A research-oriented Python pipeline for mapping scalp EEG activity into 3D voxel-level brain space using a hybrid information-theoretic and spatial-consistency framework.
Status: Methodology complete · Large-scale validation in progress
License: CC BY-NC 4.0 (Academic use only)
This repository provides an implementation of a hybrid mutual information (MI) + Dice coefficient scoring approach for estimating voxel-level EEG representations, combined with Leave-One-Subject-Out (LOSO) cross-validation and ICC reliability analyses.
The pipeline enables frequency-specific spatial inference across canonical EEG bands (δ, θ, α, β, γ), with validation procedures designed to reduce circularity and improve generalizability.
- Hybrid MI–Dice scoring for voxel-level EEG projection
- LOSO cross-validation for subject-level generalization
- ICC(2,k) reliability analysis for voxel-wise stability
- 7-step circularity control framework
- Frequency-specific mapping (1–45 Hz)
- Modular Python architecture for research use
This implementation is intended for non-commercial academic neuroscience research.
src/ # Core MI-Dice voxelization algorithms utils/ # Preprocessing + helper functions data/ # Example EEG matrices (sample subject) results/ # Output templates (NIfTI, ICC maps) docs/ # Supplementary documentation (WIP) LICENSE # CC BY-NC 4.0 CITATION.cff # Citation metadata README.md # This file
- Bandpass filtering (1–45 Hz)
- ICA artifact removal
- CSD transformation
- Hilbert envelope extraction
- Epoching (2 s windows)
- Gray-matter voxel grid (~200k voxels, MNI space)
- Geodesic Gaussian weighting
- Hemisphere normalization
- Hybrid score:
[ H = \alpha \cdot MI + (1 - \alpha) \cdot Dice ]
- Leave-One-Subject-Out optimization
- ICC(2,k) reliability maps
- Circularity reduction (noise-floor estimation, spatial boundary checks)
- Frequency-specific voxel maps
- Reliability volumes (NIfTI)
- Seed-to-voxel and ROI-level summaries
SPIS Dataset (N=10)
✔ Processing complete
✔ Full validation
✔ Circularity: r = 0.33
LEMON Dataset (N=40)
⏳ Validation in progress
⏳ LOSO weighting estimation
⏳ Final ICC maps forthcoming
Detailed usage will be added once LEMON validation is complete.
📖 Citation If you use this repository, please cite:
Kemik, K., Aykaç, C. (2025). EEG-fMRI Direct Signal Voxelization Pipeline (v3.4-dev). GitHub Repository: https://github.com/keremkem/eegdirectsignalvoxelization
BibTeX: @software{Kemik_Aykac_2025_voxelization, author = {Kerem Kemik and Cansu Aykaç}, title = {EEG-fMRI Direct Signal Voxelization Pipeline}, year = {2025}, version = {3.4-dev}, url = {https://github.com/keremkem/eegdirectsignalvoxelization}, note = {Academic research use only} }
⚖️ License (Academic Use Only) This software is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.
Permitted: Academic research
Educational use
Non-commercial studies
Methodological replication
Restricted: Commercial applications
Proprietary software integration
Clinical or diagnostic use
For commercial licensing inquiries: keremkemik9@gmail.com
👤 Authors Dr. Kerem Kemik Post-Doctoral Researcher — MD-PhD Neuroscience 📧 keremkemik9@gmail.com 🔗 github.com/keremkem
Dr. Cansu Aykaç - PsYD Neuropsychology
Vibecoded with Claude Sonnet 4.5.
🚧 Development Status Core MI–Dice algorithm
LOSO cross-validation module
SPIS pilot validation
LEMON large-scale validation (ongoing)
Manuscript submission (in preparation)
Last updated: November 2025