Influence of activation pattern estimates and statistical significance tests in fMRI decoding analysis.

Area of Science:

• Neuroimaging
• Cognitive Neuroscience
• Machine Learning in Neuroscience

Background:

• Multi-Voxel Pattern Analysis (MVPA) is increasingly used in functional magnetic resonance imaging (fMRI).
• Accurate estimation of brain responses is critical for MVPA.
• A systematic comparison of different response estimation methods in fMRI is lacking.

Purpose of the Study:

• To compare the efficiency of three pattern estimation methods: Least-Squares Unitary (LSU), Least-Squares All (LSA), and Least-Squares Separate (LSS).
• To evaluate these methods across different experimental designs, including sustained activity, block-design, and event-related designs.
• To compare the sensitivity of the t-test with non-parametric permutation testing methods.

Main Methods:

• Comparison of run-wise (LSU) and trial-wise (LSA, LSS) estimation methods in fMRI.
• Evaluation across sustained activity, block-design, and event-related fMRI paradigms.
• Sensitivity analysis using t-test versus permutation testing (Stelzer et al., 2013) and Threshold-Free Cluster Enhancement.

Main Results:

• Least-Squares Separate (LSS) demonstrated the highest accuracy for event-related designs with significant signal overlap.
• Stelzer's permutation method showed increased sensitivity across all settings, particularly in event-related designs.
• LSS effectively unmixes events with varying durations and substantial signal overlap, handling collinearity better than other methods.

Conclusions:

• LSS is the most accurate method for analyzing fMRI data with overlapping signals in event-related designs.
• Stelzer's permutation testing method enhances statistical sensitivity, improving the detection of informative brain regions.
• The combination of LSS and advanced statistical methods offers improved analysis of complex fMRI experimental designs.