Structural–Functional Connectome Generation via Diffusion-Guided Graph Transformer for Alzheimer’s Disease Analysis

Abstract

Brain networks can reveal abnormal structural and functional connection patterns induced by Alzheimer’s Disease (AD), providing critical insights for early diagnosis and intervention. However, current brain network computation tools can only construct unimodal brain networks, overlooking the complementarity between structural and functional brain connections, and suffer from limitations such as time-consuming processes and subjective parameter settings. To address these issues, we propose a structural-functional connectome generation paradigm called DiffusionBrain and apply it to the diagnosis and analysis of AD. Firstly, a Graph Prompt Fusion Module (GPFM) is designed to extract complementary structural and functional features from multimodal brain images and perform interleaved fusion as conditional prompts. Subsequently, a Dual Diffusion-guided Graph Transformer (DDGT) is developed to efficiently generate structural and functional brain networks from Gaussian noise under the guidance of conditional prompts. Lastly, a Graph Alignment Module (GAM) deeply fuses the structural and functional brain networks to obtain the structural-functional connectome. Experimental results on the ADNI dataset demonstrate that DiffusionBrain can better capture the complex interaction mechanisms of structural and functional brain networks and effectively reveal abnormal connection patterns in AD. More importantly, the proposed DiffusionBrain provides a multimodal brain network modeling paradigm, supporting further exploration of AD biomarkers and pathological mechanisms.