Diffusion Bridge Networks Simulate Clinical-grade PET from MRI for Dementia Diagnostics

TL;DR

This study presents SiM2P, a 3D diffusion-bridge framework that translates routine brain MRI and auxiliary patient data into high-fidelity simulated FDG-PET images for dementia diagnostics. By conditioning a volumetric diffusion Transformer on MRI and clinical priors, SiM2P captures disease-specific hypometabolic patterns and delivers PET-like diagnostic signals without actual PET scans. In a blinded clinical reader study, simulated PET improved diagnostic accuracy and interrater reliability over MRI alone, and a Local-Adapt workflow enables site-specific deployment using as few as 20 local cases. The approach reduces cost and radiation exposure while preserving PET's diagnostic value, with robust performance even under limited data and potential applicability to other tracers and neurodegenerative disorders.

Abstract

Positron emission tomography (PET) with 18F-Fluorodeoxyglucose (FDG) is an established tool in the diagnostic workup of patients with suspected dementing disorders. However, compared to the routinely available magnetic resonance imaging (MRI), FDG-PET remains significantly less accessible and substantially more expensive. Here, we present SiM2P, a 3D diffusion bridge-based framework that learns a probabilistic mapping from MRI and auxiliary patient information to simulate FDG-PET images of diagnostic quality. In a blinded clinical reader study, two neuroradiologists and two nuclear medicine physicians rated the original MRI and SiM2P-simulated PET images of patients with Alzheimer's disease, behavioral-variant frontotemporal dementia, and cognitively healthy controls. SiM2P significantly improved the overall diagnostic accuracy of differentiating between three groups from 75.0% to 84.7% (p<0.05). Notably, the simulated PET images received higher diagnostic certainty ratings and achieved superior interrater agreement compared to the MRI images. Finally, we developed a practical workflow for local deployment of the SiM2P framework. It requires as few as 20 site-specific cases and only basic demographic information. This approach makes the established diagnostic benefits of FDG-PET imaging more accessible to patients with suspected dementing disorders, potentially improving early detection and differential diagnosis in resource-limited settings. Our code is available at https://github.com/Yiiitong/SiM2P.

Figures (12)

• Figure 1: Overall study design, model pipeline, and evaluation. a, Clinical context. Clinical diagnostic workup typically involves the structural MRI for assessment of cerebral atrophy, cerebrovascular disease, and exclusion of secondary causes such as tumors. If the diagnosis remains unclear, FDG-PET can be performed in specialized centers. However, the routine use of PET is limited by scanner availability, high costs, and radiation exposure. Our goal is to develop an AI-supported workflow that simulates FDG-PET from routine MRI, enabling PET-informed decision support in settings where PET is unavailable. b, Our model employs a 3D diffusion bridge to simulate PET from structural MRI, conditioned on available auxiliary data such as demographics and MRI-derived segmentation volumes. We validated the diagnostic utility of our simulated PET (SimPET) in a blinded clinical reader study, where SimPET showed a higher accuracy than MRI. c, SiM2P-simulated PET closely resembled disease-specific hypometabolism patterns observed in real FDG-PET and substantially outperformed the biomarker magnitude in MRI.
• Figure 1: Comparison of simulated PET from SiM2P and real PET. For each subject (an Alzheimer's disease patient, a behavioral-variant frontotemporal dementia patient, and a normal control), we show eight evenly distributed axial slices of the PET scan. Below these slices, the corresponding 3D-SSP maps are displayed, where the first row shows a direct surface projection from different directions, and the second row provides a quantitative measure as a globally-normalized negative Z-score map.
• Figure 2: Clinical reader study pipeline and results. a, The study involves a two-stage diagnostic workflow for both neuroradiologists (for MRI) and nuclear medicine physicians (for simulated PET, SimPET in short). We report the diagnostic accuracy alongside interrater reliability (using Cohen's $\kappa$) for three tasks: dementia disorders diagnosis, differential diagnosis of dementia disorders for either AD-versus-bvFTD or CN-versus-AD-versus-bvFTD, with error bars representing within-rater standard deviation and 95% CI, respectively. Significance levels using McNemar's test are denoted on top as *$P<$ 0.05. b, Study population and diagnostic accuracy weighted by each rater’s confidence level for MRI and SimPET across three tasks. Error bars indicate the 95% CI. The percent increase in mean performance gained by SimPET is indicated alongside each task. Significance levels using the Wilcoxon signed-rank test are denoted on top as *$P<$ 0.05, **$P<$ 0.01, ***$P<$ 0.001. c, Confusion matrices of all raters for MRI (left) and SimPET (right) across three labels.
• Figure 2: Diagnostic performance of MRI (MRI-R1, MRI-R2) and simulated PET raters (PET-R1, PET-R2). We demonstrate sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) for MRI and simulated PET raters R1/R2. The results are broken down by individual diagnostic categories and also include an overall average performance.
• Figure 3: Representative success and failure cases of SiM2P in the clinical reader study. a, Two success cases in which subtle MRI atrophy patterns led to incorrect diagnosis as healthy control subjects, whereas simulated PET (SimPET) reproduced the temporoparietal hypometabolism seen on the real PET, enabling correct diagnosis of Alzheimer’s disease. b, Two failure cases in which overlapping metabolic patterns of frontal-variant AD and behavioral-variant FTD resulted in misdiagnosis of AD, despite SimPET closely matching the frontal-lobe hypometabolism observed on real PET. Abnormal regions are highlighted with red boxes on the middle slice.