TL;DR
This paper introduces a novel method using generative models and Bayesian model comparison to probabilistically determine hippocampal engagement in MEG data. By comparing models with and without an explicit hippocampal source, they demonstrate that accurate co-registration (<3mm) and sufficient signal-to-noise ratio (SNR > -20dB) are crucial for reliably detecting deep brain structures like the hippocampus.
What Problem Does This Paper Address?
Magnetoencephalography (MEG) offers excellent temporal resolution for tracking brain activity, but accurately localizing signals from deep brain structures, particularly the hippocampus, has been a persistent challenge. This difficulty stems from the hippocampus's depth, complex geometry, and proximity to neocortical sources, leading to signal attenuation and potential cancellation. While previous studies have suggested hippocampal involvement, these findings often relied on statistical peaks in group-level analyses, which can be confounded by factors like image smoothing, head movement, and co-registration errors. This paper aims to move beyond simply identifying activation peaks and instead develop a method to probabilistically infer whether the hippocampus is engaged by comparing competing generative models of brain activity.
Methods
The core of this work lies in a Bayesian model comparison framework applied to simulated MEG data. The authors constructed two generative models: one representing only cortical sources and another including both cortical and hippocampal sources.
- Generative Models: These models incorporate realistic anatomical information (cortical and hippocampal surface meshes derived from MRI) and simulate neuronal current flow. The hippocampal model leverages the morphological similarity between neocortical pyramidal cells and hippocampal pyramidal cells, assuming current flow is perpendicular to the hippocampal surface.
- Simulation: They simulated single dipole sources on the hippocampal surface (20 Hz sinusoidal waveform) and added Gaussian white noise to achieve various signal-to-noise ratios (SNRs) from 0 dB down to -20 dB. Crucially, they also simulated co-registration errors by applying displacements (0, 1, 2, or 3 mm) to the anatomical mesh used for inversion relative to the simulated source.
- Source Reconstruction: The simulated data were then inverted using three different functional priors: Minimum Norm Estimate (MNE), Empirical Bayes Beamformer (EBB), and Multiple Sparse Priors (MSP). This allowed them to test the robustness of their findings across different source modeling assumptions.
- Bayesian Model Comparison: The key metric used is Free Energy (F), which approximates model evidence. By comparing the Free Energy of the "combined" (hippocampus included) model to the "cortical" (hippocampus-free) model (ΔF = F_combined - F_cortical), they could probabilistically determine which model better explained the simulated data. A positive ΔF indicates evidence for the combined model.
Datasets
The study primarily uses simulated data. Anatomical data for creating the cortical and hippocampal surface meshes were derived from a single subject's 3T Siemens Magnetom T1-weighted MRI scan, processed using the FreeSurfer software. No empirical datasets were used for the primary analysis, but the simulation setup was designed to mimic realistic empirical conditions.
Key Insights and Ideas
- Probabilistic Inference for Deep Sources: The paper moves beyond simple localization to probabilistic inference, allowing for statements about the likelihood of hippocampal engagement rather than just identifying a peak location.
- Importance of Co-registration: Co-registration error is identified as the most critical factor limiting the reliable detection of hippocampal activity. Even small errors (around 3 mm) can significantly degrade the ability to distinguish between models.
- SNR Threshold: A sensor-level SNR of -20 dB is shown to be a critical threshold for reliable inference, suggesting that sufficient signal quality is essential.
- Model Comparison as a "Simpler Solution" Detector: The Bayesian model comparison framework inherently favors more parsimonious explanations. If the hippocampus is truly active, including it in the model provides a simpler explanation than trying to account for that activity with a complex combination of cortical sources, leading to higher Free Energy.
- Specificity of Hippocampal Signal: Control simulations with cortical sources and shifted hippocampal meshes demonstrate that the observed Free Energy improvements are specific to correctly modeled hippocampal activity and not just an artifact of adding a deep source or increasing model complexity.
Key Results
- Significant Inference: With an SNR of -5 dB and no co-registration error, the EBB and MSP algorithms showed significant evidence (ΔF > 3, meaning the combined model is >20 times more likely) for the combined model when hippocampal sources were simulated. MNE showed less sensitivity.
- Co-registration Error Impact: As co-registration error increased from 0 to 3 mm, the ΔF values decreased significantly. For EBB and MNE, reliable distinction between models was lost at 3 mm error. MSP remained more robust.
- SNR Impact: Increasing noise (decreasing SNR) also reduced ΔF, but co-registration error was found to be more detrimental than SNR.
- Cortical Neighbor Control: Simulating activity on cortical vertices closest to hippocampal locations did not yield significant evidence for the combined model, indicating specificity.
- Shifted Hippocampus Control: Shifting the hippocampal mesh by as little as 0.5 cm or 2 degrees in any dimension significantly reduced or eliminated the evidence for the combined model, highlighting the importance of accurate anatomical localization.
- Dipole Localization Error (DLE): DLE results mirrored Free Energy findings, showing increased error with higher co-registration error and lower SNR, supporting Free Energy as a valid metric even without ground truth.
Limitations and Future Work
- Idealized Simulations: The simulations represent "best-case scenarios." Real-world data may involve more complex noise, multiple simultaneous sources, and less precise anatomical parcellations.
- Prior Selection: For MSP, the "correct" priors were pre-selected. In empirical data, selecting optimal priors can be computationally intensive.
- Generic Inversion Algorithms: The inversion algorithms used were generic and not individually optimized (e.g., no depth re-weighting in MNE).
- Model Comparison Limitations: Model comparison only provides relative evidence; it cannot definitively prove a model is "correct" in an absolute sense. If a relevant structure (e.g., amygdala) is not included in any model, inference could be incorrect.
- Future Work: The authors suggest incorporating more detailed hippocampal subfield models, exploring other nearby structures (amygdala, parahippocampal cortex), and testing the method with empirical data using subject-specific headcasts for improved co-registration. They also propose investigating the impact of different types of correlated noise and exploring the use of volumetric beamformers for empirical validation.
Summary updated