Non-invasive electrophysiology—via electroencephalography (EEG) and magnetoencephalography (MEG)—offers an unparalleled window into the human brain, capturing neural dynamics on a millisecond scale. However, these techniques suffer from a fundamental spatial limitation: signal mixing.
As electrical currents travel from neurons in the cortex through the cerebrospinal fluid, skull, and scalp, they spread out. This phenomenon, known as the volume conductor effect, means that a single scalp sensor records a blurred mixture of signals from many different brain regions. As Dongyang Li notes in Visual Decoding and Reconstruction via EEG Embeddings with Guided Diffusion, operating strictly on this sensor-level data obscures the true spatial origins of neural activity. To see the brain clearly, we must map these blurred surface measurements back to their three-dimensional anatomical sources—a process known as cortical source localization or source imaging.
The Forward Model: Simulating the Physics
To solve the mystery of where a signal originated, we first need to understand how it travels. This is the forward problem. If we place a hypothetical electrical dipole at a specific coordinate on the cortex, how does its electrical or magnetic field project across the head to reach our sensors?
To answer this, researchers construct a mathematical map called a lead field matrix. This matrix models the physical and anatomical properties of the head. In toolboxes like SEREEGA: Simulating Event-Related EEG Activity, integrating lead field theory allows researchers to simulate realistic scalp maps from known ground-truth sources, providing a baseline for validation.
The accuracy of this forward model depends heavily on the anatomical constraints we feed into it. Researchers can use high-resolution anatomical templates derived from average brains, as discussed by Runhao Lu et al. in Aperiodic and oscillatory systems underpinning human domain-general cognition. Alternatively, they can use individual structural MRI data to build personalized head models, a strategy leveraged by J. Hirvonen et al. in Whole-Brain Source-Reconstructed MEG-Data Reveal Reduced Long-Range Synchronization in Chronic Schizophrenia to bypass spatial blurring. While standardized templates are sometimes used for convenience, Huixiang Yang et al. point out in Longitudinal multitask wireless electrocorticography data from two fully implanted nonhuman primates that approximating anatomy with templates can introduce subtle localization errors compared to individualized postoperative reconstructions.
The Inverse Problem: One Scalp, Infinite Brains
Mathematically, the forward model can be expressed as a linear equation:
$$Y = LX + \epsilon$$
where $Y$ is the recorded sensor data, $L$ is the lead field matrix, $X$ is the underlying cortical source activity, and $\epsilon$ is noise.
The real challenge is the inverse problem: we observe the scalp data ($Y$) and calculate the lead field ($L$), and we want to solve for the brain activity ($X$). This problem is fundamentally ill-posed. Because the human cortex is typically modeled using thousands of discrete source locations (voxels or dipoles) while we only have a few dozen or hundreds of scalp sensors, there are infinitely many different configurations of brain activity ($X$) that can produce the exact same scalp reading ($Y$).
To find a unique, biologically plausible solution, we must introduce mathematical constraints or "priors" that bias the algorithm toward certain types of solutions.
Strategies for Source Reconstruction
Different source localization algorithms represent different physical and biological assumptions about how the brain operates.
Minimum Norm Estimation (MNE)
MNE assumes that the brain is efficient and prefers solutions that minimize overall electrical energy. Mathematically, it searches for a solution that minimizes the $L_2$-norm of the source activity ($\min \|X\|_2^2$). This results in smoothly distributed estimates of brain activity across the cortex.
In Magnetoencephalography detects phase-amplitude coupling in Parkinson’s disease, Ryohei Fukuma utilizes MNE to project sensor-level magnetic signals onto a 3D anatomical model, localizing abnormal neural oscillations to specific cortical regions. Variations include weighted minimum norm solutions, used by Yin Tian et al. in Cortical Classification with Rhythm Entropy for Error Processing in Cocktail Party Environment Based on Scalp EEG Recording to mitigate volume conduction, and Bayesian formulations of MNE, which Michael Kai Petersen et al. employ in Smartphones Get Emotional: Mind Reading Images and Reconstructing the Neural Sources to handle sparse, noisy data from mobile headsets. For sharper, more focal reconstructions, $L_1$-norm minimization can be used instead of $L_2$, as demonstrated by Mingxiong Huang et al. in Somatosensory System Deficits in Schizophrenia Revealed by MEG during a Median-Nerve Oddball Task.
Tomographic Smoothness (LORETA and sLORETA)
Low-Resolution Electromagnetic Tomography (LORETA) and its standardized variant (sLORETA) add a spatial constraint: they assume that neighboring neurons fire together, meaning that source activity should vary smoothly across adjacent cortical voxels.
This approach is widely used to pinpoint the origins of event-related potentials (ERPs). For instance, Takayuki Ito et al. use sLORETA in Event-Related Potentials Associated with Somatosensory Effect in Audio-Visual Speech Perception to infer the neural sources of tactile-auditory integration. Similarly, Elvira Khachatryan et al. leverage sLORETA in Effect of word association on linguistic event-related potentials in moderately to mildly constraining sentences to separate the neural generators of sentence-level context from word-level associations.
Beamforming (Spatial Filtering)
Instead of solving for the entire brain simultaneously, beamforming takes an alternative approach: it constructs an adaptive spatial filter for one specific target location at a time. The filter is designed to pass signals originating from that target coordinate perfectly while actively nulling or suppressing interference from all other locations based on the data's covariance matrix. In Cortical Network Dynamics of Perceptual Decision-Making in the Human Brain, Markus Siegel demonstrates how linear beamforming can isolate high-frequency oscillatory networks across the cortex with millisecond precision.
Geometric-Informed Priors
Recent innovations seek to embed the physical structure of the brain directly into the mathematical basis functions. In Advancing EEG/MEG Source Imaging with Geometric-Informed Basis Functions, Song Wang introduces Geometric-informed Basis Functions (GBFs). By decomposing the cortical surface into basis functions derived from its actual physical geometry, this method ensures that reconstructed electrical activity naturally respects the structural folds and boundaries of the brain, yielding sharper and more biologically accurate localizations.
+-----------------------------------------------------------------+
| THE INVERSE PIPELINE |
| |
| [ Scalp Sensors (Y) ] --------+ |
| | |
| v |
| [ Structural MRI ] ---> [ Lead Field (L) ] ---> [ Inverse Algo ]|
| (MNE/sLORETA) |
| | |
| v |
| [ Cortical Space ] |
| (Unblurred Sources)|
+-----------------------------------------------------------------+
From Scalp Blurring to Cortical Insights
Moving from "sensor space" to "source space" fundamentally changes what researchers can see and do, unlocking two primary advantages:
1. Uncovering Hidden Networks
Source reconstruction uncovers fine-grained effects that are otherwise invisible at the scalp level due to signal mixing. As shown in Semantic Features Reveal Different Networks During Word Processing: An EEG Source Localization Study, mapping electrical activity back to thousands of discrete cortical dipoles reveals nuanced semantic processing networks that traditional scalp EEG fails to capture.
It also allows researchers to validate cognitive theories by confirming that signals originate from the expected anatomical structures. For example, Pierre W. Ferrez uses inverse modeling in EEG-Based Brain-Computer Interaction: Improved Accuracy by Automatic Single-Trial Error Detection to prove that BCI error-monitoring signals originate specifically within the brain's natural conflict-monitoring circuitry in the anterior cingulate cortex.
2. Enhancing Brain-Computer Interfaces (BCIs)
Source localization dramatically improves the performance of BCIs and neuroprosthetics. By transforming blurred channel data into localized source-space features, systems can isolate the precise intent of a user.
Baoguo Xu et al. demonstrate in Electroencephalogram Source Imaging and Brain Network Based Natural Grasps Decoding that source-space mapping mitigates the volume conductor effect, allowing a system to decode complex natural grasps using fewer, more physically meaningful features. This spatial isolation is critical for real-time applications, whether decoding movement kinematics (Reinmar J Kobler et al. in Simultaneous decoding of velocity and speed during executed and observed tracking movements: an MEG study) or providing non-invasive "test drives" for neuroprosthetic hands before patients commit to invasive surgery (Ryohei Fukuma et al. in Closed-Loop Control of a Neuroprosthetic Hand by Magnetoencephalographic Signals).
By mathematically undoing the blurring caused by the skull and scalp, source localization bridges the gap between the "when" and the "where" of human neuroimaging, turning non-invasive surface recordings into high-fidelity maps of the living brain.