimer's disease (AD) at its predementia stage would be a major breakthrough from both therapeutic and research standpoints. In this review, our focus is on markers obtained with structural imaging--especially magnetic resonance imaging (MRI)--and on studies of subjects at risk of developing AD. Among the latter, amnestic mild cognitive impairment (MCI) is currently the most commonly accepted reference, and therefore is specially targeted in this review. MCI refers to patients with significant but isolated memory impairment relative to subjects of identical age. Consistent with established histopathological data, structural imaging studies comparing patients with early probable AD to healthy aged subjects have shown that the most specific and sensitive features of AD at this stage are hippocampal and entorhinal cortex atrophy, especially when combined with a reduced volume of the temporal neocortex. MCI patients have significant hippocampal atrophy when compared to aged normal controls. When comparing patients with probable AD to MCI subjects, hippocampal region atrophy significantly extends to the neighboring temporal association neocortex. However, only longitudinal studies of MCI subjects are suited to assess (in a retrospective way) the predictive value of initial atrophy measurements for progression to AD. Few such studies have been published so far and for the most they were based on small samples. Furthermore, the comparison among studies is clouded by differences in both populations studied and MRI methodology used. Nevertheless, comparing the initial MRI data of at-risk subjects who convert to AD at follow-up to those of nonconverters suggests that a reduced association temporal neocortex volume combined with hippocampal or anterior cingulate cortex atrophy may be the best predictor of progression to AD. These data, although still preliminary, are consistent with postmortem studies describing the hierarchical progression of tau lesions in normal aging and early stages of AD, such that damage to the medial temporal lobe and association cortex would account for the memory and nonmemory cognitive impairments, respectively, the combination of which is required to operationally define probable AD. Future research in this field should capitalize on thorough methodology for brain structure delineation, and combine atrophy measurements to cognitive and/or functional imaging data.