Introduction
Network neuroscience has increasingly modeled the human brain as a "connectome," where anatomical regions serve as nodes and white matter pathways represent edges. In the context of Alzheimer’s Disease (AD), often characterized as a "disconnection syndrome," graph theoretical analysis offers a mathematical framework to quantify the breakdown of these neural networks. However, the field faces a significant reproducibility challenge: conflicting reports on AD-related topological changes. A primary suspect for this inconsistency is the "parcellation scale"—the resolution at which the brain is divided into discrete regions. This study investigates how varying parcellation scales impact the stability and sensitivity of common graph metrics in identifying AD-related connectivity changes.
Methods
The study utilized Diffusion Tensor Imaging (DTI) data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI-GO/2) cohort, comprising 173 age-controlled subjects: 76 Healthy Controls (HC), 68 with Mild Cognitive Impairment (MCI), and 29 with AD.
Networks were constructed using "fiber density" (FD) weighting, which normalizes streamline counts by the surface area of the connected regions. To assess scaling effects, the researchers applied five hierarchical scales of the Lausanne parcellation, ranging from 83 nodes (Scale 33) to 1,015 nodes (Scale 500). Four key graph metrics were evaluated:
- Segregation (Local): Average Clustering Coefficient and Transitivity.
- Integration (Global): Characteristic Path Length and Global Efficiency.
Statistical differences between groups (HC, MCI, AD) were determined using non-parametric tests with False Discovery Rate (FDR) corrections across all five scales.
Results
The analysis revealed that most graph metrics are highly sensitive to the resolution of the brain map, with one notable exception.
The Robustness of Transitivity
Transitivity emerged as the most consistent biomarker, successfully identifying significant group differences (HC vs. MCI and HC vs. AD) across all five parcellation scales. Unlike the Average Clustering Coefficient, which failed to detect differences at higher resolutions (Scales 125, 250, and 500), transitivity remained scale-invariant.
Sensitivity of Integration Metrics
Measures of global integration proved to be the most fragile. While Global Efficiency identified HC vs. MCI differences across all scales, its ability to distinguish HC from AD dropped to 3 out of 5 scales. Characteristic Path Length was the least reliable, failing to show consistent group differences across the majority of scales. As the number of nodes increased, the mathematical "distance" between them scaled inconsistently, confounding the results.
Network Topology in AD Progression
The findings support the "disconnection" hypothesis, showing increased segregation (transitivity/clustering) in AD patients compared to controls. This suggests a shift from efficient global communication to isolated local processing. Interestingly, MCI patients exhibited a non-linear trend, sometimes showing decreased segregation before the eventual increase seen in the AD stage.
Discussion and Conclusion
The choice of parcellation scale is not merely a technical detail; it fundamentally alters the interpretation of brain network health. The sensitivity of integration metrics like Path Length suggests they may be unreliable for cross-study comparisons using different atlases. Conversely, the robustness of transitivity across resolutions makes it a superior candidate for a stable AD biomarker.
Future research should explore whether these scaling effects persist across different weighting methods (e.g., fractional anisotropy) or functional parcellations. For now, the study highlights the need for standardized resolutions in connectomics to ensure that findings regarding Alzheimer's Disease are both replicable and clinically meaningful.
Summary updated