Intra-Atlas Node Size Effects on Graph Metrics in fMRI Data: Implications for Alzheimer's Disease and Cognitive Impairment

Sahithi Kolla, Haleh Falakshahi, Anees Abrol, Zening Fu, Vince D. Calhoun

TL;DR

This paper investigates how the size of brain regions (nodes) in functional brain networks affects the calculated graph metrics. They introduce "Node-Metric Coupling" (NMC) and find it's a significant factor, particularly in distinguishing between healthy controls, individuals with mild cognitive impairment, and those with Alzheimer's disease, suggesting NMC could be a valuable biomarker.

What Problem Does This Paper Address?

Understanding the brain's complex network of interconnected regions is a major goal in neuroscience. Researchers often use graph theory to model these networks, representing brain regions as "nodes" and their connections as "edges." However, a key challenge is how to define these nodes. Some approaches use individual voxels, while others use larger, predefined anatomical regions (atlases). A significant issue with atlas-based methods is that the size of these predefined regions can vary dramatically, both between different atlases and even within the same atlas. This variability in node size has been largely overlooked, yet it could significantly influence the graph metrics calculated from fMRI data, potentially leading to misinterpretations or obscuring true biological differences. This paper aims to quantify this influence and explore its implications, especially in the context of neurodegenerative diseases like Alzheimer's.

Methods

The researchers employed a data-driven approach using the Neuromark framework. They started with resting-state fMRI data and used automated Independent Component Analysis (ICA) to identify intrinsic connectivity networks (ICNs), which serve as their functional nodes. To ensure reproducibility and overcome challenges with standard ICA, they utilized a spatially constrained ICA guided by established network templates.

Functional network connectivity (FNC) matrices were computed by calculating cross-correlations between the time courses of these ICNs. To quantify node size, they counted the number of voxels within each component map that exceeded a specific threshold (3.5 times the standard deviation). This count is proportional to the volume of highly participating voxels.

They then calculated various graph metrics:

  • Local Metrics: Node degree (number of connections), participation coefficient (distribution of connections across communities), and closeness centrality (how close a node is to all others).
  • Global Metrics: Global efficiency (how efficiently information is exchanged across the network), characteristic path length (average shortest path between all node pairs), and clustering coefficient (how interconnected nodes are within their local neighborhoods).

The core of their analysis involved calculating "Node-Metric Coupling" (NMC) by computing the Pearson correlation between node size and each graph metric. They also investigated whether NMC varied across different groups (healthy controls, MCI, AD) and performed statistical tests to assess significance.

Datasets

The study utilized data from the Alzheimer's Disease Neuroimaging Initiative (ADNI). This dataset includes fMRI data from participants in ADNI-1, ADNI-GO, ADNI-2, and ADNI-3. The specific cohort used in this study comprised 213 scans from individuals diagnosed with Alzheimer's disease, 368 with mild cognitive impairment (MCI), and 947 healthy controls (HC). The data underwent standard preprocessing steps including motion correction, slice timing correction, normalization to MNI space, and smoothing.

Key Insights and Ideas

  • Node size matters, a lot: The study strongly demonstrates that the size of brain regions (nodes) significantly influences calculated graph metrics. This isn't a minor effect; it's a robust relationship.
  • Introducing Node-Metric Coupling (NMC): The paper formalizes the concept of NMC, quantifying the correlation between node size and graph metrics. This provides a novel way to assess how node characteristics impact network analysis.
  • NMC as a potential biomarker: The study found significant differences in NMC between healthy controls, MCI, and AD groups. This suggests NMC could be a valuable tool for identifying and understanding neurodegenerative conditions.
  • Disease-specific patterns: The Alzheimer's group showed distinct patterns of NMC, often exhibiting stronger correlations with graph metrics compared to MCI and HC groups. This highlights how disease processes might interact with structural node properties.
  • Metric-dependent relationships: The relationship between node size and graph metrics isn't uniform across all metrics. For example, characteristic path length and global efficiency showed different patterns of correlation with node size in the AD group compared to others.

Key Results

  • Significant correlations were consistently observed between node size and various graph metrics across different brain networks and groups.
  • The AD group showed a particularly robust correlation between node size and global graph metrics.
  • Specific metrics like closeness centrality and characteristic path length showed pronounced differences in NMC across the AD, MCI, and HC groups, suggesting their sensitivity to node size variations in disease states.
  • For instance, in the control and MCI groups, larger node size generally correlated with a greater characteristic path length, while in the AD group, larger node size correlated with a lower characteristic path length.
  • Global efficiency showed predominantly negative correlations with node size across all groups, indicating larger nodes are associated with lower global efficiency.

Limitations and Future Work

  • Limited scope of metrics: The study focused on a subset of graph metrics. Future work could explore a wider range of metrics to capture more nuanced aspects of node roles.
  • Single parcellation approach: The analysis used a single data-driven parcellation. Future research should investigate how different atlas-based approaches with varying node-size ratios affect NMC.
  • Dataset imbalance: The ADNI dataset had an imbalance between Alzheimer's and control subjects, which could influence correlation results.
  • Generalizability: The findings are based on the ADNI dataset and specific parameters. Future studies should aim to replicate these findings in diverse datasets and populations.
  • Thresholding impact: The choice of threshold for defining node size could influence results. Further investigation into different thresholding strategies is warranted.
  • Clinical subtypes: Exploring the predictive value of NMC for different subtypes of MCI (e.g., early, late, stable, progressive) is a promising avenue.
  • Voxel-level and uniformly sampled graphs: Future work could compare atlas-based NMC with voxel-level graphs or uniformly sampled graphs to further understand these relationships.

Summary updated

Abstract

Network neuroscience delves into comprehending the intricate arrangement and interactions within the neural networks of the brain. By merging insights from both neuroscience and network theory, it investigates the brain's structure and function through the lens of interconnected nodes (symbolizing brain regions) and edges (symbolizing the connections among these regions) forming a dynamic network. Graph metrics play a pivotal role in neuroscience by providing quantifiable insights into the intricate connectivity patterns within the brain's neural networks. Frequently, nodes are extracted from functional or structural atlases, which can lead to diverse shapes and sizes. Nevertheless, our understanding of how these differing node characteristics and definitions impact the computed graph metrics in neuroimaging data remains limited. This study adopts a data-driven methodology to delineate functional nodes, subsequently examining the influence of their sizes on resultant graph metrics. By employing the Neuromark framework, an entirely automated independent component analysis (ICA) is applied to resting state fMRI data. Through this, functional network connectivity (FNC) matrices are computed—capturing Pearson correlations between component-time courses—while employing a proportional threshold to streamline node connections based on these correlations. Subsequently, various global and local graph metrics are computed, offering insights into network characteristics. Global metrics provide an overarching summary of the network structure, while local measures delve into structural intricacies at the individual component level. Node sizes are computed based on voxel counts surpassing a designated threshold. Next, we calculated the Pearson correlation between the obtained node sizes and the graph metrics, which we term 'node-metric coupling' (NMC). Our findings revealed consistent and noteworthy correlations between the values of graph metrics and the dimensions of brain nodes. To delve deeper into the implications of this correlation, we examined the node-metric coupling within a dataset comprising Alzheimer's disease, mild cognitive impairment, and control subjects. The disparities observed in node-metric coupling among these groups underscore the necessity of accounting for this factor during analysis. The two principal outcomes of this study are as follows: Firstly, the substantial interplay between varying node sizes within a given atlas and the resultant graph metrics; and secondly, the potential utility of node-metric coupling as a viable biomarker for brain disorders. These discoveries hold significant ramifications. While comparing studies employing diverse atlases is already a challenge, our work highlights an additional, critical source of variability: accounting for node sizes. Consequently, the association between intra-atlas node size and graph metric should be thoughtfully acknowledged in future neuroimaging investigations.
BibTeX
@misc{kolla2024intra,
  author = {Sahithi Kolla and Haleh Falakshahi and Anees Abrol and Zening Fu and Vince D. Calhoun},
  title = {{Intra-Atlas Node Size Effects on Graph Metrics in fMRI Data: Implications for Alzheimer's Disease and Cognitive Impairment}},
  year = {2024},
  doi = {10.3390/s24030814},
}
Figure