Diffusion bridge networks simulate clinical-grade PET from MRI for dementia diagnostics: a retrospective cohort study

Background Fluorodeoxyglucose-PET (FDG-PET) outperforms structural MRI in early and differential dementia diagnostics but is challenged by scanner availability, high cost, and radiation exposure. We aimed to develop and clinically validate a generative artificial intelligence (AI) framework that simulates clinical-grade FDG-PET from routine MRI and auxiliary patient information to broaden access to metabolic imaging for dementia diagnostics. Methods In this retrospective cohort diagnostic accuracy study, we developed SiM2P, a conditional three dimensional diffusion bridge model with diffusion Transformers to simulate FDG-PET from T1-weighted MRI and 13 auxiliary clinical variables (age, sex, education, mini-mental state examination, Alzheimer's disease assessment scale–cognitive subscale 13, apolipoprotein E ε4 status, and seven MRI-derived segmentation volumes). Eligible participants had baseline paired T1-weighted MRI and FDG-PET scans with accompanying clinical records, each participant contributed one baseline scan, and no longitudinal follow-up scans were included. No further exclusion criteria were applied. The framework was trained and evaluated using data from three independent cohorts: the Alzheimer's Disease Neuroimaging Initiative (ADNI), the Japanese ADNI (J-ADNI), and an in-house Technical University of Munich University Hospital cohort. No age limit was imposed on participants and their cognitive continuum spanned from cognitively normal to mild cognitive impairment, Alzheimer's disease, and frontotemporal lobar degeneration. The primary endpoint was diagnostic accuracy in a masked clinical reader study in which two neuroradiologists independently interpreted MRI scans and two nuclear medicine physicians independently interpreted both simulated and real FDG-PET scans from a held-out in-house test set of individuals (who were cognitively normal or had Alzheimer's disease or behavioural-variant frontotemporal dementia) for dementia diagnosis and three-class differential diagnosis. Secondary outcomes included image quality metrics against four competing methods and automated three-class differential diagnosis. Findings We analysed paired MRI and FDG-PET scans from 1889 participants (n=1247 for ADNI, n=319 for J-ADNI, and n=323 for in-house dataset). The participants were aged 41–91 years. The scans were acquired between Sept 22, 2005, and Oct 14, 2019 (ADNI); Sept 3, 2008, and May 17, 2012, (J-ADNI); and May 19, 2011, and Dec 27, 2018 (Technical University of Munich University hospital data). In the masked clinical reader study including a held-out in-house test set of 62 individuals (comprising 22 cognitively normal individuals, 19 with Alzheimer's disease, and 21 with behavioural-variant frontotemporal dementia), simulated PET yielded an accuracy of 93% (95% CI 86·8–96·1) for dementia detection versus 83% for MRI (75·5–88·7; p=0·0076) and 85% (77.3–90·0) versus 75% (66·7–81·8) for three-class differential diagnosis of participants who were cognitively normal, had Alzheimer's disease, and had behavioural-variant frontotemporal dementia (p=0·044). Inter-rater reliability was substantially higher for simulated PET (Cohen's κ =0·899) than for MRI ( κ =0·644). Simulated PET also improved diagnostic confidence. A crossreading analysis further confirmed that the observed gain was attributed to the sensitive disease-relevant information provided by the simulated PET rather than reader-specific expertise. Interpretation AI-simulated PET generated from routine MRI can improve the diagnostic accuracy and inter-rater reliability of structural MRI for dementia detection and differential diagnosis. The data-efficient Local-Adapt workflow provides a practical pathway for deployment at clinical centres without PET infrastructure, thus potentially widening equitable access to metabolic PET diagnostics for dementia care. Funding German Research Foundation (DFG), Munich Center for Machine Learning, and the German Academic Exchange Service (DAAD) Konrad Zuse Schools of Excellence in Artificial Intelligence.