| name | spectralot-functional-alignment |
| description | Method for geometry-aware functional alignment of fMRI data using SpectralOT to improve cross-subject decoding by embedding cortical geometry into Laplace-Beltrami eigenmodes. |
| license | Complete terms in LICENSE.txt |
| metadata | {"arxiv_id":"2607.10931","authors":["Pierre-Louis Barbarant","Florent Meyniel","Bertrand Thirion"],"submitted":"2026-07-12","comment":"Proceedings of the 9th Conference on Cognitive Computational Neuroscience, New York, NY, USA, 2026"} |
SpectralOT Functional Alignment
When to Use This Skill
Use this skill when you need to:
- Perform functional alignment of fMRI data across subjects for improved cross-subject decoding
- Incorporate cortical geometry into functional alignment procedures
- Work with surface-based fMRI data (cortical surfaces) and functional maps
- Implement geometry-aware regularization in neuroimaging analysis pipelines
Overview
SpectralOT introduces a novel functional alignment method for fMRI data that embeds cortical geometry into Laplace-Beltrami eigenmodes to regularize the alignment process. This approach balances functional feature alignment with anatomical structure preservation while maintaining computational efficiency. The method improves cross-subject decoding performance by leveraging the spectral decomposition of the Laplace-Beltrami operator on the cortical surface.
Methodology
Core Concept
The method addresses the challenge of inter-individual variability in brain response patterns by aligning functional data across individuals before training population-level decoders. Unlike traditional approaches that may distort anatomical structure, SpectralOT incorporates geometric constraints directly into the alignment framework.
Mathematical Formulation
-
Laplace-Beltrami Eigenmodes: Compute eigenmodes of the Laplace-Beltrami operator on each subject's cortical surface mesh, capturing intrinsic geometric properties.
-
Spectral Embedding: Project functional data (e.g., fMRI activation maps) onto the Laplace-Beltrami eigenbasis to obtain spectral coefficients that represent functional patterns in a geometry-aware basis.
-
Optimal Transport Alignment: Solve an optimal transport problem in the spectral domain to find the mapping that minimizes discrepancies between subjects' functional representations while respecting geometric constraints.
-
Geometric Regularization: The use of Laplace-Beltrami eigenmodes inherently regularizes the alignment to preserve cortical topology, preventing excessive warping that could distort anatomical correspondence.
Workflow
-
Preprocessing:
- Extract cortical surface meshes from structural MRI for each subject
- Preprocess fMRI data to obtain functional maps on the cortical surface
-
Spectral Decomposition:
- Compute Laplace-Beltrami eigenmodes for each surface mesh
- Project functional maps onto the eigenbasis to obtain spectral coefficients
-
Alignment Optimization:
- Formulate optimal transport problem between spectral coefficients of source and target subjects
- Solve for optimal coupling matrix using Sinkhorn algorithm or similar
- Apply the learned transformation to align functional data
-
Validation:
- Evaluate alignment quality using cross-subject decoding accuracy
- Assess preservation of anatomical landmarks and functional boundaries
Implementation Notes
- Requires surface reconstruction tools (e.g., FreeSurfer) to obtain cortical meshes
- Functional data should be sampled on the cortical surface (e.g., via surface-based smoothing)
- Choice of truncation level for Laplace-Beltrami eigenbasis affects trade-off between geometric detail and computational cost
- The method is compatible with various functional modalities (fMRI, EEG source localization, etc.) when mapped to cortical surface
Expected Outcomes
When applied correctly, this skill should enable:
- Improved cross-subject decoding accuracy compared to alignment-free or geometry-unaware methods
- Better preservation of topographically organized functional areas
- Reduced inter-subject variability in functional responses post-alignment
- Computationally efficient alignment suitable for large datasets
References
- Barbarant, P.-L., Meyniel, F., & Thirion, B. (2026). Fast Whole-Brain, Geometry-Aware Functional Alignment for Cross-Subject Decoding. arXiv:2607.10931.
- Proceedings of the 9th Conference on Cognitive Computational Neuroscience, New York, NY, USA, 2026.
Activation Keywords
- spectralot
- functional alignment
- laplace beltrami
- cortical geometry
- fmri alignment
- cross-subject decoding
- geometry-aware