TAFNet: Trusted Multiview Associative Fusion Neural Networks for Analyzing Dynamic Brain Networks.

Dynamic functional connectivity (DFC) is crucial for analyzing brain networks, as it captures the temporal dynamics of brain regions. However, most existing methods assume uniform quality across time windows, neglecting the inherent data heterogeneity in real clinical environments. This study proposes a framework using trusted multiview associative fusion neural networks (TAFNet) for analyzing dynamic brain networks. In the TAFNet framework, different temporal windows are treated as independent views, and a local-global convolutional filtering module is introduced to extract evidence from the brain network in each view. Moreover, a multiview associative fusion mechanism, which exploits mutual information to identify homogeneous relationships between views from heterogeneous multiview data, is constructed. A $\text {top-}k$ view selection strategy is designed to retain reliable views, providing a reliability-sensitive prior for subsequent fusion. The selected views are fused using the Dempster combination rule. On this basis, a dynamic trust assessment (DTA) module is designed to align evidential beliefs and uncertainty with class probabilities and further integrate them with Softmax-based probabilistic predictions, allowing explicit quantification of predictive confidence. Meanwhile, a composite loss function is formulated to jointly constrain evidence learning and probabilistic inference. The proposed framework is verified by comprehensive experiments on three real-world schizophrenia datasets and compared with the state-of-the-art methods. The results demonstrated that the proposed TAFNet framework outperforms the existing approaches, improving diagnostic accuracy by at least 2.15%.