Enhanced Brain Graph Construction in Neuroimaging: A Data-Centric AI Approach

Loading...
Thumbnail Image

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

The construction of brain graphs from functional Magnetic Resonance Imaging (fMRI) data plays a crucial role in leveraging Graph Neural Networks (GNNs) for neuroimaging analysis. However, existing approaches are often constrained by limitations in signal selection, suboptimal topology extraction, and the omission of dynamic functional interactions. To address these limitations, a Data-Centric AI framework is introduced to systematically explore a design space for brain graph construction, optimizing high-amplitude signal retention, topological representations, and graph featurization. Specifically, a high-amplitude signal retention strategy is developed to preserve only high-amplitude blood-oxygen-level-dependent (BOLD) signal fluctuations, retaining 30% of the original data while enhancing co-activation patterns. Furthermore, alternative correlation metrics and a unified graph topology are explored to more effectively capture robust functional connectivity. For advanced graph featurization, lagged correlation is integrated to incorporate dynamic properties, capturing temporal delays between brain region activations. Additionally, edge features are introduced to encode multi-scale and multi-metric functional connectivity, enriching the structural expressiveness of the graphs. These innovations enable brain graphs to more comprehensively reflect the evolving and complex interactions between brain regions. Extensive experiments are conducted on the HCP1200 and ABIDE datasets, demonstrating that the proposed strategies generally improve downstream classification performance across multiple GNNs and ROI settings. Building on the proposed strategies, we systematically explore the data-centric design space to optimize brain graph representations, demonstrating consistent and significant improvements in classification accuracy and representation quality compared with the baseline. Our methods offer a more robust and generalizable foundation for advancing neuroimaging research with GNNs across datasets and tasks.

Description

Keywords

brain graphs, fMRI, GNNs, neuroimaging analysis, data-centric AI

Citation

Endorsement

Review

Supplemented By

Referenced By