Language-enhanced generative modeling for amyloid PET synthesis from MRI and blood biomarkers

Assessment of amyloid pathology in Alzheimer's disease (AD) often relies on amyloid-beta positron emission tomography (Aβ-PET), but its clinical use is limited by cost and accessibility. We developed a language-enhanced generative framework to synthesize Aβ-PET images from T1-weighted magnetic resonance imaging (MRI) and blood biomarkers in a cohort of 566 participants. The synthetic PET images resembled real PET scans in both structural detail (structural similarity index [SSIM] = 0.920 ± 0.003) and regional uptake patterns (Pearson's R = 0.955 ± 0.007). In physician evaluation, diagnoses based on synthetic PET showed high agreement with those based on real PET (accuracy = 0.80). In addition, models using synthetic PET improved Aβ positivity classification performance compared with models based on MRI or blood biomarkers alone. These findings show that the framework can generate clinically informative PET-like images and may support resource-limited amyloid assessment workflows.