To study the human brain as a network, neuroscientists rely on a powerful mathematical abstraction: the graph. By representing distinct brain regions as nodes and the structural or functional pathways between them as edges, researchers can apply graph theory to quantify the brain's efficiency, modularity, and health.
However, unlike a computer network with discrete routers and physical cables, the human brain does not come pre-labeled with neat, unambiguous nodes. Instead, researchers must impose a spatial map—a parcellation scheme or brain atlas—to divide the continuous cerebral cortex and subcortical structures into discrete regions of interest (ROIs).
Because there is no universal standard for drawing these boundaries, the choice of atlas introduces a profound systemic vulnerability known as parcellation sensitivity. The boundaries we draw dictate the network metrics we observe, meaning that the fundamental topological properties of a brain network are often artifacts of the chosen map rather than the underlying biology.
Why Node Definitions Distort Graph Topology
To understand why network metrics are so sensitive to parcellation, consider how graph-theoretic calculations operate. Metrics like global efficiency, clustering coefficients, and modularity are not absolute physical properties; they are deeply dependent on the number of nodes ($N$) and how signal variance is averaged within those nodes.
As Salehi et al. demonstrated, network size alone introduces a severe mathematical bias. Even when connection density is held perfectly constant, larger networks (those with more granular parcellations) mathematically skew toward higher efficiency and assortativity, but lower modularity. Because topological metrics scale non-linearly with the number of nodes, it is mathematically invalid to directly compare graph properties across studies that utilize different atlas scales.
[Continuous Brain Surface]
│
├─► Coarse Atlas (e.g., 90 nodes) ──► Low Resolution, High Modularity Artifacts
│
└─► Fine Atlas (e.g., 500 nodes) ──► High Resolution, High Efficiency Artifacts
Beyond node count, the physical boundaries of an atlas alter the time-series data extracted from fMRI or MEG scans. When an atlas boundary misaligns with a true functional boundary, signals from distinct functional tissues are averaged together within the same node. This spatial blending introduces artificial correlations and exacerbates spatial leakage, a phenomenon that Woolrich et al. showed can severely compromise the reliability of resting-state connectivity metrics.
The Spatial Overlap Problem and the Associative Divergence
A common assumption in neuroimaging is that if two different atlases contain a network labeled the "Default Mode Network (DMN)," they are measuring the same biological system. In reality, cross-atlas definitions suffer from a profound lack of spatial agreement.
In a systematic evaluation of neuroimaging results, Kong et al. uncovered what they termed a "Tower of Babel" problem in connectivity analysis. Networks sharing identical names across different parcellation schemes frequently exhibit surprisingly low spatial overlap.
Crucially, the authors revealed that this instability is not uniform across the brain:
- Sensory and Motor Networks: Show high spatial consistency and are relatively robust to atlas selection.
- Higher-Order Associative Networks: (such as those governing executive control or emotion) vary dramatically depending on the specific atlas used.
Because complex cognitive systems are more topologically variable across individuals, relying on a single fixed reference atlas can mask inherent spatial ambiguities and lead to overconfident, unreplicable localizations of cognitive functions.
Downstream Impacts on Machine Learning and Biomarkers
Parcellation sensitivity does not just impact theoretical graph metrics; it directly determines the success or failure of clinical diagnostic models. When building machine learning pipelines to identify disorders like Autism Spectrum Disorder (ASD) or Alzheimer's disease, the initial spatial partitioning acts as a critical feature filter.
For instance, Liu et al. demonstrated that changing the anatomical or functional atlas used for feature extraction can swing the classification accuracy of a Graph Attention Network by over ten percent. If the boundaries of a node fail to isolate the precise sub-regions affected by a pathology, the downstream classifier receives diluted, noisy features.
This sensitivity complicates the deployment of advanced deep learning architectures. Modern models, such as the community-aware graph frameworks proposed by Pei et al., are designed to exploit the brain's natural modular structure. However, because these models depend on an initial parcellation to define their graph structures, their diagnostic predictions and interpretability maps remain tethered to the assumptions of the underlying atlas.
Mitigation Strategies: Building Robustness
To prevent parcellation sensitivity from undermining the reproducibility of connectomics, the field has shifted away from single-atlas dependency toward multi-scale and cross-atlas validation frameworks.
Multi-Atlas Validation
Rather than reporting findings from a single arbitrary map, researchers execute their entire analysis pipeline across multiple distinct anatomical or functional atlases. As seen in work by Wee et al. and Yin et al., if a topological abnormality or biomarker remains statistically significant across entirely different spatial partitioning schemes, it provides strong evidence that the finding reflects a true biological signal rather than an artifact of node definition.
Multi-Resolution Scaling
Instead of viewing scale as a nuisance variable, multi-resolution network analysis treats granularity as a continuous dimension. By validating connectivity findings across a spectrum ranging from coarse anatomical structures to high-density node models, authors like Bai et al. and Wu et al. ensure that observed topological properties (such as small-worldness or rich-club organization) are invariant to the spatial resolution of the graph.
Ultimately, understanding parcellation sensitivity means recognizing that a brain network is always a joint product of biology and geometry. By treating the atlas as an active variable rather than a fixed constant, researchers can isolate true neural architectures from the cartographic boundaries used to map them.
Key Foundational Work
Network modelling methods for FMRI (2011) — contributed 3 concept elements to this direction.