Related Experiment Video
Updated: Apr 19, 2026
05:55
Modeling the Functional Network for Spatial Navigation in the Human Brain
Published on: October 13, 2023
1.7K
Geo-Mamba: Geometry-informed state-space learning of functional brain organization
Yuwei Cao1, Tingting Dan2, Yang Yang1
1School of Information Science and Technology, Yunnan Normal University, Kunming, 650500, Yunnan, China.
Medical Image Analysis
|April 17, 2026
Summary
Geo-Mamba introduces a novel geometric approach for analyzing brain connectivity data from functional magnetic resonance imaging (fMRI) and electroencephalography (EEG). This method enhances accuracy and robustness in neuroimaging analysis, paving the way for clinical applications.
Area of Science:
- Neuroimaging
- Machine Learning
- Computational Neuroscience
Background:
- Functional connectivity (FC) from fMRI is often represented on non-Euclidean spaces, such as Riemannian manifolds of symmetric positive-definite (SPD) matrices.
- Conventional Euclidean sequence models are ill-suited for these complex, non-Euclidean data structures, limiting the analysis of brain networks.
Purpose of the Study:
- To introduce Geo-Mamba, a novel geometric variant of the Mamba model designed for neuroimaging data residing on Riemannian manifolds.
- To develop a dual-path selective state-space model that effectively handles high-dimensional, non-Euclidean neuroimaging data like fMRI and EEG.
Main Methods:
- Geo-Mamba utilizes a dual-path design: a stacked path for hierarchical feature aggregation and a distillation path for geometry-aware dimensionality reduction.
- A GeoMix operator fuses complementary outputs, creating compact and discriminative SPD representations essential for manifold-based analysis.
- The model was evaluated on seven fMRI datasets (Alzheimer's, Parkinson's, Autism, contact sports study) and three EEG datasets.
Main Results:
- Geo-Mamba demonstrated consistently competitive accuracy and robustness across diverse neuroimaging benchmarks, including fMRI and EEG data.
- The model effectively captures short- and long-range dependencies while managing high-dimensional SPD inputs through geometry-aware reduction.
- Validation on clinical datasets suggests significant potential for detecting subtle brain changes and disease-related alterations.
Conclusions:
- Geo-Mamba offers a powerful new tool for analyzing complex neuroimaging data by leveraging Riemannian geometry.
- The dual-path manifold modeling approach provides a robust and scalable solution for functional connectivity analysis.
- This work highlights the potential of geometric deep learning models for advancing neuroimaging research and clinical translation.
