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

Area of Science:

• Neuroimaging Analysis
• Computational Neuroscience
• Data Fusion

Background:

• Open-source neuroimaging datasets offer large sample sizes but suffer from cross-study comparability issues.
• Inconsistent functional connectome (FC) data arises from the use of different brain atlases, hindering multi-site data fusion.
• Existing FC remapping methods often neglect functional semantics, compromising analytical validity.

Purpose of the Study:

• To propose a novel framework, Vision Transformer-based Autoencoder with Hypergraph Optimal Transport (ViTAE-HGOT), for cross-atlas FC remapping.
• To preserve functional semantics during FC remapping to improve cross-study comparability and data fusion.
• To enable standardized multi-center connectomics and unlock the potential of legacy neuroimaging datasets.

Main Methods:

• The ViTAE-HGOT framework utilizes a Vision Transformer-based Autoencoder (ViTAE) to derive brain region representation features in a latent space.
• Hypergraph Optimal Transport (HGOT), guided by Yeo-7 functional semantics, computes a group-level optimal transport plan for inter-atlas correspondence.
• Latent features are remapped across atlases using the computed group-level optimal transport plan.

Main Results:

• ViTAE-HGOT demonstrated superior performance on the CamCAN dataset, outperforming state-of-the-art methods in correlation coefficient (CC) and mean absolute error (MAE).
• The framework showed strong generalization on the independent ICBM dataset.
• Remapped FC data achieved clinical utility comparable to real data in brain age prediction analyses.

Conclusions:

• ViTAE-HGOT effectively preserves functional semantics during cross-atlas FC remapping, addressing a critical limitation in current methods.
• The framework enhances the analytical validity of multi-site and cross-study neuroimaging analyses.
• ViTAE-HGOT facilitates standardized multi-center connectomics by converting data heterogeneity into an analyzable resource.