Positron emission tomography (PET) measures physiological and chemical processes in the brain, but its limited spatial resolution and reader-dependent interpretation complicate quantification. Brain atlases supply the anatomical and functional scaffolding needed for cross-subject normalisation, region-of-interest definition, and image segmentation. Here we survey atlases widely used in PET, from classical anatomical references (Brodmann, MNI, AAL) to functional and metabolic atlases (FDG normative cohorts, receptor maps, connectivity parcellations), and review the registration pipeline and partial volume effects, which can underestimate tracer uptake by up to 30% in small structures.
Segmentation approaches are compared and linked to clinical applications in neurology (neurodegenerative diseases, epilepsy, tumours) and psychiatry. We highlight a shift toward deep learning-based segmentation approaches and the continued necessity of partial volume correction, underscoring challenges such as atlas mismatch in pathological brains and the lack of standardized benchmarks.