Topology-aware hybrid graph-transformer network for Alzheimer's disease diagnosis from structural magnetic resonance imaging.

Introduction: Alzheimer's disease (AD) is a progressive neurodegenerative disorder characterized by localized cortical atrophy and large-scale disruption of brain connectivity. Although deep learning (DL) methods have shown promise for neuroimaging-based diagnosis, many approaches fail to jointly capture localized structural changes and global network-level degeneration. Methods: We propose a topology-aware hybrid DL framework for AD classification from structural MRI. The model integrates (1) a 3D convolutional neural network (CNN) to extract volumetric morphometric features, (2) a dynamic graph attention network (GAT) to infer patient-specific structural connectivity without predefined atlases, and (3) a topology-biased Vision Transformer (Topo-ViT) that incorporates this connectivity into global attention. The framework is trained under strict subject-level data segregation and optimized using focal loss with an AUC-driven strategy. Results: Evaluated on a structural MRI dataset derived from the OASIS cohort, the proposed model achieved a test ROC-AUC of 0.857, with an overall accuracy of 85% and high sensitivity in detecting demented cases. Ablation studies show that topology-guided attention improves performance over CNN and hybrid baselines. Additional analyses reveal stable connectivity patterns and well-separated latent representations. Discussion: The results demonstrate that integrating topology-aware mechanisms enables more coherent modeling of AD as a network-level disorder. The proposed framework captures both local and global structural patterns, offering improved diagnostic reliability. Further validation on larger datasets is required for clinical deployment.