ViT-CMN: a vision transformer framework based on fusion of low- and high-order dynamic central moment networks for autism spectrum disorder diagnosis

Autism spectrum disorder (ASD) is a complex neurodevelopmental condition characterized by social communication deficits and repetitive behaviors. Current neuroimaging-based diagnostic methods often rely on functional connectivity features extracted from a single perspective, either low-order interactions or high-order co-fluctuations, limiting their capacity to capture the hierarchical and dynamic properties of brain networks. We propose ViT-CMN, a novel ASD diagnosis framework that integrates multilevel dynamic connectivity information from both low-order and high-order perspectives. First, dynamic functional connectivity networks are constructed using a sliding window approach to capture temporal variations. To enrich data diversity, a temporal reorganization-based augmentation strategy is introduced, which generates additional dynamic sequences by shifting their starting time points. For each sequence, seventh-order central moment features are extracted to enhance statistical stability over time. These multiview features are then structurally reorganized via a jigsaw-style fusion strategy into a unified 2D representation. This fused input is modeled using a Vision Transformer (ViT) to extract discriminative representations across spatial and hierarchical dimensions through self-attention mechanisms. Experiments on the autism brain imaging data exchange (ABIDE) dataset demonstrate that ViT-CMN outperforms existing baseline methods, achieving a top classification accuracy of 79.8%. The model also successfully identifies ASD-related brain regions that align with known neuropathological findings. ViT-CMN effectively addresses the limitation of single-view modeling in previous studies by structurally fusing heterogeneous dynamic features into a ViT-compatible form. The proposed approach provides a powerful and interpretable solution for ASD diagnosis, with strong potential for broader applications in neuroimaging-based disorder classification.