BrainCL: Transformer-Based Brain Network Contrastive Learning with Multi-Order Topology and Salience Masking.
Authors
Abstract
Brain network analysis based on functional magnetic resonance imaging (fMRI) is crucial for the diagnosis of neurological disorders. Recently, Transformers have been adopted for brain network analysis to mitigate the over-smoothing issue in GNNs. However, they often fail to account for complex topological properties of brain networks and tend to rely on a limited set of regions of interest (ROIs) for neurological disorder prediction. This makes these models highly sensitive to site differences and inter-subject variability, leading to suboptimal performance. In this paper, we propose a novel brain network contrastive learning framework (BrainCL) to enhance Transformer-based brain network analysis. Specifically, we first design a multi-order topology-aware Transformer (MoT-Former) that leverages a preferential random walk scheme (PRWS) and a hop-wise gated attention (HWGA) module to adaptively introduce multi-order topological inductive biases, thereby capturing informative multi-hop functional interactions. In addition, to overcome the limitation of relying on a few ROIs, we propose a salience-informed dynamic masking strategy that deliberately occludes salient ROIs to encourage MoTFormer to continuously mine subtle yet critical functional abnormalities from relatively unactivated ROIs for complementary learning, thereby generating more comprehensive representations. Finally, we develop synergistic dual-level contrastive learning to promote semantically meaningful representation invariance and construct a class-discriminative feature space, further improving model robustness. Experimental results demonstrate that BrainCL significantly outperforms existing methods and achieves excellent cross-site generalization. Furthermore, visualization results show that BrainCL can leverage information from a broader set of ROIs, which may offer a novel perspective for exploring more diverse biomarkers in future research.