Embedding Functional Brain Networks in Low Dimensional Spaces Using Manifold Learning Techniques.

Area of Science:

• Neuroscience
• Data Science
• Computational Biology

Background:

• Functional magnetic resonance imaging (fMRI) data is high-dimensional and challenging to visualize.
• Manifold learning techniques reduce dimensionality, projecting functional brain networks onto lower-dimensional spaces for analysis.
• Non-linear manifold learning methods like t-SNE and UMAP are explored for brain network representation.

Purpose of the Study:

• To investigate the potential of t-SNE and UMAP for representing functional brain networks.
• To compare the performance of t-SNE and UMAP in discriminating between different brain states (resting vs. working memory).
• To evaluate the topological preservation and cross-study alignment capabilities of these manifold learning techniques.

Main Methods:

• fMRI data from the Human Connectome Project (HCP) and an aging study were used.
• Functional brain networks were generated from resting-state and working memory task data.
• t-SNE and UMAP were applied to project networks onto 2D manifolds, with performance assessed via discrimination and topology metrics.

Main Results:

• Both t-SNE and UMAP successfully discriminated between resting and working memory states.
• UMAP achieved higher classification accuracy in distinguishing between tasks.
• t-SNE demonstrated superior preservation of the high-dimensional data's topology.
• Combined data from different studies showed correct alignment of resting and memory networks.

Conclusions:

• UMAP is a powerful tool for visualizing functional brain networks, offering excellent discrimination.
• t-SNE provides better topological fidelity, which may be crucial for certain analyses.
• Manifold learning techniques can effectively align brain network data across different studies and acquisition protocols.