Comparison of two exploratory data analysis methods for fMRI: unsupervised clustering versus independent component...

Area of Science:

• Neuroimaging
• Data Science
• Computational Neuroscience

Background:

• Exploratory data-driven methods are crucial for hypothesis generation in functional magnetic resonance imaging (fMRI).
• Unsupervised clustering and Independent Component Analysis (ICA) are key complementary techniques to hypothesis-led statistical inference in fMRI.

Purpose of the Study:

• To systematically compare the performance of unsupervised clustering techniques against ICA methods in fMRI data analysis.
• To evaluate the effectiveness of these methods in generating hypotheses and identifying task-related patterns.

Main Methods:

• Comparative quantitative evaluation of three clustering techniques (Self-Organizing Map, Neural Gas, Fuzzy Clustering) and three ICA methods (FastICA, Infomax, Topographic ICA).
• Assessment based on task-related activation maps, associated time-courses, and receiver operating characteristic (ROC) analysis.

Main Results:

• Independent Component Analysis (ICA) methods effectively extract features for a limited number of components but are constrained by linear mixture assumptions.
• Unsupervised clustering methods demonstrated superior classification results compared to ICA.
• ICA methods generally required less processing time than unsupervised clustering techniques.

Conclusions:

• Unsupervised clustering provides better classification performance for fMRI data compared to ICA.
• The choice between clustering and ICA depends on the specific research question, data characteristics, and computational resources.
• Both methods are valuable for hypothesis generation in fMRI, but clustering offers higher accuracy in classification tasks.