ViTAE-HGOT: Vision Transformer-based Autoencoder with Hypergraph Optimal Transport for cross-atlas functional connectome remapping.

Abstract

The utilization of open-source neuroimaging datasets, despite their unprecedented sample scale, poses significant challenges to cross-study comparability and multi-site data fusion due to the use of different brain atlases that generate inconsistent functional connectome (FC). Current FC remapping fusion methods often fail to account for functional semantics, resulting in compromised analytical validity due to incorrect information transport. Thus, we propose Vision Transformer-based Autoencoder with Hypergraph Optimal Transport (ViTAE-HGOT) for functional semantics preserving cross-atlas FC remapping. Specifically, the ViTAE-HGOT framework includes three stages. Firstly, the representation features of brain region are derived in a latent space by reconstruction of a given atlas FC using the ViTAE. Subsequently, the proposed HGOT (where Yeo-7 functional semantics serve as hyperedges to guide transport) is employed to compute a group-level optimal transport plan that can capture the inter-atlas correspondence. Finally, the latent features generated by ViTAE are remapped individually across different atlases under the constraints defined by the group-level optimal transport plan precomputed above. Experimentally evaluated on 583 subjects from CamCAN dataset, the proposed ViTAE-HGOT framework outperforms the state-of-the-art methods in correlation coefficient (CC = 0.5471 ± 0.0498; minimum 13.3% increase) and mean absolute error (MAE = 0.1153 ± 0.0108; minimum 21.2% reduction), and generalizes well in the independent ICBM dataset (CC = 0.5186 ± 0.0053). The remapped FC achieves clinical utility comparable to that of real data in downstream brain age prediction analyses. By converting heterogeneity into an analyzable resource, this framework unlocks legacy datasets and enables standardized multi-center connectomics with minimal data sharing.