Forecasting Alzheimer's disease progression via identity-preserved denoising diffusion generative adversarial network

Forecasting the progression of Alzheimer's disease (AD) is essential for evaluating secondary prevention measures thought to modify the disease trajectory. However, accurate prediction of longitudinal MRIs remains challenging, particularly in preserving subject identity, as deep generative models may potentially generate plausible future MRIs of different individuals from a single baseline scan. In the present study, we developed a novel identity-preserved denoising diffusion generative adversarial network (IP-DDGAN) capable of rapidly generating subject-specific longitudinal MRIs conditioned on metadata. Concretely, we developed an identity-preservation strategy incorporating a metadata-guided module and identity-preserved regularization terms to maintain subject identity in synthetic longitudinal MRIs. Furthermore, we comprehensively integrated morphometric, subject-identity-consistency, and image-level quality metrics to evaluate the fidelity and biological plausibility of synthetic longitudinal MRIs. The results demonstrate that the synthetic MRIs generated by IP-DDGAN retain biological and disease-related phenotypes and exhibit sufficient realism to support downstream applications. Our proposed model effectively captures temporal biological and disease-related changes and predicts distinct disease progression trajectories, including the clinically important transitions from cognitively normal (CN) to mild cognitive impairment (MCI) and from MCI to AD.