Hippocampal atrophy, as evidenced using magnetic resonance imaging (MRI), is one of the most validated, easily accessible and widely used biomarkers of Alzheimer's disease (AD). However, its imperfect sensitivity and specificity have highlighted the need to improve the analysis of MRI data. Based on neuropathological data showing a differential vulnerability of hippocampal subfields to AD processes, neuroimaging researchers have tried to capture corresponding morphological changes within the hippocampus. The present review provides an overview of the methodological developments that allow the assessment of hippocampal subfield morphology in vivo, and summarizes the results of studies looking at the effects of AD and normal aging on these structures. Most studies highlighted a focal atrophy of the CA1 subfield in the early (predementia or even preclinical) stages of AD, before atrophy becomes more widespread at the dementia stage, consistent with the pathological literature. Preliminary studies have indicated that looking at this focal atrophy pattern rather than standard whole hippocampus volumetry improves diagnostic accuracy at the Mild Cognitive Impairment (MCI) stage. However, controversies remain regarding changes in hippocampal subfield structure in normal aging and regarding correlations between specific subfield volume and memory abilities, very likely because of the strong methodological variability between studies. Overall, hippocampal subfield analysis has proven to be a promising technique in the study of AD. However, harmonization of segmentation protocols and studies on larger samples are needed to enable accurate comparisons between studies and to confirm the clinical utility of these techniques. Highlights:-Hippocampal subfield structure can be assessed in vivo with (high-resolution) MRI-AD-related atrophy is initially focal (in CA1) before spreading to other subfields-This pattern of atrophy could be a sensitive biomarker for early AD detection-The effect of age and specific memory-volume correlations are less clear-Variations in methods and segmentation protocols cause important discrepancies