Similarity-guided multi-view functional brain network fusion

Abstract

Understanding the intricate patterns and interactions within functional brain networks (FBNs) is crucial for the accurate diagnosis and analysis of mental disorders. Brain function can be represented through different brain networks, each providing complementary insights into underlying neural processes. Integrating data from these different sources enables a more comprehensive and precise understanding of brain function. However, effectively combining these heterogeneous data while maintaining the structural integrity of each modality remains a critical challenge. To address this challenge, we propose an innovative fusion model for multi-view FBNs that emphasizes the preservation of shared geometric structures across views. The novelty of our model lies in two main aspects: First, we design a novel manifold regularization term that ensures the common geometric structure of FBNs is accurately captured and preserved across all views, providing strong theoretical support for robust graph construction with heterogeneous neuroimaging data. Second, we introduce a pairwise regularization function that maximizes the similarity between different views, effectively integrating complementary information while managing data heterogeneity. This dual-regularization framework uniquely addresses challenges such as small sample sizes and high-dimensional feature spaces, showcasing its distinct advantages in analyzing complex brain networks. Extensive experiments on the ABIDE dataset demonstrate that our model outperforms current state-of-the-art diagnostic methods and highlights the importance of preserving geometric structures in improving diagnostic accuracy. Additionally, our framework successfully identifies key biomarkers related to Autism Spectrum Disorder (ASD), particularly within the primary visual cortex, aligning with recent findings published in Nature in 2023. These results further validate the critical role of maintaining shared geometric structures in the effective fusion of multi-view FBN data and their pivotal contribution to mental disorder diagnosis and biomarker discovery.