Performance evaluation of nonnegative matrix factorization algorithms to estimate task-related neuronal activities...

Area of Science:

• Neuroimaging
• Signal Processing
• Machine Learning

Background:

• Nonnegative matrix factorization (NMF) is a blind source separation technique.
• NMF's nonnegativity constraint is suitable for analyzing functional magnetic resonance imaging (fMRI) data.
• Task-related neuronal activity estimation is crucial for fMRI data interpretation.

Purpose of the Study:

• To evaluate the efficacy of four popular NMF algorithms for fMRI data analysis.
• To compare the performance of least-squares and multiplicative update NMF methods.
• To identify the most effective NMF algorithm for extracting task-related neuronal activities from fMRI data.

Main Methods:

• Applied four NMF algorithms: Alternating Least-Squares NMF (ALSNMF), Projected Gradient Descent NMF, Euclidean distance-based NMF, and Divergence-based NMF.
• Tested algorithms on simulated and real fMRI data from single subjects and group datasets.
• Evaluated algorithm performance in estimating task-related neuronal activities.

Main Results:

• All four NMF algorithms detected task-related activation in simulated data at the single-subject level.
• Multiplicative update NMF algorithms showed poor performance on real visuomotor task fMRI data.
• ALSNMF outperformed other NMF algorithms in group-level analysis for both simulated and real fMRI data.

Conclusions:

• ALSNMF is the most promising NMF algorithm for extracting task-related neuronal activities from fMRI data.
• NMF algorithms, particularly ALSNMF, offer a viable approach for fMRI data analysis.
• Further investigation into NMF for neuroimaging applications is warranted.