Finding imaging patterns of structural covariance via Non-Negative Matrix Factorization.

Area of Science:

• Neuroimaging Analysis
• Computational Neuroscience
• Biomedical Data Science

Background:

• Structural neuroimaging data analysis aims to uncover brain networks and common influencing factors like genetics and pathologies.
• Traditional methods like PCA and ICA can produce dispersed components, limiting interpretability.
• Non-Negative Matrix Factorization (NNMF) offers a data-driven approach for localized component extraction.

Purpose of the Study:

• To investigate the application of Non-Negative Matrix Factorization (NNMF) for analyzing structural neuroimaging data.
• To identify co-varying brain regions that may represent underlying networks or be influenced by common mechanisms.
• To leverage NNMF's parts-based representation for interpretable dimensionality reduction.

Main Methods:

• Applied Non-Negative Matrix Factorization (NNMF) to structural neuroimaging datasets.
• Derived brain decompositions that partition the brain into consistently varying regions.
• Validated NNMF using split-sample experiments for generalization and compared it with PCA and ICA.

Main Results:

• NNMF produced sparse, parts-based representations of structural neuroimaging data at various resolutions.
• These representations align with known functional brain organization and capture pathological processes.
• NNMF-derived low-dimensional representations favorably compared to those from PCA and ICA.

Conclusions:

• NNMF provides a powerful, interpretable, and generalizable method for analyzing structural neuroimaging data.
• The technique effectively identifies localized, co-varying brain regions relevant to network structure and pathology.
• NNMF offers advantages over traditional methods for brain network and disease-related imaging studies.