Regression models for partially localized fMRI connectivity analyses.

Area of Science:

• Neuroscience
• Data Science
• Medical Imaging

Background:

• Resting-state fMRI connectivity analysis often assumes consistent brain region localization across subjects in template space.
• Existing methods include one-edge-at-a-time analyses, dimension reduction, and hyperalignment, each with varying localization assumptions.
• Alternative approaches like density-based methods eschew localization, while others like hyperalignment focus on functional and structural alignment.

Approach:

• Proposes simple regression models to characterize brain functional connectivity using subject-level Fisher transformed regional connection matrices.
• Utilizes geographic distance, homotopic distance, network labels, and region indicators as covariates to explain connection variations.
• Analyzes data in template space but envisions use in multi-atlas registration settings, keeping subject data in its native geometry.

Key Points:

• Network labels and regional characteristics significantly explain more variation in connections than geographic or homotopic relationships.
• Visual regions demonstrated the highest explanatory power, indicated by the largest regression coefficients.
• Subject repeatability is largely recovered with subject-level regression models, comparable to fully localized models.

Conclusions:

• The proposed regression method offers a way to characterize connectivity variations by covariate type.
• Results suggest fMRI connectivity analysis may be feasible in subject-space with reduced registration requirements (e.g., affine or multi-atlas).
• Even fully exchangeable models retain significant repeatability information, highlighting the robustness of connectivity analysis.