M/EEG Fundamentals: Hardware, Physiology, and Source Localization

Posted on Nov 7, 2025 in Technology

M/EEG Hardware Components and Principles

1. M/EEG Basics

• EEG (Electroencephalography): Measures electrical potential (Volts) via scalp electrodes.
• MEG (Magnetoencephalography): Measures magnetic flux (Teslas) using sensors close to the scalp.
• Invasive Methods: Electrocorticography (ECoG), depth recordings.
• Quote: MEG “sees less but sees better” – David Cohen.

2. EEG Details

Signal Strength and Electrodes

• Scalp EEG: 10–100 µV (peak-to-peak).
• Dura (Brain Surface) EEG: 10–20 mV.
• Electrodes: Silver (Ag) + Silver chloride (AgCl). Rely on electrochemical reactions (oxidation/reduction).

Electrode Placement Systems

• 10-20 System: 21 electrodes, typically used for clinical applications.
• 10-10 System: Approximately 64 electrodes, typically used for research.

3. MEG and SQUIDs

• SQUID (Superconducting Quantum Interference Device): Based on the Josephson junction (quantum tunneling).
• Requirement: Requires supercooling (cryogenics) to detect tiny magnetic fields.

Quantum Mechanics Key Concepts

• Wavefunction describes electron probability.
• Tunneling allows electrons to cross barriers.
• Wavefunction collapse occurs upon measurement.

4. Noise Issues and Shielding

• Signal Strength: Approximately 50 femtoTesla (fT).
• Environmental Noise: Up to 50 million fT—a huge noise problem.

Noise Reduction Solutions

• Magnetically-Shielded Room (MSR):
  • Aluminum: High-frequency shielding.
  • Mu-metal: Low-frequency shielding.
• Active Shielding: Coils cancel external magnetic fields using a 180° phase shift.

5. MEG System Examples

Sensor Types and Function

• Pickup Coils:
  • Magnetometer: Measures the magnetic field directly.
  • Gradiometer: Measures the magnetic field gradient to cancel distant noise.

Commercial Systems Comparison

• CTF-275: 275 axial gradiometers + 28 reference channels.
• Elekta 306: Triple sensors (magnetometers + planar gradiometers).

Axial vs. Planar Gradiometers

• CTF (Axial): Maximizes Signal-to-Noise Ratio (SNR) using axial sensors and a reference array.
• Elekta (Planar): Maximizes cortical measurements using planar sensors and magnetometers.
• 4D Systems: Captures everything (magnetometers + reference array).

6. New Technology: Optically-Pumped Magnetometers (OPMs)

• Advantages: No cryogenics needed, higher SNR (closer to scalp), lightweight, head-mountable.
• How OPMs Work: Rubidium vapor absorbs laser light, becoming bulk magnetized. An external magnetic field disturbs this magnetization. A photodetector picks up the resulting changes.

Physiological Origins of M/EEG Signals

1. Key Historical Figures

• Hans Berger (1920s): Conducted the first EEG recordings.
• David Cohen (1967–68, MIT): Conducted the first MEG recordings.

2. Brain Current Sources

M/EEG signals originate from ion flow (Na⁺, K⁺, Cl⁻) across neuronal membranes.

• Driving Forces: Ionic gradients and membrane permeability changes (due to neurotransmitter binding or voltage changes).
• Regulation Sites: Synapses (dendrites) and axons.

3. PSPs vs. Action Potentials (APs)

• Postsynaptic Potentials (PSPs): Ion flow occurring primarily at dendrites. These are the main contributors to M/EEG signals due to their spatial summation.
• Action Potentials (APs): Ion flow along axons. These are generally too brief and asynchronous to be detected by M/EEG.

4. Primary and Secondary Currents

• Primary (Intraneuronal) Current: Flows inside the neuron (MEG sensitive).
• Secondary (Volume) Current: Flows outside the neuron, through tissue (EEG sensitive).

5. Current Dipoles and Pyramidal Neurons

• Current Dipoles: Formed by a source (positive) and a sink (negative). The separation of this dipole produces detectable fields.
• Pyramidal Neurons: Ideal for generating detectable dipoles. They constitute 85% of cortical neurons and are oriented perpendicular to the cortex surface, allowing for effective summation.

6. EEG and MEG Signal Origins Compared

• EEG: Detects volume currents through tissue layers (conductivity matters significantly).
• MEG: Detects magnetic fields generated by primary (intraneuronal) currents.

7. Skull Effects and Volume Conduction

• The skull significantly affects EEG signals (blurring and attenuation).
• The skull has little impact on MEG signals.

8. Current Orientation and Detection Sensitivity

• Radial Currents: Outward from the cortex (EEG is better at detecting these).
• Tangential Currents: Parallel to the surface (MEG is better at detecting these).
• The brain’s irregular shape and neuron bending allow MEG to pick up both orientations to some extent.

9. Source Depth Sensitivity

• Deeper sources result in weaker signals.
• MEG is generally more sensitive to shallow (cortical) sources.
• EEG is better at detecting deeper sources, but the resulting signal is blurrier.

10. Noise Sources in M/EEG

Intrinsic Noise (Thermal)

• Random current fluctuations originating from electrodes, gel, and tissue.

External Noise (Electromagnetic)

• Power line interference (60 Hz or 50 Hz).

11. Biological Artifacts

• Movement: A significant problem for MEG (due to fixed sensors), less so for EEG.
• Ocular: Eye blinks and movements (EOG).
• Cardiac: Heartbeat artifacts (ECG).
• Muscle Activity: Electromyography (EMG) noise.
• Metal Artifacts: Jewelry, dental work, or tattoos.

12. Artifact Removal Methods

• Electrooculography (EOG): Used to track and model eye movement/blinks.
• Independent Components Analysis (ICA): A computational method used to separate and remove artifacts from the neural signal.

Signal Acquisition and Averaging Techniques

1. Signal Acquisition and Digitization

Analog signals are continuous and must be digitized for computer analysis.

Analog-to-Digital Conversion (ADC) Process

• Sampling: Measuring amplitude at fixed time intervals (sampling frequency $f_s = 1/\Delta t$).
• Quantization: Converting continuous amplitudes into discrete levels (e.g., 24-bit resolution, approximately 60 nV).

2. Sampling Rate and Aliasing

• Nyquist Theorem: The sample rate ($f_s$) must be greater than $2 \times$ the maximum signal frequency ($f_m$) to avoid aliasing.
• Undersampling: Leads to distortion (aliasing), where high frequencies appear as false low frequencies.

3. Signal Strength Measures

• Root Mean Square (RMS): Represents the average signal strength over time. $RMS = \sqrt{[(x_1^2 + x_2^2 + \dots)/n]}$.
• Peak-to-Peak: The difference between the highest positive and lowest negative peaks. Useful for detecting transient events like spikes and seizures.

4. Signal-to-Noise Ratio (SNR)

• $SNR = \text{Signal Power} / \text{Noise Power}$.
• Power ($P$) is the average of squared amplitudes.
• $SNR (\text{dB}) = 20 \log_{10} (\text{RMS}_{\text{signal}} / \text{RMS}_{\text{noise}})$.

5. The Signal Averaging Process

Signal averaging is used to extract weak, time-locked neural responses (Evoked Potentials/Fields) from background noise.

• The stimulus (visual, auditory, etc.) is presented repeatedly.
• The brain response is assumed to be identical across trials.
• Epochs are time-locked to the stimulus onset and averaged.
• The signal reinforces, while random noise averages out.
• SNR improves proportionally to the square root of the number of epochs averaged.

6. Residual Noise Estimation

Odd-even epoch averaging and subtraction is a method used to estimate the residual noise or errors remaining after the averaging process.

7. Evoked Potentials Examples

• Auditory Evoked Potential (AEP):
  • Stimulus: Clicks or tone bursts.
  • Short latency: Brainstem (~1–10 ms).
  • Middle latency: Subcortical (~10–50 ms).
  • Long latency: Cortical (>50 ms).
• Visual Evoked Potential (VEP):
  • Stimulus: Flash or checkerboard reversal.
  • Recorded primarily from occipital regions.

Oscillations and Frequency Analysis in M/EEG

1. The Fourier Transform (FT)

The Fourier Transform converts time-domain signals into the frequency domain (representing the signal as a sum of sinusoids). It is key for understanding M/EEG frequency content.

• Discrete Fourier Transform (DFT): FT adapted for sampled signals.
  • Frequency resolution $= 1 / \text{Total Time} (T) = f_s / n$.
• Fast Fourier Transform (FFT): A fast algorithm for computing the DFT.
  • Computational complexity is $n \log_2 n$.

2. Preparing EEG for Frequency Analysis

• Stationarity: Features must be stable over time. Typically, segments of approximately 2 seconds of EEG are used.
• The FFT is applied to estimate the power spectra of these segments.

3. Digital Filters

Filters are used to remove noise or isolate desired frequency ranges.

Types of Filters

• Low-pass: Keeps low frequencies, blocks high frequencies.
• High-pass: Keeps high frequencies, blocks low frequencies.
• Band-pass: Passes a specific frequency range.
• Band-stop (Notch): Blocks a specific frequency range (e.g., 60 Hz power line noise).

Anti-Aliasing Filter

This is a low-pass analog filter applied before sampling to prevent high-frequency aliasing (cutoff frequency must be less than $f_s/2$).

4. Filters and Phase Distortion

Filters may alter the signal phase, which is critical for timing analyses.

• Zero-phase: No phase distortion introduced.
• Linear-phase: Introduces a constant time delay across all frequencies.
• Nonlinear-phase: Introduces variable delay and distortion.

5. Time-Frequency Analysis

This technique studies how frequency content evolves over time. It is important because brain activity is often nonstationary.

• Stockwell Transform: A method that clearly shows both time and frequency changes simultaneously.

6. Evoked Responses vs. Phase Resetting Models

Time-domain averaging can lead to signal cancellation if responses are not perfectly phase-locked.

• Up to 90% of the M/EEG response may be non-phase-locked.

Two Models Explaining Event-Related Activity (ERP/ERF)

• Additive Model: The ERP/ERF is a new signal added to the background ongoing activity.
• Phase-Reset Model: The ERP/ERF results from the phase re-alignment of ongoing oscillations, often without a net power increase.

7. Response Types

• Evoked: Phase locked to the stimulus, consistent timing across trials.
• Induced: Not phase locked to the stimulus, timing varies, and cancels out during simple averaging.
• Spontaneous: Not related to any specific stimulus.

M/EEG Source Reconstruction and Localization

1. The Goal of Source Reconstruction

• Measurements: 2D sensor data (scalp/helmet).
• Sources: 3D brain space.
• Goal: Reconstruct 3D brain activity from 2D recordings.

2. The Inverse Problem vs. The Forward Problem

The core challenge in M/EEG source localization is the Inverse Problem.

• The Inverse Problem: Inferring the sources from the sensor data. This is an ill-posed problem because there is no unique solution (Helmholtz Principle, 1853). Infinitely many source configurations can explain the same sensor data.
• The Forward Problem: Predicting the sensor data given known sources. This has an analytic solution.

3. Key Components for Source Modeling

• Sensor Locations: Defined relative to anatomical landmarks (Nasion, Left Preauricular Point (LPA), Right Preauricular Point (RPA)).
• Volume Conductor Model: Models the conductivity of different tissue layers (e.g., scalp, skull, brain).
• Source Model: Defines the possible locations and orientations of brain current sources (e.g., dipoles distributed across the cortex).

4. Fiducials and Head Tracking

Induction coils placed on fiducials (Nasion, LPA, RPA) allow for continuous head position tracking, which is crucial, especially in MEG systems like CTF.

5. MRI and MEG Coregistration

This process aligns the anatomical MRI scan with the M/EEG measurements. It is more complex than fMRI alignment due to the different modalities involved.

6. Head Models (Volume Conductor Models)

• MEG Models: Single sphere, local spheres, or realistic single shell models.
• EEG Models: Multi-layer Boundary Element Models (BEM) or Finite Element Models (FEM).

7. The Lead Field Matrix (L)

The Lead Field Matrix maps source locations to the resulting sensor patterns.

• Dimensions: $m \times n$ (where $m$ = number of sensors, $n$ = number of sources).
• It relates the source activity vector ($n \times 1$) to the sensor measurement vector ($m \times 1$).

8. Dipole Fitting (Equivalent Current Dipoles)

This method assumes the activity is generated by one or a few focal sources.

1. Guess the dipole parameters (location, orientation, amplitude).
2. Calculate the forward prediction based on the guess.
3. Minimize the difference from the observed data (using least squares).

Challenges and Best Use

• Challenges: Susceptible to local minima traps, unstable with too many dipoles, difficult to apply consistently across multiple subjects.
• Best for: Time-averaged data (ERP/ERF) where sources are known to be focal (e.g., early sensory responses).

9. Distributed Source Solutions

These methods estimate activity across a large number of sources covering the entire brain surface, rather than just a few dipoles.

• L2 Minimum Norm Estimation (MNE): Finds the solution with the minimum overall energy (smallest current magnitude).
• L1 Minimum Current Estimate (MCE): Favors a sparse solution (fewer active sources).
• sLORETA (Standardized Low-Resolution Electromagnetic Tomography): Provides smooth, standardized localization.
• dSPM (Dynamic Statistical Parametric Mapping): A noise-normalized version of MNE.
• Beamformers: Adaptive spatial filters (often used for power estimation, not strictly true inverse solutions).

10. Typical M/EEG Processing Sequence

1. Clean M/EEG data (artifact removal).
2. Sensor-space analysis (optional publication point).
3. Acquire MRI (critical for timing).
4. Extract scalp and brain surfaces (especially for EEG).
5. Coregister MRI and M/EEG data.
6. Fit the head model (volume conductor).
7. Perform source analysis (Dipole fitting or distributed models).

Advanced Source Localization: MNE and Beamforming

1. Review: The Lead Field Matrix (L)

The Lead Field Matrix $L$ ($m \times n$) maps brain sources ($n$) to sensor signals ($m$).

2. Minimum Norm Estimation (MNE)

MNE solves for the source activity $Q(t)$ from the sensor data $B(t)$ using the formula: $Q(t) = (L^T L)^{-1} L^T B(t)$. This method finds the solution with the smallest overall energy that fits the measured data.

Key Characteristics of MNE

• Underdetermined: Typically, there are far more sources ($n$) than sensors ($m$).
• Minimization: Minimizes the difference between measured and predicted fields.
• Depth Bias: Superficial sources are inherently favored unless specific depth weighting corrections are applied.
• Regularization: Used to account for measurement noise and stabilize the solution.
• Best for: High-SNR, phase-locked early sensory responses (Evoked activity).
• Poor for: Induced, non-phase-locked activity.

3. Limitations of Time-Domain Averaging

• Jitter: Timing variability (jitter) between trials reduces the strength of the averaged signal.
• Cancellation: Non-phase-locked activity (induced responses) is canceled out during averaging.

4. Event-Related Desynchronization (ERD)

• Definition: A decrease in power in specific frequency bands following a stimulus.
• Detection: Often detected effectively using Beamforming techniques (less so with MNE).
• Relationship to fMRI: fMRI BOLD signals often show an increase in activity at sites exhibiting ERD.

5. Adaptive Beamforming (Spatial Filtering)

Beamforming uses a weighted sum of sensor signals to estimate source power at a specific point in the brain.

The source estimate $Q_k(t)$ at location $k$ is calculated as $Q_k(t) = W_k B(t)$, where $W_k$ is the spatial filter weight vector and $C$ is the data covariance matrix.

Key Characteristics of Beamforming

• Spatial Filter: Focuses on one location while suppressing signals originating from other locations.
• Scanning: The entire brain is scanned by moving the focal point across a predefined grid.
• Assumption: Assumes sources are uncorrelated (though some correlation is tolerable).
• Best for: Non-phase-locked (induced) activity.
• Advantage: Does not require predefined source locations.
• Requirements: Requires multiple time samples to accurately estimate the covariance matrix ($C$). Requires depth normalization to correct for increased noise at deeper sources.

6. Summary Comparison of Source Localization Methods