Node merging in Kohonen's self-organizing mapping of fMRI data

Area of Science:

• Neuroimaging
• Data Analysis
• Machine Learning

Background:

• Functional magnetic resonance imaging (fMRI) generates complex datasets requiring advanced analysis techniques.
• Traditional methods may not fully capture the intricate patterns within fMRI data.
• Data-driven approaches are increasingly valuable for exploring neuroimaging findings.

Purpose of the Study:

• To develop and validate a novel data-driven technique for analyzing fMRI data.
• To enhance the identification of meaningful temporal patterns in brain activity.
• To integrate machine learning with established neuroimaging analysis methods.

Main Methods:

• Application of Kohonen's self-organizing map (SOM) for unsupervised clustering of fMRI data.
• Development of a cluster merging strategy based on fMRI data reproducibility across epochs.
• Generation of 'super nodes' representing time course templates for further analysis.

Main Results:

• The SOM analysis successfully identified distinct patterns within the fMRI datasets.
• The cluster merging technique effectively consolidated similar SOM nodes into representative templates.
• Demonstrated satisfactory results on fMRI data from motor and visual paradigms.

Conclusions:

• Kohonen's SOM combined with cluster merging offers a robust data-driven approach for fMRI analysis.
• The generated time course templates facilitate subsequent template-based analyses, improving activation pattern detection.
• This integrated technique shows promise for advancing the interpretation of functional brain imaging studies.