Xavier Beltran-Urbano, Katie R Jobson, Ilya M Nasrallah, Ajay Kumar, Manuel Taso, Sandhitsu R Das, Christopher A Brown, Sipei Li, Long Xie, Corey T McMillan, Paul A Yushkevich, Dave A Wolk, Greg Zaharchuk, Sudipto Dolui, John A Detre, Alzheimer's Disease Neuroimaging Initiative
Both models achieved high fidelity with real FDG PET (structural similarity > 93%) and strong diagnostic performance in AD-specific regions, with ASL providing modest improvements. Results generalized to an independent AD external cohort.
INTRODUCTION: Fluorodeoxyglucose (FDG) positron emission tomography (PET) is widely used for detecting metabolic changes associated with neurodegeneration in Alzheimer's disease (AD) and other dementias but is costly and involves ionizing radiation. Here we developed SynthPET, a multimodal deep learning framework that synthesizes FDG PET images from T1-weighted (T1w) and arterial spin labeling (ASL) perfusion magnetic resonance imaging (MRI) data.
METHODS: Two conditional generative models were trained on paired MRI and FDG PET acquisitions: one using T1w alone (n = 1170) and one integrating T1w with ASL via cross-modality transfer learning (n = 220). We evaluated synthetic FDG PET through voxel-wise fidelity metrics, regional correlation and receiver operating characteristic analyses, blinded reader studies, and differential diagnosis across independent external cohorts.
RESULTS: Both models achieved high fidelity with real FDG PET (structural similarity > 93%) and strong diagnostic performance in AD-specific regions, with ASL providing modest improvements. Results generalized to an independent AD external cohort.
DISCUSSION: SynthPET generates clinically plausible FDG PET from routinely acquired MRI, potentially expanding access to metabolic neuroimaging where PET is unavailable.