Multimodal Diffusion Model for Synthetic Alzheimer’s Disease Data Generation Using MRI, PET, and Clinical Biomarkers

Resumo

The scarcity, imbalance, and privacy constraints of high-quality multimodal datasets hinder progress in machine-learning for Alzheimer’s disease (AD). Structural MRI, PET imaging and clinical biomarkers each capture complementary aspects of AD pathophysiology, but multimodal collections that include all three modalities are limited. To address data scarcity and enable privacy-preserving research, we propose a Multimodal Diffusion Generative Model (MDGM) that jointly synthesizes high-resolution structural MRI, PET, and tabular clinical biomarkers. MDGM uses modality-specific encoders, a latent-space fusion mechanism, and a conditional score-based diffusion model operating in a compact latent space; the decoder yields harmonized MRI, PET volumes and clinical vectors consistent with a conditioning disease profile. We train and evaluate MDGM on the Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohort, quantify image fidelity and a domain-adapted FID, assess clinical realism by correlation of biomarkers and cognitive scores, and test downstream utility by training multimodal diagnostic classifiers with and without synthetic augmentation. Our results show that MDGM generates clinically plausible multimodal samples that preserve disease patterns (e.g., hippocampal atrophy, elevated amyloid signal) and substantially improve classifier sensitivity for under-represented early-disease classes when used for augmentation. MDGM demonstrates how diffusion-based generative models can produce multimodal synthetic cohorts to accelerate AD research while respecting patient privacy.