Structured dissociative PCA methods for high dimensional neuroimaging signal decomposition.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Signal Processing

Background:

• Sparse Principal Component Analysis (SPCA) and Independent Component Analysis (ICA) are common for fMRI blind source separation.
• Isolated sparsity or independence constraints can distort spatial coherence and degrade source recovery in fMRI data, especially with overlapping networks.

Purpose of the Study:

• To propose a novel unified structured dissociative PCA (SDPCA) framework for improved fMRI source separation.
• To jointly learn dissociation and representation matrices within a single decomposition for enhanced network recovery.

Main Methods:

• Integrated spatiotemporal priors (DCTs, splines, hemodynamic models) into Singular Value Decomposition (SVD).
• Developed two algorithms, SDPCAG (block coordinate descent) and SDPCAC (coordinate descent), using adaptive row-wise sparsity and correlation-guided least-squares reconstruction.
• Employed an iterative dual-decomposition strategy for precise recovery of spatially coherent brain networks.

Main Results:

• SDPCA framework demonstrated superior performance over state-of-the-art methods (PMD, ACSDBE, SICA) across synthetic, block-design, and event-related fMRI datasets.
• SDPCAG achieved a 22% improvement in source recovery accuracy compared to ACSDBE.
• SDPCAG was 1.6 times faster than SDPCAC with comparable results.

Conclusions:

• The proposed SDPCA framework effectively disentangles overlapping sources and denoises fMRI data, maintaining spatial coherence and temporal fidelity.
• SDPCAG offers a computationally efficient and accurate solution for fMRI blind source separation.