Optimising beamformer regions of interest analysis

Ashwini Oswal, Vladimir Litvak, Peter Brown, Mark W. Woolrich, Gareth R. BarnesNeuroImage 2014 · DOI

TL;DR

This paper introduces a novel two-step approach to improve the accuracy and resolution of brain activity localization using MEG/EEG data. By combining a region-of-interest (ROI) projection with Bayesian Principal Component Analysis (bPCA), the method effectively reduces the dimensionality of the data, leading to more robust covariance matrix estimation and better source localization, especially with limited data or in the presence of artifacts.

What Problem Does This Paper Address?

Estimating brain activity from MEG/EEG data often relies on beamforming, a spatial filtering technique. A critical component of beamforming is accurately estimating the data's covariance matrix. However, accurately estimating this matrix requires a large amount of data, especially as the number of sensors increases. This becomes a significant challenge when dealing with short data segments or specific phenomena like transient brain activity. Furthermore, traditional beamforming methods can be susceptible to artifacts and noise from irrelevant brain regions. This paper addresses the need for more accurate and efficient source localization in scenarios with limited data or when focusing on specific brain regions of interest (ROIs), aiming to improve robustness against artifacts and enhance spatial resolution.

Methods

The paper proposes a two-step dimensionality reduction framework:

  1. Region of Interest (ROI) Projection: This step leverages prior knowledge about the location of interest. It involves projecting the sensor data onto a subspace defined by the lead fields of the specified ROI. This effectively reduces the number of channels considered, making the subsequent covariance estimation more robust. The projection matrix is derived from the singular value decomposition of the lead field matrix.
  2. Bayesian Principal Component Analysis (bPCA): Following the ROI projection, bPCA is applied to further refine the dimensionality estimate. bPCA uses a generative model and Bayesian inference (specifically Bayesian Model Selection) to estimate the optimal number of latent components that best represent the data. This step acts as an adaptive regularizer for the covariance matrix.

The novelty lies in the synergistic combination of these two techniques. The ROI projection provides an initial, anatomically informed reduction, while bPCA offers an adaptive, data-driven refinement of dimensionality.

Datasets

The study utilizes simulated data for validation. Specifically:

  • Simulation 1: Data generated for 3 and 27 source configurations to assess the ability of the combined method to recover the true number of sources.
  • Simulation 2: A more realistic simulation involving background brain activity and a specific source of interest (primary motor cortex) to compare the performance of bPCA, ROI projection alone, and the combined approach across varying signal-to-noise ratios (SNRs) and trial lengths.
  • Real Patient Data: Data from a Parkinsonian patient with externalized DBS electrodes, recorded simultaneously with MEG and local field potentials (LFPs). This dataset was used to demonstrate the method's effectiveness in a real-world scenario with significant artifacts.

No new benchmark datasets were introduced; the focus was on evaluating the proposed method against existing techniques using simulated and publicly available (or previously published) real data.

Key Insights and Ideas

  • ROI Projection as a Pre-processor: Using an ROI projection before bPCA is more effective than bPCA alone, especially for reducing artifacts and improving statistical sensitivity. This is because the ROI projection can be tailored to suppress signals from irrelevant brain areas.
  • Synergistic Dimensionality Reduction: Combining ROI projection with bPCA offers a powerful way to estimate the optimal data dimensionality. The ROI projection provides an initial, anatomically guided reduction, while bPCA adaptively refines this estimate, leading to more robust covariance matrix estimation.
  • Improved Performance with Limited Data: The combined approach significantly outperforms standard bPCA when data is scarce (e.g., short trial lengths). It allows for achieving comparable spatial specificity to a beamformer with much more data.
  • Artifact Suppression: The ROI projection, particularly when combined with bPCA, demonstrates improved artifact suppression compared to bPCA alone, as evidenced by the patient data analysis.
  • Trade-off in ROI Projection: There's a clear trade-off in ROI projection: increasing the number of components improves the representation of sources within the ROI but also increases the spatial extent of the effective ROI, potentially including irrelevant areas. bPCA helps to mitigate this by adaptively selecting the most relevant components.

Key Results

  • Simulation 1: The combined ROI projection and bPCA method accurately recovered the true data dimensionality, even when more components were selected from the ROI than the actual number of sources.
  • Simulation 2:
    • The combined approach (ROI-27 + bPCA) consistently outperformed bPCA alone and often ROI projection alone in terms of peak t-statistic, reduced FWHM (spatial resolution), lower localization error, and improved reconstruction of source amplitude, especially at lower SNRs and shorter trial lengths.
  • For short trial lengths (700 ms), the combined method achieved similar spatial specificity to a beamformer using ten times more data.
  • The dimensionality estimate from the combined approach was more stable across varying SNRs and time windows compared to bPCA alone.
  • Patient Data: The combined ROI projection and bPCA approach yielded more focal t-images and a greater peak T statistic, with a significant number of voxels surviving family-wise error correction, indicating increased statistical sensitivity and improved artifact suppression compared to bPCA alone.

Limitations and Future Work

  • Experimenter-Defined Parameters: The choice of the anatomical ROI and the number of leadfield components to select from it are still determined by the experimenter. While guidance is provided (e.g., based on filter response profiles), this remains a subjective step.
  • ROI Selection: Selecting an ROI where no signal exists could lead to erroneous inferences.
  • Correlated Sources: The study did not explicitly examine the effect of correlated sources, particularly those slightly outside the defined ROI.
  • bPCA Assumptions: bPCA assumes data samples are independent and drawn from a Gaussian distribution, which may not always hold for real experimental data. The authors suggest kernel density estimation as an alternative for non-Gaussian data.
  • Grid Resolution vs. Beamformer Resolution: The signal-to-noise gain from subspace methods might push beamformer resolution beyond the grid resolution, potentially leading to undersampling. This might necessitate reducing grid spacing.

Future work could explore more automated methods for ROI selection, investigate the impact of correlated sources, and adapt bPCA for non-Gaussian data distributions. The authors also suggest that combining leadfields from multiple spatially distinct ROIs could be a valuable extension.

Summary updated

BibTeX
@article{oswal2014optimising,
  author = {Ashwini Oswal and Vladimir Litvak and Peter Brown and Mark W. Woolrich and Gareth R. Barnes},
  title = {{Optimising beamformer regions of interest analysis}},
  journal = {NeuroImage},
  year = {2014},
  doi = {10.1016/j.neuroimage.2014.08.019},
}
Figure