Large-scale Probabilistic Functional Modes from resting state fMRI

Samuel J. Harrison, Mark W. Woolrich, Emma C. Robinson, Matthew F. Glasser, Christian F. Beckmann, Mark Jenkinson, Stephen M. SmithNeuroImage 2015 · DOI

TL;DR

This paper introduces a novel Bayesian modeling approach, Probabilistic Functional Modes (PFMs), to identify large-scale brain networks from resting-state fMRI data. Unlike traditional methods that often assume independence, PFMs can capture complex spatio-temporal interactions and individual differences, offering a more nuanced and neuroscientifically plausible representation of brain activity.

What Problem Does This Paper Address?

Resting-state fMRI (rs-fMRI) generates rich data about intrinsic brain activity, but characterizing this activity in a simplified yet interpretable way is a significant challenge. Existing methods like Independent Component Analysis (ICA) often rely on assumptions of independence (spatial or temporal) which may not accurately reflect the brain's interconnected nature. Furthermore, these methods can struggle to robustly capture individual variations in brain networks and their dynamic interactions, leading to potential inaccuracies in subsequent analyses. The paper aims to develop a method that can identify "modes" of coherent activity that are not restricted to being orthogonal or non-overlapping, and can account for subject variability.

Methods

The core of the paper is the introduction of Probabilistic Functional Modes (PFMs), a Bayesian generative model. This model represents the fMRI data as a product of spatial maps and time courses, akin to matrix factorization approaches like PCA and ICA. However, PFMs differ by:

  • Probabilistic Framework: Using a Bayesian approach with variational inference allows for flexible modeling and efficient computation on large datasets.
  • Generative Model: It explicitly models the haemodynamic response function (HRF) in the temporal prior, linking observed BOLD signals to underlying neural activity.
  • Subject Variability Modeling: A hierarchical model explicitly accounts for differences in spatial maps across subjects, going beyond simple group averages.
  • Flexible Spatial Priors: The model uses a delta-Gaussian mixture model for spatial weights, allowing for modes that can overlap and contain anti-correlated regions, reflecting a less restrictive view of brain networks.
  • No Hard Orthogonality Constraints: Unlike some ICA methods, PFMs do not enforce orthogonality between modes, allowing for more complex relationships.

Datasets

  • Simulated Data: The authors generated synthetic rs-fMRI data to rigorously test their method against ground truth and compare it with existing techniques. This simulated data mimicked Human Connectome Project (HCP) acquisition protocols in terms of spatial and temporal resolution, but was reduced in scale for computational manageability.
  • Human Connectome Project (HCP) Data: The developed method was evaluated on high-quality rs-fMRI data from the HCP, comprising data from 209 subjects. This dataset is characterized by high spatial and temporal resolution and extensive preprocessing.

Key Insights and Ideas

  • Brain Networks are Spatially and Temporally Interacting: The paper strongly advocates for the idea that brain networks are not isolated entities but exhibit complex spatio-temporal correlations, a property that traditional ICA methods often struggle to capture due to their independence assumptions.
  • Subject Variability is Crucial and Modelable: Differences between individuals in brain network organization are significant and can be robustly modeled, providing a more personalized understanding of brain function.
  • Bayesian Inference Offers Flexibility and Tractability: The use of a variational Bayesian framework allows for the construction of complex, neuroscientifically plausible models while remaining computationally feasible for large datasets.
  • Test-Retest Reliability as an Indicator of Accuracy: For their method, the authors found a strong correlation between test-retest reliability and ground truth accuracy, suggesting that this metric can be a useful indicator of performance on real data where ground truth is unknown.

Key Results

  • Superior Accuracy on Simulated Data: PFMs outperformed PCA, sICA, tICA, and stICA in recovering ground truth spatial maps, time courses, and correlation structures from simulated data.
  • Robustness to Dimensionality Mismatch: The PFM method showed consistent relative accuracy even when the number of inferred modes did not match the ground truth, suggesting its reliability on real data where the true number of modes is unknown.
  • Stable Convergence on HCP Data: The algorithm demonstrated stable convergence to a consistent set of PFMs on real HCP data, with Bayesian regularization effectively eliminating less reliable components.
  • Identification of Spatio-Temporal Interactions: The method successfully identified complex spatial and temporal correlations between identified brain modes, such as the dorsal attention (DA) and fronto-parietal control (FPC) systems, which were not trivially predictable from each other.
  • Modeling Subject-Specific Network Variations: The model successfully captured subject-specific variations in the spatial extent and location of brain modes, highlighting the importance of individual differences.

Limitations and Future Work

  • Computational Demands: While efficient for a Bayesian approach, the algorithm still requires significant computational resources (memory and time) for very large datasets, posing a challenge for extremely large cohorts.
  • HRF Simplification: The model uses a canonical double-gamma HRF, acknowledging that real HRFs are more complex and variable. However, the probabilistic nature of the prior allows some flexibility.
  • Model Identifiability: While the non-Gaussian spatial prior ensures identifiability, the temporal prior is rotationally invariant, meaning it helps identify the subspace of neural activity but not necessarily unmix modes perfectly.
  • Future Work: The authors suggest further investigation into identifying larger numbers of modes, exploring the relationship between their method and functional registration algorithms, and developing more sophisticated spatial templates. They also highlight the need for further research into how subject variability influences the dimensionality of functional parcellations.

Summary updated

BibTeX
@article{harrison2015large,
  author = {Samuel J. Harrison and Mark W. Woolrich and Emma C. Robinson and Matthew F. Glasser and Christian F. Beckmann and Mark Jenkinson and Stephen M. Smith},
  title = {{Large-scale Probabilistic Functional Modes from resting state fMRI}},
  journal = {NeuroImage},
  year = {2015},
  doi = {10.1016/j.neuroimage.2015.01.013},
}
Figure