Individualized parcellation reveals functional boundaries in human prefrontal cortex

Individualized parcellation reveals hidden functional boundaries in the human prefrontal cortex

The human prefrontal cortex (PFC)—the brain's supposed command center—is organized much more specifically than traditional maps suggest. While standard brain atlases show smooth, continuous transitions of function across the frontal lobes, looking at individual brains reveals sharp, discrete boundaries between different functional zones. This suggests that our cognitive control centers are not uniform expanses. Instead, they are a complex mosaic of specialized, idiosyncratic parts.

The failure of the group average

For decades, neuroscientists have debated the fundamental architecture of the PFC. One school of thought suggests the region is organized into large-scale, continuous gradients. These are smooth transitions where function shifts gradually, such as a rostro-caudal axis (an axis running from the front to the back of the brain) moving from abstract to concrete processing. Another school argues for regional specialization. They suggest discrete "patches" of cortex respond selectively to specific stimuli, implying sharp functional boundaries.

The problem is that current state-of-the-art tools struggle to resolve this tension. Most functional brain atlases, such as the widely used Glasser atlas, are derived from group-averaged data. By averaging the activity of many people together, these maps essentially "smooth out" the unique features of any single brain. If two people have similar functional peaks located just a few millimeters apart, the group average shows a single, blurry peak in the middle. As the authors demonstrate, this process obscures the very thing we are looking for: the sharp edges that define specialized functional units. Consequently, group-averaged maps may only ever show the "gradient" because the "boundaries" have been washed away by the math of averaging.

Decoding the idiosyncratic brain

To move beyond the blurred consensus of the group, the researchers implemented a hierarchical Bayesian parcellation (HBP) framework. This method combines a group atlas with individual task-evoked fMRI data. Instead of assuming everyone shares the same functional map, this approach treats the individual's brain as a unique entity. It uses the group-level atlas as a biological "prior" (a starting assumption used to guide statistical estimation).

The mechanism functions in three primary stages: 1. Variance Decomposition: The authors first decompose task-evoked neural activity into three orthogonal (independent) components. These include a group signal ($g$) consistent across subjects, a subject-specific signal ($s$) unique to the individual, and noise ($\epsilon$). This allows them to quantify how much of the brain's activity is "shared" versus "idiosyncratic." 2. Bayesian Fusion: Using an Expectation-Maximization (EM) algorithm, the HBP framework combines the group atlas with the individual's specific task-evoked fMRI data. This allows the boundaries of functional parcels (defined regions of the brain) to shift and adapt to the specific geometry of a single person's response profiles. 3. Boundary Detection: To prove these shifted boundaries were real, the authors employed the Distance-Controlled Boundary Coefficient (DCBC). This metric compares the similarity of neural activity between voxels (the smallest controllable elements of a 3D volume in imaging) inside a parcel versus those across a boundary. Crucially, it strictly controls for the physical distance between them. This ensures a detected "edge" is not just a byproduct of nearby voxels naturally looking more alike.

Evidence for a mosaic architecture

The results of this individualized approach are stark. The researchers found that the PFC is far more idiosyncratic than previously realized. The authors report that approximately 67.2% of the explainable variance in the PFC is subject-specific [Figure 1b]. This means more than two-thirds of the reliable functional signal in your prefrontal cortex is unique to you.

When the authors applied the individualized parcellation to the MDTB (Multi-Domain Task Battery) dataset, the difference in boundary detection was massive. While the Glasser group atlas showed almost no evidence of functional boundaries (a DCBC near zero), the individualized atlases revealed significant, sharp transitions [Figure 2c,e]. The individualized maps successfully uncovered a "mosaic" of regions. These regions are stable across different types of cognitive tasks, suggesting these boundaries are a fundamental feature of individual brain architecture.

Furthermore, the PFC was found to be more "fine-grained" than other association areas. By measuring the spatial autocorrelation—how quickly neural activity patterns change as you move across the cortex—the authors found that the PFC's functional profiles drop off much more rapidly than those in the parietal cortex [Figure 3a,b]. This implies a denser packing of distinct computational units in the frontal lobes. This density provides the high-resolution toolkit required for complex cognitive control.

The cost of smoothing reality

While these findings are transformative, they come with important caveats. First, the study is limited by the inherent spatial resolution of fMRI technology. This resolution is typically around 2 mm. The authors note that even finer-grained organizational structures might exist at scales smaller than current scanners can resolve.

Second, the research highlights a significant risk in how we interpret "domain-general" systems. One major construct in neuroscience is the "multiple-demand" (MD) system. This refers to regions thought to be recruited for almost any difficult task. However, the authors show that group-averaging significantly overestimates how much these tasks actually overlap in a single person. When looking at individuals, the correlation between different executive tasks was substantially lower than group maps suggested .

Specifically, the correlation between task-switching and stop-signal tasks was 25.1% lower at the individual level than at the group level. This means what looks like a single, unified "command center" in a textbook may actually be a collection of distinct, interdigitated sub-regions in a real human brain.

Verdict: Move toward precision mapping

The verdict is clear: the era of relying solely on group-averaged brain maps for understanding high-level cognition is reaching its limit. If we want to understand how the human mind truly organizes its most complex functions, we must embrace individualized parcellation.

The study proves that the PFC is not a smooth gradient, nor a static set of group-defined blocks. Instead, it is an individualized mosaic of fine-grained subdivisions embedded within broader gradients. For researchers, this means a necessary re-evaluation of "universal" cognitive networks like the multiple-demand system. For clinicians, it serves as a warning. An assessment based on a group atlas may completely miss the unique functional boundaries that define a patient's cognitive profile. Code for the hierarchical Bayesian framework is reportedly available; see the paper for the canonical link.

Figures from the paper