A Bayesian particle filtering method for brain source localisation

Xi Chen, Simo Särkkä, Simon J. Godsill

TL;DR

This paper introduces a novel Bayesian particle filtering method for accurately pinpointing the location of multiple electrical sources within the brain using MEG data. By developing a more realistic head model and an advanced filtering technique, the method improves both the accuracy of source localization and the estimation of the number of active sources, which is crucial for understanding brain activity.

What Problem Does This Paper Address?

Understanding brain activity non-invasively is a major goal in neuroscience, with techniques like MEG and EEG offering valuable insights. A key challenge is "brain source localization," which aims to determine the origin of the electrical signals detected outside the skull. This is an "ill-posed inverse problem," meaning many different source configurations can produce the same measured signals, making it difficult to find a unique, accurate solution. Furthermore, the number of active sources can change over time, adding another layer of complexity. Existing methods often struggle with accurately identifying the number of sources and localizing them simultaneously, especially when multiple sources are active and their locations are unknown.

Methods

The authors propose a Gibbs Multiple Particle Filtering (GMPF) algorithm. Here's a breakdown of their key innovations:

  • Continuous Realistic Head Model: Instead of relying on discrete points, they developed a method to model the brain's cortical surface as a continuous manifold. This is achieved by interpolating between discrete mesh nodes of a realistic head model (Nolte model), allowing for more precise localization of dipoles on the cortex.
  • Gibbs Multiple Particle Filter (GMPF): This is the core of their approach. Unlike standard multiple particle filters that process each source independently, GMPF iteratively refines the state estimates of all sources within a single time step. This "Gibbs sampling" approach accounts for the statistical dependencies between sources, leading to more accurate joint state estimation.
  • Dynamic Dipole Number Model: The algorithm can estimate an unknown and time-varying number of dipoles. It uses a probabilistic model that allows for dipoles to "appear" (birth) or "disappear" (death) at each time step.
  • Probabilistic ROI Estimation: Before the main filtering step, a noise-normalized Minimum Norm Estimation (MNE) approach is used to identify potential "Regions of Interest" (ROIs) where sources might be located. This provides an initial estimate of the number of sources and helps guide the subsequent filtering.
  • Adaptive Filtering: The algorithm adaptively adjusts the number of particles and the state transition range based on the estimated localization error. This helps to reduce computational cost without sacrificing accuracy, especially when the sources are well-localized.
  • Selection Criterion: To choose the most likely dipole number and state estimate from multiple possibilities generated by the dynamic model, a selection criterion based on posterior probability is employed.

Datasets

The authors used synthetic MEG data for their evaluation. They generated this data using a realistic 1-layer head model and a 204-magnetometer sensor setup. The ground-truth dipole locations, amplitudes, and orientations were known for these synthetic datasets, allowing for a direct comparison with the algorithm's estimates.

Key Insights and Ideas

  • Continuous modeling matters: Treating the brain's cortical surface as a continuous manifold, rather than a set of discrete points, significantly improves localization accuracy.
  • Inter-source dependencies are key: Standard particle filters often assume source independence, which is not realistic. The GMPF's iterative Gibbs sampling approach effectively captures these dependencies, leading to better joint estimation.
  • Adaptive particle count is efficient: Dynamically adjusting the number of particles based on localization error is a smart way to balance computational cost and accuracy.
  • Estimating the number of sources is as important as their location: The paper highlights that accurately determining how many sources are active is critical for overall performance, and their proposed method makes significant strides in this area.

Key Results

  • The proposed GMPF algorithm demonstrated improved performance metrics in terms of both dipole number estimation and dipole localization compared to existing particle filtering methods.
  • In simulations with known dipole numbers, the continuous head model consistently outperformed the spherical and discrete head models.
  • For unknown dipole number scenarios, the GMPF algorithm showed better dipole number estimation and localization accuracy compared to the standard MPF algorithm.
  • The computational cost was reduced by the adaptive filtering scheme, with GMPF being faster than MPF in experiments. For example, in a five-dipole localization task, GMPF achieved a processing time of 28 seconds per iteration compared to MPF's 56 seconds.
  • The algorithm showed good performance in dynamic unknown number of dipole examples, where the number of sources varied over time.

Limitations and Future Work

  • Deterministic ROI estimation: The initial ROI estimation step is deterministic, which can be a limitation. The authors suggest exploring a fully probabilistic dipole number estimation method for future work.
  • Computational burden: While adaptive filtering helps, the computational cost can still be high, especially for a large number of dipoles or complex head models. Future work aims to further decrease computational time, potentially through parallel computing and Rao-Blackwellization.
  • Short-term changes in dipole number: The current method might not perform optimally when the dipole number changes dramatically in a very short time period.
  • Continuous modeling of brain sources: The authors plan to explore better continuous modeling schemes to interpret brain current sources more realistically.

Summary updated

Abstract

In this paper, we explore the multiple source localisation problem in the cerebral cortex using magnetoencephalography (MEG) data. We model neural currents as point-wise dipolar sources which dynamically evolve over time, then model dipole dynamics using a probabilistic state space model in which dipole locations are strictly constrained to lie within the cortex. Based on the proposed models, we develop a Bayesian particle filtering algorithm for localisation of both known and unknown numbers of dipoles. The algorithm consists of a region of interest (ROI) estimation step for initial dipole number estimation, a Gibbs multiple particle filter (GMPF) step for individual dipole state estimation, and a selection criterion step for selecting the final estimates. The estimated results from the ROI estimation are used to adaptively adjust particle filter's sample size to reduce the overall computational cost. The proposed models and the algorithm are tested in numerical experiments. Results are compared with existing particle filtering methods. The numerical results show that the proposed methods can achieve improved performance metrics in terms of dipole number estimation and dipole localisation.
BibTeX
@article{chen2015bayesian,
  author = {Xi Chen and Simo Särkkä and Simon J. Godsill},
  title = {{A Bayesian particle filtering method for brain source localisation}},
  journal = {Digit. Signal Process.},
  year = {2015},
  doi = {10.1016/j.dsp.2015.06.007},
}
Figure