invertmeeg: A Benchmark and Unified Python Library for EEGInverse Solvers

This paper introduces "invertmeeg," a unified Python library and comprehensive benchmark that evaluates 106 EEG inverse solvers across diverse source configurations, revealing that while no single solver dominates all scenarios, flexible subspace and hybrid methods generally perform best overall.

The Gist

Imagine your brain is a giant, dark concert hall filled with thousands of tiny musicians (neurons) playing different instruments. You can't see inside the hall, but you have 32 microphones (electrodes) placed on the ceiling (your scalp) recording the sound.

The big mystery is: Who is playing what, and where are they standing?

This is the "EEG Inverse Problem." It's incredibly hard because:

1. There are thousands of musicians but only 32 microphones.
2. The sound bounces around the walls (the skull and skin), making it muddy.
3. Many different arrangements of musicians could create the exact same sound on the microphones.

For decades, scientists have invented different "mathematical detectives" (called inverse solvers) to guess where the musicians are. But here's the problem: these detectives live in different cities, speak different languages, and use different rulebooks. It's impossible to fairly compare them.

Enter invertmeeg, the new paper and software library by Lukas Hecker.

The "Taste Test" (The Benchmark)

Think of this paper as the ultimate blind taste test for these mathematical detectives.

• The Setup: The author created a "frozen" simulation. Imagine a virtual concert hall where we know exactly where the musicians are standing and what they are playing (the "Ground Truth").
• The Contest: He invited 106 different detectives (algorithms) to listen to the same 4 types of "muddy" recordings and guess the location of the musicians.
• The Scenarios:
  1. The Soloist: One musician playing loudly (Easy).
  2. The Band: A few musicians playing at once (Medium).
  3. The Choir: A large group of musicians spread out over a section (Hard).
  4. The Storm: Musicians playing in a hurricane of noise (Very Hard).

The New Tool: invertmeeg

Before this, if you wanted to test a new detective, you had to learn three different programming languages and download five different software packages. It was like trying to compare a Ferrari, a tractor, and a bicycle by building a new road for each one.

invertmeeg is like a universal adapter.

• It puts all 118 different detectives into one single, easy-to-use Python toolbox.
• It speaks the same language as the most popular brain-mapping software (MNE-Python).
• It forces every detective to use the exact same microphone setup and the exact same rules.

What Did They Find? (The Results)

The big takeaway is: There is no single "best" detective. It depends on the situation.

• The "Swiss Army Knives" (Hybrid & Subspace Methods):
  Methods like FLEX-GreedyML and Hydra were the overall winners. They are like detectives who can switch hats instantly. If the sound is sharp and local, they act like a sniper. If the sound is spread out, they act like a wide-net fisherman. They adapt to the situation better than anyone else.

• The "Noise Fighters" (Bayesian Methods):
  When the "storm" (noise) was loud, Subspace-SBL (a Bayesian method) was the champion. These detectives are great at ignoring the static and focusing on the signal, kind of like a noise-canceling headphone that actually understands the music.

• The "Old School" (Minimum Norm):
  The classic, go-to methods (like MNE and eLORETA) did okay, but they often got confused. They tend to spread the sound out too much, like a detective who says, "The music is coming from everywhere," even when it's just one person. They ranked lower than the newer, smarter methods in this specific test.

• The "AI Students" (Deep Learning):
  The paper also tested some AI neural networks. They were decent, but they didn't beat the top classical detectives. It's like a new student who studied hard but hasn't quite mastered the art of the trade yet. They need more training data and bigger brains to catch up.

Why Should You Care?

If you are a researcher or a doctor trying to figure out what's happening in a patient's brain:

1. Stop guessing: You no longer have to guess which software to use. This paper gives you a "menu" based on your specific problem.
2. Save time: You can use the invertmeeg library to try out 100+ different methods in minutes instead of months.
3. Better accuracy: By using the "Hybrid" or "Bayesian" detectives recommended for your specific scenario, you get a clearer picture of the brain's activity.

The Bottom Line

This paper didn't just build a better tool; it built a fair arena. It showed us that while some old methods are still useful, the future of brain imaging lies in flexible, adaptive detectives that can change their strategy depending on whether the brain is whispering, shouting, or screaming in the noise.

In short: invertmeeg is the universal translator and referee that finally lets all the brain-mapping detectives compete on a level playing field, helping us hear the brain's music more clearly than ever before.

Technical Summary

1. Problem Statement

Electroencephalography (EEG) source imaging involves reconstructing the distribution of neural source currents from sensor-level measurements. This is a severely underdetermined inverse problem (thousands of potential sources vs. hundreds of sensors) requiring regularization or prior assumptions.

Despite decades of research yielding a rich landscape of inverse methods (e.g., Minimum Norm, LORETA, Beamformers, Bayesian, Deep Learning), the field suffers from fragmentation:

• Implementation Scattering: Algorithms are distributed across incompatible software toolboxes (MNE-Python, Brainstorm, FieldTrip, SPM) with different data formats and interfaces.
• Lack of Standardization: Many algorithms lack public implementations or exist only as standalone scripts tied to specific datasets.
• Comparison Difficulty: The absence of a unified framework makes systematic, controlled comparison of methods under identical conditions (forward models, noise, metrics) nearly impossible.

2. Methodology

The paper introduces invertmeeg, a unified open-source Python library, and a frozen benchmark to address these issues.

A. Software Architecture (invertmeeg)

• Unified API: The library exposes 118 inverse solvers through a consistent two-step interface built on the MNE-Python ecosystem:
  1. make_inverse_operator(): Computes the inverse operator from a forward model and regularization parameters.
  2. apply_inverse_operator(): Applies the operator to EEG data (Evoked, Epochs, or Raw) to produce source estimates.
• Solver Factory: A Solver() factory function allows users to instantiate any solver by name (e.g., Solver("eloreta")), handling lazy loading and alias resolution.
• Common Infrastructure: All solvers inherit from a BaseSolver class, ensuring shared handling of:
  • Forward model preparation (depth weighting, channel matching).
  • Regularization parameter selection (via GCV, L-curve, or Product methods).
  • Output conversion to MNE source estimates.
• Novel Solvers: The library includes over 30 novel solver variants developed in collaboration with AI coding agents (e.g., Subspace-SBL, Chimera, Hydra), which combine geometric methods with statistical refinement.

B. The Frozen Benchmark

A standardized evaluation framework was established to compare 106 solvers across four distinct scenarios:

1. Focal Source: Single point dipole at moderate-to-high SNR (3–5 dB).
2. Multi-Source: 2–4 simultaneous focal dipoles at low-to-moderate SNR (1–3 dB).
3. Extended Source: 2–4 spatially extended contiguous Gaussian patches at moderate SNR (1–3 dB).
4. Noisy: 1–3 sources with variable extent under challenging low-SNR conditions (-5 to 0 dB).

Simulation Framework:

• Ground Truth: Uses two spatial models: Diffusion-basis patches (graph-based smoothing) and Contiguous Gaussian patches (irregular, anatomically plausible shapes grown via BFS).
• Noise Model: A "realistic" noise model combining low-rank artifacts, banded spatial correlations, and white noise, rather than simple i.i.d. Gaussian noise.
• Hardware/Setup: BioSemi-32 electrode montage, ico3 source space (1,284 fixed-orientation dipoles).

C. Evaluation Metrics

• Primary Metric: Earth Mover's Distance (EMD) (Wasserstein-1 distance) to quantify the spatial "work" required to transform the predicted source map to the ground truth.
• Secondary Metrics: Average Precision (AP) for ranking active dipoles, Mean Localization Error (MLE), and Temporal Correlation.

3. Key Contributions

1. Unified Library (invertmeeg): A single Python package providing access to 118 solvers (including Minimum Norm, LORETA, Bayesian, Beamformers, Subspace, Matching Pursuit, Deep Learning, and Hybrid methods) via a standardized API.
2. Frozen Benchmark: The first large-scale, systematic comparison of 106 solvers across four diverse source configurations using a shared simulation pipeline and evaluation protocol.
3. Novel Algorithmic Development: Introduction of new hybrid and AI-assisted solvers (e.g., Subspace-SBL, Chimera, Hydra) that dynamically switch strategies or combine subspace scanning with Bayesian refinement.
4. Reproducibility: The entire benchmark, simulation framework, and evaluation suite are open-source, allowing researchers to reproduce results and extend the library.

4. Key Results

The benchmark reveals that no single solver dominates all regimes; performance is highly dependent on source geometry and noise levels.

• Overall Ranking:

  • Top Performers: Flexible subspace scanners and hybrid switching methods achieved the best global ranks. FLEX-GreedyML and Hydra tied for the top global rank.
  • Bayesian Methods: Subspace-SBL+ and Subspace-SBL performed exceptionally well, particularly in extended-source and low-SNR scenarios.
  • Classical Defaults: Standard Minimum Norm (MNE) and eLORETA ranked significantly lower than top-tier subspace and hybrid methods in this specific setup.
  • Deep Learning: Benchmarked neural networks (CovCNN-KL, CovCNN, FC) outperformed random noise baselines but lagged substantially behind the best classical solvers, likely due to constrained training budgets in the study.

• Scenario-Specific Findings:

  • Focal/Multi-Source: Explicit source enumeration methods (GreedyML, Subspace scanners) excelled. Classical beamformers struggled with correlated sources.
  • Extended Sources: Bayesian methods with flexible priors (Subspace-SBL+) and FLEX variants outperformed sparse methods, which failed to capture extended activity.
  • Low SNR: Bayesian methods with noise learning (Subspace-SBL) were most robust. Beamformers and Minimum Norm methods degraded significantly.

• Performance vs. Speed:

  • Fast (<2s): Subspace scanners (SSM, FLEX-MUSIC) and Matching Pursuit (OMP).
  • Moderate (2–10s): Champagne variants, ESMV family.
  • Slow (>30s): Iterative Bayesian methods (VB-SBL) and Hybrid methods (Hydra, FLEX-GreedyML).
  • Deep Learning: Fast inference but high upfront training costs.

5. Significance and Implications

• Solver Selection Guidelines: The paper provides a practical guide for practitioners. For example, if sources are expected to be focal, fast subspace methods are recommended; if sources are extended or SNR is low, Bayesian refinements (Subspace-SBL) or hybrid methods are superior.
• Paradigm Shift: The results challenge the reliance on classical defaults (like MNE/eLORETA) for general-purpose imaging, suggesting that flexible subspace and hybrid approaches offer superior accuracy across diverse conditions.
• Community Standard: By unifying the interface and benchmarking protocol, invertmeeg removes the friction of comparing methods, enabling rapid prototyping and fair evaluation of future algorithms.
• AI in Research: The inclusion of solvers developed with AI coding agents demonstrates the potential of human-AI collaboration in generating novel, high-performing algorithmic combinations.

In conclusion, invertmeeg establishes a new standard for EEG source imaging research, demonstrating that the choice of inverse solver is critical and that flexible, hybrid, and Bayesian approaches currently offer the best trade-offs between accuracy and robustness across varied physiological conditions.