Multiclass fMRI data decoding and visualization using supervised self-organizing maps.

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Multivariate pattern decoding in fMRI often simplifies multi-class problems into binary ones.
• Current decoding analyses primarily report classification accuracy, neglecting insights from activation pattern topology.

Purpose of the Study:

• To introduce and evaluate a supervised variant of self-organizing maps (SSOMs) for decoding and visualizing fMRI data with multiple experimental conditions.
• To compare the performance of SSOMs against established classifiers like k-nearest neighbor and support vector machines.

Main Methods:

• Developed and applied SSOMs to simulated and real fMRI datasets for pattern decoding.
• Evaluated SSOM performance across varying signal-to-noise and contrast-to-noise ratios.
• Compared SSOMs with k-nearest neighbor and support vector machines (SVMs) using different feature set sizes.

Main Results:

• SSOMs demonstrated competitive classification performance against k-nearest neighbor and SVMs, particularly for multi-class fMRI decoding.
• SSOMs excelled in visualizing the topology of activation patterns, providing deeper insights into brain representations.
• Successfully decoded speaker identity from auditory cortical patterns and visual category representations in the ventral visual cortex.

Conclusions:

• SSOMs are highly suitable for decoding fMRI datasets with more than two classes.
• Combining SSOMs with feature reduction techniques (e.g., region-of-interest or searchlight) optimizes classification performance.
• The visualization capabilities of SSOMs enhance the comprehensive understanding of decoding outcomes in neuroimaging.