Contrastive Learning-Based Multi-Atlas Feature Fusion for Brain Disorder Diagnosis

Resting-state functional magnetic resonance imaging (rs-fMRI) serves as a widely used non-invasive brain imaging tool for investigating neurological disorders through functional brain networks (FBN) modeling among brain regions of interest (ROIs). Deep learning frameworks, notably graph neural networks and Transformers, have advanced FBN analysis capabilities. However, existing research still faces two major challenges: (1) deep models require large labeled datasets, whereas rs-fMRI data are scarce and heterogeneous, and (2) most existing studies rely on a single brain atlas for FBN, ignoring complementary multi-scale information. To tackle these challenges, we propose MAFCF, a novel Multi-Atlas Feature Contrastive Fusion framework. Specifically, we construct multi-atlas FBN and design an adaptive graph augmentation strategy that perturbs both edges and node features to enhance the robustness of the representations. A hybrid GIN-Transformer encoder is then employed to jointly capture local subgraph patterns and global dependencies, while contrastive loss is used to guide representation learning. Furthermore, semantic preservation and cross-atlas alignment mechanisms are incorporated to ensure the consistency of feature representations. Extensive evaluations on two public datasets indicate that MAFCF consistently outperforms state-of-the-art methods in brain disorder diagnosis.