Brain-Source Imaging: From sparse to tensor models

Hanna Becker, Laurent Albera, Pierre Comon, Rémi Gribonval, Fabrice Wendling, Isabelle MerletIEEE Signal Process. Mag 2015 · OpenReview

TL;DR

This paper tackles the challenging problem of pinpointing the origin of brain activity from noisy EEG/MEG signals. It provides a comprehensive taxonomy and comparative analysis of existing brain source imaging techniques, categorizing them by their underlying assumptions (sparse vs. tensor models) and evaluating their performance on simulated data. The work offers valuable insights for researchers choosing the right tool for their specific brain imaging needs.

What Problem Does This Paper Address?

The core challenge addressed is brain source imaging, which involves reconstructing the location and activity of neural sources within the brain using external measurements like Electroencephalography (EEG) and Magnetoencephalography (MEG). These signals are recorded at the scalp, and the relationship between the source activity and the scalp measurements is described by a linear system. However, the number of potential source locations (thousands) far exceeds the number of sensors (hundreds), making this an underdetermined inverse problem. To overcome this ill-posedness and achieve unique solutions, researchers must make assumptions about the nature of the brain activity. This paper focuses on two major classes of assumptions: those that assume the activity is sparse (only a few sources are active at any given time) and those that leverage tensor-based models to capture multi-dimensional relationships in the data. The paper aims to provide a structured overview and comparison of methods based on these different assumptions.

Methods

The paper's primary methodological contribution is a taxonomy of brain source imaging algorithms. It categorizes methods based on the hypotheses they employ regarding the spatial, temporal, or spatio-temporal distribution of brain sources. These hypotheses include:

  • Spatial Hypotheses: Minimum energy, sparsity (in spatial or transformed domains), separability, and parameterized spatial covariance.
  • Temporal Hypotheses: Smoothness, sparsity in a transformed domain, pseudo-periodicity, and separability in time-frequency domains.
  • Spatio-Temporal Hypotheses: Synchronous dipoles.

The paper then discusses four main families of algorithms that exploit these hypotheses:

  1. Regularized Least Squares Approaches: These methods aim to find a solution that minimizes the difference between the measured and reconstructed data, while incorporating prior information through regularization terms (e.g., L1 or L2 norms). Examples include MNE, LORETA, and MCE.
  2. Tensor-Based Approaches: These methods leverage multi-dimensional data structures (tensors) and techniques like Canonical Polyadic (CP) decomposition to separate and localize multiple sources. They often exploit assumptions about the structure of the data, such as separability or pseudo-periodicity. Examples include STWV-DA.
  3. Bayesian Approaches: These methods frame the problem probabilistically, estimating the posterior distribution of the sources given the data. They utilize prior distributions for the sources and noise, often incorporating hyper-parameters that can be learned from the data. Examples include Champagne.
  4. Extended Source Scanning Methods: These approaches involve searching a dictionary of potential source locations and estimating activity based on a metric. This includes beamforming techniques and subspace-based methods (like MUSIC).

The paper also discusses identifiability and convergence properties of these methods and presents a simulation study to compare the performance of seven representative algorithms.

Datasets

The paper does not introduce new datasets. Instead, it uses simulated EEG data for evaluating the performance of the source imaging algorithms. The simulations are designed to mimic epileptic EEG activity, involving two or three quasi-simultaneous active patches with varying distances and signal strengths.

Key Insights and Ideas

  • No One-Size-Fits-All: The paper strongly emphasizes that the choice of brain source imaging algorithm depends heavily on the specific problem, the desired properties of the solution, and the validity of the underlying assumptions.
  • Hypotheses Drive Performance: The performance of any given algorithm is intrinsically linked to how well its assumptions (e.g., sparsity, smoothness, tensor structure) match the true underlying brain activity.
  • Trade-offs Exist: There are inherent trade-offs between computational complexity, accuracy, and the assumptions made. For instance, methods that identify spatially extended sources (like VB-SCCD, STWV-DA) are often more computationally demanding but can capture more complex activity patterns.
  • Identifiability is Crucial but Challenging: While theoretical conditions for unique solutions exist, they are often not fully met in real-world brain imaging scenarios, particularly for sparse source distributions due to ill-conditioned lead-field matrices. Tensor-based methods, however, tend to have better identifiability under mild conditions.
  • Pre-whitening Matters: Many algorithms benefit from pre-whitening the data or having a good estimate of the noise covariance matrix. This not only addresses assumptions about noise but also helps decorrelate lead-field vectors, improving the conditioning of the matrix and thus source localization.

Key Results

The paper presents a simulation study comparing seven algorithms (sLORETA, MCE, MxNE, VB-SCCD, STWV-DA, Champagne, and 4-ExSo-MUSIC) on simulated epileptic EEG data.

  • Localization Error (DLE): VB-SCCD generally performed best in terms of minimizing the Distance of Localization Error (DLE), especially in more complex scenarios with multiple patches. STWV-DA also showed strong performance for distributed source localization. MCE was noted as the best among focal source imaging algorithms.
  • Source Extent: Algorithms like VB-SCCD, STWV-DA, and 4-ExSo-MUSIC were better at estimating the spatial extent of sources, while methods designed for focal sources (sLORETA, MCE, MxNE, Champagne) tended to produce more localized estimates.
  • Computational Cost: CPU runtimes varied significantly, with sLORETA being the fastest and Champagne being the slowest among the tested algorithms (excluding 4-ExSo-MUSIC, which had partial C implementation).
  • Challenges: Most algorithms struggled with identifying mesial sources (close to the midline) and multiple quasi-simultaneous active sources.

Limitations and Future Work

  • Limitations:
    • The identifiability conditions for sparse source distributions are often not met in practice due to ill-conditioned lead-field matrices.
  • Tensor decomposition algorithms can suffer from multiple local minima, making it difficult to guarantee convergence to the global optimum.
  • The simulation study, while informative, is based on specific scenarios and may not fully represent the complexity of real-world brain activity.
  • Some methods require careful parameter tuning, which can be tedious.
  • Future Work:
    • Further development of approaches that impose sparsity in suitable spatial transform domains.
  • Exploring combinations of successful strategies from different established source imaging approaches (e.g., tensor-based, subspace-based, and sparsity).
  • Integrating steps of two-step procedures (like STWV-DA) into a single step to process all available information and constraints simultaneously.
  • More detailed studies are needed to confirm the observed performances of the tested algorithms.

Summary updated

Abstract

A number of application areas such as biomedical engineering require solving an underdetermined linear inverse problem. In such a case, it is necessary to make assumptions on the sources to restore identifiability. This problem is encountered in brain-source imaging when identifying the source signals from noisy electroencephalographic or magnetoencephalographic measurements. This inverse problem has been widely studied during recent decades, giving rise to an impressive number of methods using different priors. Nevertheless, a thorough study of the latter, including especially sparse and tensor-based approaches, is still missing. In this article, we propose 1) a taxonomy of the algorithms based on methodological considerations; 2) a discussion of the identifiability and convergence properties, advantages, drawbacks, and application domains of various techniques; and 3) an illustration of the performance of seven selected methods on identical data sets. Directions for future research in the area of biomedical imaging are eventually provided.
BibTeX
@inproceedings{becker2015brain,
  author = {Hanna Becker and Laurent Albera and Pierre Comon and Rémi Gribonval and Fabrice Wendling and Isabelle Merlet},
  title = {{Brain-Source Imaging: From sparse to tensor models}},
  booktitle = {IEEE Signal Process. Mag},
  year = {2015},
  url = {https://openreview.net/forum?id=N38LOunHS3},
}
Figure