Blind fMRI source unmixing via higher-order tensor decompositions

Area of Science:

• Neuroimaging
• Biomedical Data Analysis
• Computational Neuroscience

Background:

• Neuroimaging generates high-dimensional biomedical data, necessitating advanced analysis techniques.
• Tensor-based methods effectively leverage the multi-way nature of brain imaging data.
• Existing tensor methods for fMRI source localization face limitations with noise and overlapping activations, often assuming strict multilinear models.

Purpose of the Study:

• To explore the benefits of exploiting the 4D nature of brain images via higher-order tensorization of fMRI signals.
• To apply less restrictive generative models for improved fMRI blind source separation.
• To introduce and evaluate novel tensor decomposition techniques for fMRI data.

Main Methods:

• Application of higher-order block term decomposition (BTD) for fMRI blind source separation.
• Utilization of the PARAFAC2 tensor model in fMRI analysis.
• Proposal and implementation of a novel BTD extension (BTD2) combining BTD and PARAFAC2 properties.

Main Results:

• BTD and BTD2 demonstrated effectiveness in challenging fMRI scenarios.
• Simulations showed robustness against noise and spatial overlap of activation regions.
• The methods were validated on both synthetic and real fMRI datasets.

Conclusions:

• BTD and BTD2 offer significant advantages for fMRI source separation, particularly in noisy conditions.
• The proposed BTD2 model successfully addresses limitations of existing methods, including inter-subject variability.
• These advanced tensor methods provide a more effective approach to analyzing complex fMRI data.