Delta-Diffusion AI tracks Alzheimer's amyloid buildup from a single PET scan

Longitudinal PET imaging is the gold standard for tracking Beta-amyloid accumulation in Alzheimer's, but high costs and cumulative radiation limit its clinical use. Existing deep generative models for longitudinal synthesis suffer from identity drift and bias toward replicating baseline signals, failing to model true pathological transitions. To address this, researchers Sun, Yu, Wu, Kohi, and Liu from the University of North Carolina propose Delta-Diffusion, a progression-aware framework that reframes longitudinal PET synthesis as a conditional Poisson Diffusion Bridge (PDB) process. Unlike standard diffusion models that start from Gaussian noise, the bridge is mathematically anchored to the subject's baseline PET scan, transforming the task into a conditional distribution transition of the amyloid trajectory. A physically-grounded Poisson perturbation is integrated within a Diffusion Transformer (DiT), using adaptive scale-shift modulation to calibrate synthesis based on elapsed clinical interval and structural MRI context.

Delta-Diffusion also introduces a volume-of-interest balanced objective that emphasizes sparse, high-risk regions of amyloid accumulation, improving sensitivity to early pathological changes. The model was validated on two independent cohorts totaling 542 subjects, outperforming state-of-the-art methods in capturing longitudinal variations of amyloid deposition. This robust computational framework could enable synthetic follow-up PET scans from baseline data alone, cutting both cost and radiation exposure. It holds significant potential for clinical trials, early diagnosis, and research into disease-modifying therapies for Alzheimer's.