Analysis of connectivity between local multi-variate patterns of functional MRI data

Volker FischerUniversity 2011 · DOI

TL;DR

This paper introduces "pattern connectivity," a novel method for analyzing brain connectivity using fMRI data. It moves beyond traditional approaches by analyzing patterns of activity within brain regions, rather than just a single representative time-course, enabling a more nuanced understanding of complex brain interactions.

What Problem Does This Paper Address?

Understanding how different regions of the brain communicate is crucial for neuroscience. Traditional methods in fMRI analysis often represent a brain region's activity with a single time-series, typically derived from Principal Component Analysis (PCA). This approach, while useful, can oversimplify complex brain activity, especially when a region exhibits multiple distinct patterns of activation related to different stimuli or tasks. This limitation hinders the ability to study more intricate connectivity patterns. This paper addresses this by proposing "pattern connectivity," which leverages Independent Component Analysis (ICA) to extract multiple "patterns of interest" (POIs) from each brain region, allowing for a more detailed and nuanced analysis of brain connectivity.

Methods

The core of this paper is the introduction of pattern connectivity. Instead of using the first principal component from PCA to represent a brain region's temporal activity, pattern connectivity utilizes Independent Component Analysis (ICA) to decompose a region's data into multiple independent components. These components are hypothesized to represent distinct "patterns of interest" (POIs) within the region. The temporal activity of these POIs is then used for connectivity analysis, employing established methods like restricted empirical correlation and Dynamic Causal Modeling (DCM).

Key methodological contributions include:

  • Pattern Selection Algorithm: A crucial aspect of pattern connectivity is selecting the most relevant POIs from the ICA decomposition. The paper proposes a fitness function based on the correlation between the POI's time-course and the associated hemodynamic response function (HRF) onset. This function is used to estimate optimal ICA parameters and select the most relevant components.
  • Component Selection Connectivity Baseline (CSCB): To address potential artifacts introduced by component selection, the authors introduce a method to estimate a statistical baseline for connectivity measures. This involves computing connectivity between randomly positioned regions to establish a reference point.
  • Separately Selected Components: To further mitigate selection artifacts, the paper proposes splitting the data into training and testing sets for component selection and connectivity analysis, respectively.

Datasets

The paper utilizes two primary datasets:

  • BMS Paradigm: This experiment involved visual stimuli (blank, motion, static) and participants pressing a button. It was designed to have regions of interest (VA and MA) with only one POI each, allowing for comparison with the standard PCA method.
  • ORIENT Paradigm: This experiment used differently oriented Gabor patches and motor responses. Crucially, one region of interest (VA) was designed to have two POIs, showcasing the advantage of pattern connectivity in analyzing regions with complex activity.
  • Synthetic fMRI Data: To rigorously validate the method, synthetic fMRI data was generated based on a known underlying network structure, allowing for direct comparison of pattern connectivity results with the ground truth.

Key Insights and Ideas

  • Beyond Single Time-Series: The fundamental idea is that brain regions can exhibit multiple distinct patterns of activity (multi-variate patterns or MVPs) that are not captured by a single representative time-series. Pattern connectivity allows us to analyze the interactions between these individual patterns.
  • ICA for Pattern Extraction: ICA is presented as a more "natural" decomposition method than PCA for extracting these distinct patterns due to its reliance on statistical independence rather than just variance maximization.
  • Addressing Component Selection Artifacts: The paper highlights and provides solutions for the critical challenge of selecting the correct components from ICA, ensuring that the identified connectivity is not an artifact of the selection process.
  • Reproducing and Extending Standard Methods: Pattern connectivity, by using established connectivity analysis techniques (correlation, DCM) on extracted patterns, can reproduce results from standard PCA-based methods while also extending their applicability to more complex scenarios.

Key Results

  • BMS Paradigm: Pattern connectivity, when applied to regions with single POIs, yielded comparable results to the standard PCA method, validating its ability to replicate existing findings.
  • ORIENT Paradigm:
    • Real fMRI Data: Restricted empirical correlation results were weak and close to the connectivity baseline, suggesting limitations in this specific analysis for event-related designs with short stimuli.
  • Synthetic Data: Pattern connectivity successfully identified the known connectivity structure, with large connectivity changes observed between patterns, consistent with the simulated data.
  • DCM Analysis: DCM analysis on real fMRI data revealed hypothesized connectivity changes, particularly in the region with two POIs (VA), supporting the effectiveness of pattern connectivity in more complex scenarios. A clear model hierarchy emerged, favoring models that included instruction modulations.
  • ICA vs. PCA: For regions with multiple patterns, ICA generally achieved higher fitness (correlation with HRF onset functions) than PCA, indicating its ability to extract more relevant components.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the real fMRI data from the ORIENT paradigm yielded weak results for empirical correlation, potentially due to the short stimulus presentation times and the HRF's influence.
  • For human data in the ORIENT paradigm, connectivity changes were small and close to the connectivity baseline, suggesting potential issues with component selection when using only half of the data.
  • The interpretation of DCM parameters can be sensitive to the underlying data and model complexity.
  • Future Work:
    • Exploring other ICA decomposition methods or alternative decomposition techniques like Non-negative Matrix Factorization (NMF).
  • Investigating the use of time-courses derived from Support Vector Machines (SVMs) instead of ICA components, which could represent information content about stimuli more directly and potentially allow for overlapping onset functions.
  • Further research into the role of the HRF and its impact on connectivity analysis, especially in event-related designs.
  • Applying pattern connectivity to analyze more complex neuronal networks and explore a wider range of connectivity properties.

Summary updated

Abstract

During the last century, many innovations in medicine, chemistry, and physics allowed a deeper insight into organization and functionality of the brain. One of the newest of these methods is functional magnetic resonance imaging (fMRI), which, compared to other techniques, achieves a high spatial resolution. Aim of this thesis is to introduce, illustrate, and discuss a new way of investigating connections between brain areas using fMRI data. The key idea of this thesis is to replace the state of the art representation of a brain region's temporal behavior by its first PCA component with time-series associated to multi-variate patterns (MVP) of a region. Now each region can be represented by several time-courses, each one associated with the activation of a multi-variate pattern of the area. To identify multi-voxel patterns and their activation courses, independent component analysis (ICA) was used. During my thesis, this concept is referred to as pattern connectivity. Pattern connectivity states not only a generalization of the existing method, but also yields a more intuitive representation of brain regions and allows to map and study more complex behaviors of brain regions. The analysis of two fMRI experiments and synthetic fMRI data illustrates the proposed method. The first experiment validated pattern connectivity by comparing it to the state of the art method. The second experiment in which the behavior of one of the regions can not be represented by a single time course is described. Here, pattern connectivity was able to validate an intuitive hypothesis about underlying connections. In addition, pattern connectivity revealed the used connectivity structure used to generate a synthetic fMRI data set. The major goal of the present work was to show that, for fMRI data, it is possible to study neuronal connectivity between multi-voxel patterns instead of brain regions by providing, validating, and illustrating the new concept of pattern connectivity, and by identifying and solving problems arising with its implementation.
BibTeX
@inproceedings{fischer2011analysis,
  author = {Volker Fischer},
  title = {{Analysis of connectivity between local multi-variate patterns of functional MRI data}},
  booktitle = {University},
  year = {2011},
  doi = {10.5283/epub.20311},
}
Figure