Structural-Guided Latent Diffusion for MRI-to-PET Synthesis and Alzheimer’s Disease Diagnosis

Abstract

Multimodal neuroimaging is crucial for Alzheimer’s disease (AD) diagnosis, yet positron emission tomography (PET) is often unavailable due to its cost and limited accessibility. We propose a structural-guided latent diffusion framework for synthesizing PET from structural MRI while jointly learning disease-discriminative representations. A pretrained 3D variational autoencoder (VAE) encodes PET into a compact latent space, where conditional diffusion is guided by unified gray- and white-matter structural priors to enable structural-aware generation. To enhance sampling stability and distribution consistency, diffusion with spherical Gaussian (DSG) constraint is introduced. In addition, a task-aware diagnosis branch reuses diffusion features for AD classification, enabling joint generative and diagnostic optimization. Experiments on three public datasets demonstrate that model achieves superior synthesis quality and diagnosis performance using only MRI at inference. The proposed framework provides an effective solution for multimodal data completion and highlights the clinical potential of structural-guided diffusion models.