What Is the Desikan-Killiany Atlas & Why Is It Important?

The Desikan-Killiany atlas is a standardized map that divides the outer surface of the human brain into 34 labeled regions per hemisphere, giving researchers and clinicians a shared vocabulary for identifying and measuring specific parts of the cortex. Published in 2006 by Rahul Desikan, Florent Ségonne, Bruce Fischl, and colleagues, the atlas was built from manual tracings on 40 MRI scans and then encoded into software so that any new brain scan could be automatically parcellated into the same set of regions. Its importance stems from something deceptively simple: when thousands of studies across dozens of countries refer to “the superior frontal gyrus” or “the entorhinal cortex,” the Desikan-Killiany atlas is often the system ensuring they all mean the same chunk of tissue.

How the Atlas Was Built

The atlas emerged from a painstaking process at the Martinos Center for Biomedical Imaging at Massachusetts General Hospital. Trained neuroanatomists manually outlined 34 cortical regions of interest on each hemisphere of 40 MRI scans, using visible landmarks on the brain’s folds and grooves (gyri and sulci) to draw boundaries. That manual labeling was then translated into an automated system: software learned the spatial patterns from those 40 hand-labeled brains and could apply the same labels to any new scan without a human tracer touching it.¹ The regions are defined by surface anatomy rather than by the microscopic cell structure underneath, which makes them identifiable on a standard clinical MRI without specialized scanning protocols.

This design choice has practical consequences. Because the boundaries follow gyri and sulci visible on any decent-quality structural MRI, the atlas is usable across different scanner manufacturers, field strengths, and imaging centers. A research team in Seoul and a clinical group in Boston can both feed their scans through the same pipeline and get measurements for the same 68 cortical labels (34 per side). That consistency is what made the atlas a lingua franca for brain morphometry.

FreeSurfer and the Spread of the Atlas

The Desikan-Killiany atlas would not have become as dominant as it has without FreeSurfer, the free, open-source neuroimaging software suite developed at Harvard and MIT. FreeSurfer adopted the atlas as its default cortical parcellation, meaning that every time a researcher runs the standard FreeSurfer processing pipeline on a brain MRI, the output includes cortical thickness, surface area, and volume measurements for each of the 68 Desikan-Killiany regions. Because FreeSurfer is by far the most widely used tool for cortical morphometry, the atlas effectively ships with the software.

Normative reference data have been built specifically around this pairing. Researchers have published population norms for FreeSurfer-derived morphometric estimates using the Desikan-Killiany parcellation, accounting for factors like age, sex, and scanner characteristics.² Those norms let clinicians and researchers compare an individual brain to a reference population, region by region. If someone’s entorhinal cortex is thinner than expected for their age, that stands out as a quantifiable deviation rather than a vague impression.

Where the Atlas Shows Up in Research

The atlas’s footprint in neuroscience research is enormous, and a few areas stand out for how heavily they depend on it.

In Alzheimer’s disease and dementia research, the Desikan-Killiany atlas is a workhorse. Large-scale studies comparing cortical thinning and volume loss across disease stages routinely use FreeSurfer’s native Desikan-Killiany parcellation to generate their measurements.³ The atlas makes it straightforward to ask questions like “which regions thin the fastest in early Alzheimer’s?” and compare answers across research groups. Regions like the entorhinal cortex and the inferior temporal gyrus consistently show up as early markers of neurodegeneration, and the atlas provides the standardized labels that make those findings comparable.

In psychiatry, massive multi-site consortia like ENIGMA (Enhancing Neuro Imaging Genetics through Meta-Analysis) have processed tens of thousands of brain scans through the same FreeSurfer and Desikan-Killiany pipeline. One ENIGMA study on major depressive disorder, pooling data from multiple sites, processed all T1-weighted MRI scans with FreeSurfer and the Desikan-Killiany atlas to examine brain structural asymmetry.⁴ Another study used the atlas’s 68 regions to run a principal component analysis on cortical thickness from over 41,000 neurotypical participants and then compared the resulting pattern to thickness differences associated with numerous psychiatric and neurological disorders across roughly 36,000 additional cases and controls.⁵ These are the kinds of studies that only become possible when everyone uses the same map.

The atlas has also been central to genetics research on brain structure. Studies quantifying how genes influence the shape of the cortex have used the Desikan-Killiany parcellation to divide the brain into regions and then estimate the heritability and genetic overlap of cortical thickness and surface area across those regions.⁶ The shared parcellation scheme lets different genetic studies compare their maps of which regions are most heritable, building a cumulative picture that no single study could provide.

How It Compares to Other Brain Atlases

The Desikan-Killiany atlas is far from the only way to carve up the cortex, and understanding its alternatives helps clarify what it does well and where it falls short.

The most common comparison is with the Destrieux atlas, which is also available in FreeSurfer. Where the Desikan-Killiany atlas divides each hemisphere into 34 regions, the Destrieux atlas splits each hemisphere into 74 regions by distinguishing between sulcal and gyral subregions.⁷ The finer parcellation provides more spatial detail, but that detail comes with a cost. Because the Destrieux atlas uses more labels and a more complicated classification system, maintaining accuracy and consistency in the automated segmentation is harder. When researchers examined scan-rescan reproducibility, the Destrieux atlas showed significantly more variability than the Desikan-Killiany atlas.⁸ For many research purposes, the Desikan-Killiany atlas hits a practical sweet spot: enough spatial detail to be useful, not so much that the automated labeling becomes unreliable.

At the other end of the spectrum sits the Glasser multimodal parcellation, which divides each hemisphere into 180 regions based not just on anatomy but also on functional MRI signals, cortical myelin maps, and connectivity patterns. The Glasser atlas captures much finer functional boundaries, but its regions do not map neatly onto Desikan-Killiany regions. A study comparing the two found that while certain Desikan-Killiany regions have high correspondence with specific Glasser regions, the relationship is often asymmetric. For example, nearly all of the Desikan-Killiany pericalcarine region falls within the Glasser atlas’s primary visual cortex (V1), but V1 in the Glasser atlas spans four different Desikan-Killiany regions.⁹ This means findings from one atlas cannot simply be translated to the other by renaming labels. Researchers who need to compare results across studies using different atlases have to account for these spatial mismatches.

The DKT Revision

The original Desikan-Killiany atlas had known weaknesses, particularly regions where the anatomical boundaries were ambiguous on a standard MRI. In response, Arno Klein and Jason Tourville developed a revised version called the Desikan-Killiany-Tourville (DKT) atlas. The DKT atlas removed some regions with unclear boundaries and revised many existing boundaries to follow more anatomically consistent landmarks in the depths of the brain’s sulci.¹⁰ The changes were designed to reduce ambiguity in manual labeling and improve the reliability of automated parcellation.

The DKT atlas has gradually been adopted alongside the original, and some major projects now use it as their standard. For instance, the UK Biobank’s neuroimaging-derived structural phenotypes include cortical thickness values averaged over DKT regions.¹¹ In practice, the DKT atlas retains the overall spirit and most of the regional labels of the original while cleaning up the edges. Many researchers refer to both versions somewhat interchangeably, which can cause confusion when comparing studies. If precise boundaries matter for your analysis, it is worth checking which version a given study actually used.

Reliability and the Software Version Problem

One of the reasons the atlas became so popular is that its measurements are reasonably reproducible. When the same person is scanned twice and the images are run through FreeSurfer’s pipeline, the resulting Desikan-Killiany measurements tend to agree well. In a study examining test-retest reliability, cortical thickness, surface area, and volume from the atlas showed intraclass correlation coefficients around 0.81 to 0.88 after a visual quality-approval step.¹² Surface area and volume tended to be slightly more reliable than cortical thickness, and reliability dropped somewhat when comparing scans taken at different sites rather than the same scanner.

A subtler but important issue is that the version of FreeSurfer used to process the scans matters more than many researchers realize. A multi-site study comparing FreeSurfer versions 5.3, 6.0, and 7.1 found that cortical thickness from version 7.1 was notably less compatible with older versions, especially along the cingulate gyrus, where version compatibility dropped to intraclass correlations as low as 0.37.¹³ Surface area in regions like the temporal pole and frontal pole also showed low to moderate version compatibility. This means that a study processed with FreeSurfer 5.3 and a study processed with FreeSurfer 7.1 might produce meaningfully different thickness values for the same brain regions, even using the same atlas labels. For meta-analyses or longitudinal studies where scans are processed years apart, this version drift can introduce noise or bias that has nothing to do with actual brain changes.

The practical takeaway is that researchers need to report which FreeSurfer version they used and, ideally, reprocess all scans with the same version when combining datasets. The atlas itself is stable, but the software that applies it has changed in ways that affect the numbers coming out the other end.

Connectomics and Network-Level Uses

Beyond measuring thickness and volume in individual regions, the Desikan-Killiany atlas serves as a scaffold for mapping brain connectivity. In connectomics research, the atlas’s 68 cortical regions act as nodes, and researchers use diffusion MRI tractography to estimate the white-matter fiber bundles connecting them. The result is a structural “connectome,” a network diagram showing how strongly different brain regions are physically linked.

The atlas’s moderate number of regions makes it a practical starting point for this work. Researchers investigating how many tractography streamlines are needed for reliable connectome construction have used the Desikan-Killiany atlas to create parcellations at various resolutions, testing within-individual variability across different analysis choices.¹⁴ The atlas’s coarser grain compared to something like the Glasser parcellation is actually an advantage here, because fewer nodes means more streamlines per connection, which reduces noise in the connectivity estimates. For studies that need finer-grained networks, the atlas can be subdivided further, but the 68-region version remains a common baseline.

What the Atlas Cannot Tell You

The Desikan-Killiany atlas maps the brain’s visible surface anatomy, not its functional organization or microstructural architecture. A given atlas region might contain tissue with very different cellular compositions, connectivity profiles, and functional roles. The atlas treats the “superior frontal gyrus” as one region, but neuroscientists know that different parts of that gyrus do different things. For questions that demand functional precision, a gyral atlas is inherently limited.

This matters in clinical contexts. A probabilistic atlas based on direct electrocortical stimulation during neurosurgery found that the functional geography of the brain is far more variable across individuals than a fixed atlas would suggest. Language disruption, for instance, did not cluster neatly in the classic textbook locations. Instead, the likelihood of disrupting language followed graded probabilistic gradients with high variability between patients, and the middle frontal gyrus emerged as a consistent site for naming and speech arrest rather than the traditionally emphasized inferior frontal gyrus alone.¹⁵ Motor responses extended into parietal association cortex, and higher-order experiences like forced thoughts were reproducibly evoked from frontal and temporoparietal sites. None of this variability is captured by a fixed anatomical parcellation like the Desikan-Killiany atlas.

For neurosurgical planning, epilepsy localization, or any scenario where the question is “what does this specific piece of brain do in this specific patient,” the atlas is a starting framework at best. Its value lies in population-level comparisons, not in individual functional mapping.

Studying the Brain Across the Lifespan

One of the atlas’s more quietly important contributions is enabling researchers to track how the brain changes from childhood through old age. Because the same 68 labels can be applied to any brain regardless of the person’s age, the atlas provides a consistent measurement framework for lifespan studies. Research using FreeSurfer and the Desikan-Killiany atlas on children and adolescents has found that cortical thickness tends to decrease linearly with age within most regions, while surface area and cortical volume follow curved trajectories, reaching their peak at different developmental stages. On average, surface area peaks later in development than volume does.¹⁶

These trajectories matter because they establish what “normal” looks like at each age, which in turn helps researchers identify when development goes off track in conditions like autism, ADHD, or schizophrenia. Without a standardized parcellation, every study would have to define its own regions, and comparing developmental findings across studies would be nearly impossible. The atlas doesn’t solve the harder problem of understanding why the cortex thins or how surface area expansion relates to cognitive maturation, but it gives everyone the same ruler.

Why a 2006 Atlas Still Dominates

It may seem strange that an atlas built from just 40 brains nearly two decades ago remains the default in a field that now routinely works with tens of thousands of scans. The answer has less to do with the atlas being optimal and more to do with network effects. Tens of thousands of published studies have reported results using Desikan-Killiany labels. The normative databases, the ENIGMA consortium protocols, the UK Biobank pipelines, the brain-age prediction models, and countless disease-specific analyses all use these labels. Switching to a different parcellation would break comparability with that entire body of literature.

There is also an underappreciated value in simplicity. The atlas uses 68 labels based on surface anatomy that any trained neuroanatomist can identify. Finer atlases require multimodal data, specialized acquisition protocols, or assumptions about functional organization that may not hold across populations. The Desikan-Killiany atlas works on a basic structural MRI, which is available everywhere from research centers to rural hospitals. For large-scale epidemiological studies or clinical trials where thousands of scans need to be processed efficiently, that simplicity is a feature, not a limitation. The atlas is not the most detailed map of the brain anyone has drawn, but it is the one most people can agree on and actually use, which is why it remains the starting point for so much of modern neuroimaging.