Structure as Prior: Heat Kernel Diffusion for Anatomically Constrained Brain Functional Network Generation 3 views

Early prediction of neurological disorders can support diagnosis and treatment planning, but multimodal neuroimaging models still struggle to use structural connectivity (SC) and functional connectivity (FC) in a biologically meaningful way. Many deep learning methods fuse the two modalities only after independent feature extraction, which weakens the link between white-matter pathways and functional signal propagation. This thesis introduces SFDiff (Structure-guided Functional Diffusion), a generative model for brain functional networks that incorporates anatomical structure directly into the diffusion process. SFDiff replaces the isotropic Gaussian forward process used in standard diffusion models with a heat diffusion process driven by the SC Laplacian. An SC-conditioned denoiser then uses cross-attention to combine structural and functional embeddings while reversing this process. For downstream diagnosis, the trained SC encoder is frozen and its node embeddings are mean-pooled for latent classification with an MLP head. Experiments on the ABCD and PPMI datasets show that SFDiff is competitive with generic graph learning methods and recent brain-network architectures. In particular, SFDiff achieves the highest PPMI AUC among the compared methods, while remaining close to the strongest ABCD baseline. These results suggest that structural priors can improve discriminative feature learning, especially when SC--FC coupling is relatively consistent across subjects.

MS (Master of Science)

Brain; Diffusion model; Multimodal

English

All rights reserved by the author (no additional license for public reuse)

2026-04-15