Unsupervised spatiotemporal analysis of fMRI data using graph-based visualizations of self-organizing maps

Area of Science:

• Neuroimaging
• Machine Learning
• Data Visualization

Background:

• Unsupervised learning methods like self-organizing maps (SOMs) are used for functional magnetic resonance imaging (fMRI) analysis.
• Interpreting SOMs requires postprocessing to delineate clusters and identify brain features.
• Existing methods may struggle with subtle differences in brain response timings.

Purpose of the Study:

• To introduce novel graph-based visualizations for SOMs in fMRI data analysis.
• To enhance the interpretation of SOMs by capturing data distribution and temporal similarities.
• To improve the identification and classification of task-related brain regions.

Main Methods:

• Utilized graph-based visualizations to represent SOMs for fMRI data.
• Employed density-based connectivity based on data distribution across prototype receptive fields.
• Used correlation-based connectivity to capture temporal similarities between prototypes.
• Applied the approach to an fMRI reaction time experiment.

Main Results:

• Graph-based SOM visualizations effectively captured fMRI data features.
• The method successfully identified task-related brain areas in a visuo-manual response task.
• Visualization of SOMs outperformed independent component analysis and voxelwise univariate linear regression.
• Correlated time-to-peak fMRI responses with reaction time.

Conclusions:

• Graph-based SOM visualizations offer advanced cluster boundary visualization in fMRI data.
• This technique enables better separation of brain regions with minor timing differences in responses.
• The approach provides superior identification and classification of relevant brain regions compared to traditional methods.