BrainNetGFM: A Graph-Based Foundation Model for Brain Network Construction Integrating Individualized Geometry and Joint Self-Supervised Learning

Abstract

The emergence of foundation models has revolutionized neuroimaging analysis, offering universal representations for decoding complex brain functions. However, existing fMRI-based foundation models predominantly rely on standard group-level atlases neglecting inter-subject heterogeneity. They often operate on voxel-wise time series with inherent low signal-to-noise ratios (SNR) and employ single-objective pre-training strategies that fail to capture fine-grained local structures and global discriminative representations. To bridge these gaps, we present BrainNetGFM, a novel graph-based foundation model for brain network construction that integrates individualized geometry with joint self-supervised learning, pre-trained on a massive dataset of 44,476 samples. First, we introduce an individualized brain atlas construction method utilizing region growing on functional priors to ensure precise alignment of subject-specific functional topology. Brain network graphs integrating functional connectivity strength with individualized spatial geometry is constructed. Second, we propose a joint self-supervised learning (SSL) framework that synergizes generative and contrastive paradigms. Specifically, we combine graph masked autoencoders (GMAE) to reconstruct masked node and edge features for learning fine-grained local structures, alongside multi-level graph contrastive learning (GCL) spanning both node and graph levels to enforce global discriminative representations. Extensive experiments on downstream tasks including the diagnosis of 8 brain disorders, age and cognition prediction, demonstrate that BrainNetGFM outperforms 9 state-of-the-art methods. Moreover, BrainNetGFM exhibits superior interpretability, identifying neurologically meaningful network biomarkers for clinical analysis. The code is publicly available at https://github.com/BrainLabA/BrainNetGFM.