| name | geometric-brain-dynamics-mapping-v6 |
| description | Geometric Basis Functions (GBF) framework for noninvasive whole human brain dynamics mapping using participant-specific eigenmodes derived from cortical geometry. Use when working with EEG/MEG source imaging, brain dynamics reconstruction, neuroimaging inverse problems, or cortical geometry-based neural activity mapping. Enables high-fidelity spatiotemporal reconstruction of neural sources using geometric constraints. |
Geometric Brain Dynamics Mapping v6 (GBF Framework)
Geometric Basis Functions (GBF) methodology for noninvasive mapping of whole human brain dynamics using participant-specific cortical geometry eigenmodes.
Overview
The GBF framework addresses the fundamental limitation of EEG/MEG source imaging: low-dimensional, indirect observations of high-dimensional neural dynamics. By embedding participant-specific geometric basis functions (eigenmodes derived from each individual's cortical surface), GBF provides powerful anatomical constraints that resolve the inverse problem and improve reconstruction fidelity.
Core Innovation
Traditional EEG/MEG source imaging relies on simplistic or biologically implausible priors. GBF reconstructs neural sources as linear combinations of geometric basis functions, aligning source estimates with the geometric organization of neural dynamics.
Key Capabilities
- Participant-Specific Modeling: Uses each individual's cortical surface geometry
- High Localization Accuracy: Validated across multiple benchmarks and datasets
- Fast Spatiotemporal Dynamics: Captures dynamics consistent with anatomical pathways
- Compact Representation: Describes whole-brain activity with hundreds of geometric modes
- Versatile Applications: Scientific research and clinical applications
Theoretical Foundation
Geometric Basis Functions
GBFs are eigenmodes derived from the Laplace-Beltrami operator on each participant's cortical surface mesh:
Δφ_k = -λ_k φ_k
Where:
Δ is the Laplace-Beltrami operator on the cortical surface
φ_k are the eigenfunctions (geometric basis functions)
λ_k are the corresponding eigenvalues
Forward Model
The EEG/MEG forward model relates neural sources to sensor measurements:
Y = LX + ε
Where:
Y is the sensor data (channels × timepoints)
L is the leadfield matrix (channels × sources)
X is the neural source activity (sources × timepoints)
ε is noise
GBF Source Representation
Neural sources are represented as linear combinations of GBFs:
X = Φα
Where:
Φ is the GBF matrix (sources × modes)
α are the GBF coefficients (modes × timepoints)
Inverse Solution
The inverse problem becomes estimating GBF coefficients:
α̂ = argmin_α ||Y - LΦα||²_F + λR(α)
Where R(α) is a regularization term (e.g., L2 norm).
Workflow
1. Data Preparation
Requirements:
- Individual T1-weighted MRI scan
- Cortical surface reconstruction (e.g., FreeSurfer)
- EEG/MEG sensor positions co-registered to MRI space
- EEG/MEG recordings
Surface Processing:
vertices, faces = load_cortical_surface('lh.pial')
L = compute_laplacian_beltrami(vertices, faces)
eigenvalues, eigenvectors = eigsh(L, k=500, sigma=0)
2. Forward Model Computation
leadfield = compute_leadfield(
subjects_dir='subjects/',
subject='subject_id',
src=source_space,
bem=boundary_element_model,
meg=True,
mindist=5.0
)
leadfield_gbf = leadfield @ gbf_matrix
3. GBF Coefficient Estimation
from scipy.linalg import solve
lambda_reg = 0.1
alpha_hat = solve(
leadfield_gbf.T @ leadfield_gbf + lambda_reg * np.eye(n_modes),
leadfield_gbf.T @ sensor_data
)
source_activity = gbf_matrix @ alpha_hat
4. Advanced Estimation Methods
Weighted Minimum Norm:
w = np.linalg.norm(leadfield_gbf, axis=0)
W_inv = np.diag(1.0 / (w + epsilon))
alpha_hat = W_inv @ solve(
W_inv.T @ leadfield_gbf.T @ leadfield_gbf @ W_inv + lambda_reg * np.eye(n_modes),
W_inv.T @ leadfield_gbf.T @ sensor_data
)
Mixed Norm (MNE + L1):
from sklearn.linear_model import MultiTaskLasso
alpha_hat = MultiTaskLasso(
alpha=0.01,
max_iter=10000
).fit(leadfield_gbf, sensor_data).coef_.T
Implementation Details
Surface Eigenmode Computation
import numpy as np
from scipy.sparse import csr_matrix
from scipy.sparse.linalg import eigsh
def compute_cotangent_laplacian(vertices, faces):
"""
Compute cotangent Laplacian for triangular mesh.
Parameters:
-----------
vertices : ndarray (n_vertices, 3)
Vertex coordinates
faces : ndarray (n_faces, 3)
Triangle indices
Returns:
--------
L : sparse matrix
Cotangent Laplacian
"""
n_vertices = len(vertices)
v0 = vertices[faces[:, 0]]
v1 = vertices[faces[:, 1]]
v2 = vertices[faces[:, 2]]
e0 = v2 - v1
e1 = v0 - v2
e2 = v1 - v0
cot0 = np.sum(e1 * e2, axis=1) / np.linalg.norm(np.cross(e1, e2), axis=1)
cot1 = np.sum(e2 * e0, axis=1) / np.linalg.norm(np.cross(e2, e0), axis=1)
cot2 = np.sum(e0 * e1, axis=1) / np.linalg.norm(np.cross(e0, e1), axis=1)
row = np.concatenate([faces[:, 0], faces[:, 1], faces[:, 2],
faces[:, 1], faces[:, 2], faces[:, 0],
faces[:, 0], faces[:, 1], faces[:, 2]])
col = np.concatenate([faces[:, 1], faces[:, 2], faces[:, 0],
faces[:, 0], faces[:, ], faces[:, ],
faces[:, ], faces[:, ], faces[:, ]])
data = np.concatenate([cot2, cot0, cot1,
cot2, cot0, cot1,
-(cot0 + cot1 + cot2)] * )
L = csr_matrix((data, (row, col)), shape=(n_vertices, n_vertices))
L
():
L = compute_cotangent_laplacian(vertices, faces)
eigenvalues, eigenvectors = eigsh(L, k=n_modes + , sigma=)
eigenvalues = eigenvalues[:]
eigenvectors = eigenvectors[:, :]
eigenvalues, eigenvectors
Sensor-Space to Source-Space Mapping
def map_sensors_to_gbf_space(sensor_data, leadfield_gbf, method='mne', lambda_reg=0.1):
"""
Map sensor-space EEG/MEG data to GBF coefficients.
Parameters:
-----------
sensor_data : ndarray (n_sensors, n_timepoints)
EEG/MEG sensor recordings
leadfield_gbf : ndarray (n_sensors, n_modes)
Leadfield projected to GBF space
method : str
'mne' (minimum norm), 'wmne' (weighted), 'lasso'
lambda_reg : float
Regularization parameter
Returns:
--------
alpha : ndarray (n_modes, n_timepoints)
GBF coefficients
"""
if method == 'mne':
LTL = leadfield_gbf.T @ leadfield_gbf
alpha = np.linalg.solve(LTL + lambda_reg * np.eye(LTL.shape[0]),
leadfield_gbf.T @ sensor_data)
elif method == 'wmne':
w = np.linalg.norm(leadfield_gbf, axis=0)
W = np.diag(1.0 / (w + 1e-6))
LTL = W @ leadfield_gbf.T @ leadfield_gbf @ W
alpha = W @ np.linalg.solve(LTL + lambda_reg * np.eye(LTL.shape[0]),
W @ leadfield_gbf.T @ sensor_data)
elif method == 'lasso':
from sklearn.linear_model import MultiTaskLasso
alpha = MultiTaskLasso(alpha=lambda_reg, max_iter=10000).fit(
leadfield_gbf, sensor_data.T).coef_.T
return alpha
Source Activity Visualization
def visualize_gbf_reconstruction(vertices, faces, gbf_coefficients, gbf_matrix,
timepoint=0, cmap='coolwarm'):
"""
Visualize reconstructed source activity on cortical surface.
Parameters:
-----------
vertices : ndarray
Surface vertices
faces : ndarray
Surface faces
gbf_coefficients : ndarray (n_modes, n_timepoints)
GBF coefficients over time
gbf_matrix : ndarray (n_vertices, n_modes)
GBF matrix
timepoint : int
Timepoint to visualize
cmap : str
Colormap name
"""
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
activity = gbf_matrix @ gbf_coefficients[:, timepoint]
fig = plt.figure(figsize=(12, 10))
ax = fig.add_subplot(111, projection='3d')
triang = ax.plot_trisurf(vertices[:, 0], vertices[:, 1], vertices[:, 2],
triangles=faces, cmap=cmap,
linewidth=0, antialiased=False)
triang.set_array(activity)
plt.colorbar(triang, ax=ax, label='Activity')
ax.set_title(f'Source Activity at t={timepoint}')
plt.tight_layout()
return fig
Validation Results
The GBF framework has been validated across multiple datasets:
Meta-Source Benchmark
- High localization accuracy for simulated sources
- Improved spatial precision compared to traditional methods
Task-Evoked Data
- Accurate reconstruction of evoked responses
- Consistent with known functional neuroanatomy
Resting-State Networks
- Captures canonical resting-state networks
- Spatial patterns match fMRI-derived networks
Intracranial Stimulation
- Validated against ground-truth intracranial recordings
- High correlation with stimulated cortical regions
Epilepsy Data
- Accurate localization of epileptogenic zones
- Clinical utility for pre-surgical evaluation
Best Practices
1. Mode Selection
How many GBF modes?
- Default: 200-500 modes capture most relevant spatial scales
- More modes: Higher spatial resolution, more computation
- Fewer modes: Smoother solutions, faster computation
- Rule of thumb: Include modes up to spatial frequency relevant for your research question
2. Regularization
L2 regularization (λ):
- Start with λ = 0.1 (SNR-dependent)
- Higher λ: Smoother solutions, less overfitting
- Lower λ: Sharper reconstructions, risk of noise amplification
- Cross-validation recommended for optimal selection
3. Surface Processing
Cortical surface quality:
- Ensure accurate surface reconstruction
- Check for topological defects
- Smooth surface if needed (e.g., 10 iterations)
- Use inflated surfaces for visualization
4. Sensor Co-registration
Accurate head positioning:
- Use fiducial-based co-registration
- Verify alignment with scalp surface
- Consider using ICP refinement
- Check for systematic displacements
Common Pitfalls
1. Insufficient Modes
Problem: Using too few modes loses spatial detail
Solution: Start with 500 modes, validate with localization metrics
2. Over-regularization
Problem: Excessive smoothing obscures true activity patterns
Solution: Use cross-validation to optimize λ
3. Leadfield Errors
Problem: Inaccurate forward model causes systematic biases
Solution: Careful BEM model construction, validation with phantom data
4. Surface Misalignment
Problem: Mismatch between electrode positions and anatomy
Solution: Rigorous co-registration, visual inspection
Comparison with Other Methods
| Method | Spatial Resolution | Biological Plausibility | Computational Cost | Individual Variability |
|---|
| GBF | High | High (geometry-based) | Medium | Participant-specific |
| MNE | Medium | Low | Low | Generic |
| dSPM | Medium | Low | Low | Generic |
| sLORETA | Medium | Medium | Low | Generic |
| Beamformer | High | Low | Medium | Participant-specific |
Applications
Research Applications
- Cognitive Neuroscience: Mapping task-related brain activity
- Clinical Research: Studying pathological brain dynamics
- Brain-Computer Interfaces: High-resolution feature extraction
- Network Neuroscience: Studying large-scale brain networks
Clinical Applications
- Epilepsy: Localization of epileptogenic zones
- Stroke: Mapping peri-lesional activity
- Brain Tumors: Pre-surgical functional mapping
- Neurodegenerative Diseases: Tracking disease progression
References
Primary Source
- Wang, S., Lou, K., Wei, C., et al. (2026). A geometry aware framework enhances noninvasive mapping of whole human brain dynamics. arXiv:2604.25592 [q-bio.NC].
Related Methods
- Laplace-Beltrami eigenfunctions for cortical surface analysis
- Minimum norm estimation (MNE) for EEG/MEG source imaging
- Eigenmode-based brain mapping (BrainEigen)
- Geometric constraints in neural mass models
Software Dependencies
- MNE-Python: EEG/MEG analysis
- FreeSurfer: Cortical surface reconstruction
- NumPy/SciPy: Numerical computations
- Scikit-learn: Machine learning utilities
Further Reading
See references/ directory for:
mathematical_formulation.md: Detailed mathematical derivations
validation_studies.md: Comprehensive validation results
implementation_examples.md: Complete code examples
clinical_applications.md: Clinical use cases and protocols
Updates
v6 (April 2026):
- Initial skill creation based on arXiv:2604.25592
- Comprehensive workflow documentation
- Validation results from Meta-Source Benchmark
- Clinical application guidelines