Bayesian brain source imaging based on combined MEG/EEG and fMRI using MCMC

Sung C. Jun, John S. George, Woohan Kim, E. Juliana Paré-Blagoev, Sergey M. Plis, Douglas M. Ranken, David M. Schmidt

TL;DR

This paper introduces a novel Bayesian inference framework to combine Magnetoencephalography (MEG) and functional Magnetic Resonance Imaging (fMRI) data for more accurate brain source localization. By treating fMRI data as a spatial prior within a probabilistic model, the method effectively leverages the complementary strengths of both modalities, leading to improved detection and localization of neural activity.

What Problem Does This Paper Address?

Understanding brain function relies on accurately pinpointing the location and timing of neural activity. Various neuroimaging techniques like MEG, EEG, and fMRI offer different perspectives, each with its own strengths and weaknesses. MEG and EEG excel at capturing the millisecond-level temporal dynamics of neural electrical currents, but their spatial resolution is limited. Conversely, fMRI provides better spatial resolution by measuring blood flow changes associated with neural activity, but its temporal resolution is much coarser (seconds). Integrating these complementary data sources is crucial for a comprehensive understanding of brain function. However, effectively combining these disparate data types, each with inherent uncertainties, poses a significant challenge. Previous attempts often involved simpler integration methods or relied on assumptions that didn't fully exploit the rich information available in both modalities.

Methods

The core of this paper is a Bayesian inference framework that integrates MEG and fMRI data. The authors employ Markov Chain Monte Carlo (MCMC) sampling, a powerful technique for exploring complex probability distributions. The key innovation lies in how fMRI data is incorporated: it's used as a spatial prior for the MEG source localization problem. This means that regions with higher fMRI activation are given a higher probability of containing neural sources in the MEG analysis.

Specifically, the method formulates a combined spatiotemporal model. It assumes neural activity can be modeled by equivalent point dipoles, characterized by their locations, orientations, and time courses. The fMRI data, transformed into statistical parametric maps (T-maps), informs the prior distribution of dipole locations. This approach is "asymmetric" because fMRI primarily contributes spatial information, while MEG provides the spatiotemporal data. The authors also introduce the concept of "active time range" for dipoles, allowing for more flexible modeling of transient neural activity. They also explore both fixed and rotating dipole source models.

Datasets

The paper utilizes both simulated and empirical datasets for evaluation.

  • Simulated MEG data: Generated to mimic empirical MEG data, allowing for a "ground truth" to assess the algorithm's performance. These simulations were created by localizing sources from empirical MEG data, generating simulated sensor activation, and adding empirical MEG noise.
  • Empirical data: Acquired from a single subject undergoing median nerve stimulation. Both MEG and fMRI data were collected under identical stimulation conditions on the same day. This dataset is used to validate the combined analysis in a real-world scenario.

Key Insights and Ideas

  • fMRI as a spatial prior is powerful: Treating fMRI activation maps as a prior probability distribution for source locations in MEG analysis significantly improves localization accuracy and helps filter out spurious sources.
  • Asymmetric integration is justified: The authors argue that the fundamental differences in temporal resolution and the nature of the signals detected by MEG and fMRI make an asymmetric approach (fMRI for spatial prior, MEG for spatiotemporal data) more practical and effective than a fully symmetrical one.
  • Bayesian inference handles uncertainty well: The probabilistic framework of Bayesian inference naturally accommodates the uncertainties inherent in both MEG and fMRI data, leading to more robust solutions.
  • Improved detectability of weak sources: The combined approach can detect neural sources that are too weak to be identified by MEG alone but are visible in fMRI.
  • Noise covariance estimation is critical: Sophisticated noise covariance modeling, particularly the multi-pair Kronecker product model, is essential for preventing spurious dipole sources, especially when fMRI information is limited.

Key Results

  • Improved localization accuracy: The combined MEG/fMRI analysis consistently showed better localization performance compared to MEG-only analysis, especially when fMRI information was appropriately weighted.
  • Reduced spurious sources: In empirical data analysis, higher fMRI weighting led to fewer spurious dipole sources, suggesting fMRI's role in disambiguating true neural activity from noise.
  • Separation of nearby sources: The method demonstrated an improved ability to separate closely located dipoles, a known challenge for MEG analysis alone.
  • Successful integration of simulated and empirical data: The proposed method was successfully tested on both simulated data with known ground truth and real empirical data, demonstrating its feasibility and capability.

Limitations and Future Work

  • Temporal resolution mismatch: The paper acknowledges the inherent time scale discrepancy between MEG (milliseconds) and fMRI (seconds) and treats fMRI as purely spatial data. Future work could explore methods to better incorporate fMRI's dynamic information if the coupling between BOLD response and neural activity is better understood.
  • Computational cost: MCMC sampling, while powerful, can be computationally intensive, especially for larger time windows or more complex models (like the rotating dipole model). The authors mention transforming to eigentime variables to speed up computation.
  • Parameter tuning: The effectiveness of the fMRI prior depends on parameters like $\alpha$ and $\eta$. While they found optimal values in simulations, these might be data-dependent. Future work could involve estimating these hyperparameters within the Bayesian framework.
  • Model complexity: The paper primarily uses a dipole model. Exploring more complex source models or incorporating anatomical constraints more directly could be future directions.
  • Symmetry in MEG/EEG integration: While the current approach is asymmetric for MEG/fMRI, the authors suggest that a symmetrical approach might be more beneficial for combining MEG and EEG due to their similar underlying physics.

Summary updated

BibTeX
@article{jun2008bayesian,
  author = {Sung C. Jun and John S. George and Woohan Kim and E. Juliana Paré-Blagoev and Sergey M. Plis and Douglas M. Ranken and David M. Schmidt},
  title = {{Bayesian brain source imaging based on combined MEG/EEG and fMRI using MCMC}},
  journal = {NeuroImage},
  year = {2008},
  doi = {10.1016/j.neuroimage.2007.12.029},
}
Figure