fMRI-EEG integrated cortical source imaging by use of time-variant spatial constraints

Zhongming Liu, Bin HeNeuroImage 2008 · DOI

TL;DR

This paper introduces a novel "adaptive Wiener filter" method to fuse fMRI and EEG data for more accurate brain activity imaging. By developing a way to quantify fMRI signals and create "time-variant" spatial constraints, it overcomes limitations of previous methods that struggled with the different time scales of fMRI and EEG, leading to improved spatial and temporal resolution in brain imaging.

What Problem Does This Paper Address?

Researchers often want to combine the strengths of functional Magnetic Resonance Imaging (fMRI) and Electroencephalography (EEG) for neuroimaging. fMRI offers high spatial resolution, showing where in the brain activity is happening, but its temporal resolution is slow (seconds). EEG, on the other hand, has excellent temporal resolution (milliseconds), showing when activity occurs, but it suffers from poor spatial localization. A major challenge in integrating these modalities is the "spatiotemporal mismatch": fMRI signals are sluggish and reflect averaged activity over time, while EEG captures rapid, instantaneous neural events. This mismatch makes it difficult to accurately map EEG sources using fMRI as a spatial guide, leading to issues like fMRI "false positives" (areas showing fMRI activity but not actual EEG sources) and "false negatives" (areas with EEG activity missed by fMRI). This paper aims to solve this by developing a framework that accounts for these dynamic mismatches.

Methods

The core of this paper is the development of an "adaptive Wiener filter" for fMRI-EEG integration. Key methodological advancements include:

  • Quantifying fMRI Signals: The authors propose a method to quantify fMRI signals, assuming a linear neurovascular coupling. This quantification estimates the fMRI signal in each voxel as being proportional to the time integral of the power of local electrical current during the event-related potential (ERP) period. This provides a more physically interpretable link between fMRI and EEG signals.
  • Time-Variant Spatial Constraints: Instead of using a static fMRI activation map as a spatial prior (which can lead to mismatches), the proposed method derives "time-variant" spatial constraints. These constraints are in the form of source covariance matrices that are updated dynamically.
  • Adaptive Wiener Filter: The EEG inverse problem (estimating source activity from scalp potentials) is solved using an adaptive Wiener filter. This filter incorporates the time-variant source covariance matrices, which are estimated by fusing information from both fMRI (providing a "static" constraint on the integrated source variance) and segmented EEG data (providing temporal dynamics). This data-driven approach allows the spatial constraints to adapt to the evolving neural activity.
  • Singular Value Decomposition (SVD): SVD is used on segmented EEG epochs to estimate the temporal variation of source variances, which is then combined with the fMRI-derived integrated variance to form the time-variant prior.

Datasets

The paper uses two types of datasets:

  • Simulated Data: Extensive computer simulations were performed to evaluate the proposed method. These simulations involved generating realistic fMRI and EEG signals based on known cortical source activities, allowing for controlled testing of the algorithm's performance under various conditions, including fMRI false positives and negatives, and complex temporal dynamics.
  • Real Experimental Data: Real fMRI and EEG data were collected from a single healthy subject during a unilateral checkerboard visual stimulation experiment. This dataset was used to demonstrate the practical applicability of the adaptive Wiener filter.

Key Insights and Ideas

  • Bridging the Time Scale Gap: The most significant insight is the development of a method that explicitly addresses the temporal mismatch between fMRI and EEG. By quantifying fMRI as an integral of electrical current power and using this to inform a time-variant spatial prior, the algorithm can better align the spatial information from fMRI with the temporal information from EEG.
  • fMRI as a "Static" Constraint on Integrated Variance: The paper reframes the role of fMRI. Instead of a direct spatial map, fMRI quantifies the time-integral of source variance, providing a "static" constraint that is then dynamically updated by EEG data.
  • Adaptive Priors for Robustness: The "adaptive" nature of the spatial constraints is crucial. It allows the algorithm to be more robust to fMRI false positives and negatives, as the prior information is not fixed but evolves with the data.
  • Improved Spatial Specificity: By leveraging fMRI's spatial resolution and EEG's temporal resolution, the method achieves a better balance, leading to more localized and accurate source imaging compared to using either modality alone or traditional fMRI-constrained methods.

Key Results

  • Simulation Performance: In simulations, the adaptive Wiener filter demonstrated robust performance in reconstructing cortical current density and estimating source time courses with millisecond temporal resolution. It was effective even with fMRI false positives and false negatives.
  • Waveform Estimation: The method reliably estimated source waveforms for temporally correlated or uncorrelated, sustained or transient, and complex temporal dynamics.
  • Real Data Application: Applying the method to real visual stimulation data yielded a time series of reconstructed cortical current density images that aligned with the known hierarchical organization of visual processing.
  • Spatial Resolution: The proposed method achieved spatial resolution nearly as high as fMRI, while significantly improving upon the spatial resolution of EEG-alone imaging.
  • Temporal Resolution: The method successfully captured millisecond temporal dynamics, a feat not possible with fMRI alone.

Limitations and Future Work

  • Linearity Assumption: The method relies on the assumption of a linear relationship between neural activity and fMRI response. The authors acknowledge that nonlinearities can occur, particularly under repetitive stimulation or due to vascular adaptation, though they argue the linear model is a good approximation for many experimental paradigms.
  • Co-registration Accuracy: While not a limitation of the algorithm itself, the authors note that errors in co-registering fMRI and EEG data can still impact results. They suggest advanced geometric techniques to mitigate this.
  • Noise Covariance: While the SVD approach can estimate time-variant noise covariance, the authors suggest it may not be necessary as noise is often assumed to be time-invariant and can be estimated from pre-onset EEG data.
  • Future Work: The paper suggests that the proposed fMRI quantification method could be beneficial even in fMRI-only studies. They also mention the potential for further exploration of symmetric Bayesian approaches for fMRI-EEG fusion. The authors also note that it remains unclear how recent findings on the correlation of BOLD signals with frequency components of EEG power spectrum could be integrated.

Summary updated

BibTeX
@article{liu2008fmri,
  author = {Zhongming Liu and Bin He},
  title = {{fMRI-EEG integrated cortical source imaging by use of time-variant spatial constraints}},
  journal = {NeuroImage},
  year = {2008},
  doi = {10.1016/j.neuroimage.2007.10.003},
}
Figure