Multimodal MRI-to-PET image translation recovers disease-specific hypometabolism in epilepsy and mild cognitive impairment

nd: Glucose-metabolism imaging with positron emission tomography (FDG-PET) helps localize the seizure focus in epilepsy and diagnose Alzheimer’s disease and mild cognitive impairment, but its radioactive tracer limits repeat scanning. Arterial spin labeling, a magnetic resonance (MR) method, measures blood flow without radiation but detects these changes less reliably. We aimed to generate metabolic-like images from magnetic resonance imaging to address these limitations. From paired metabolic and magnetic resonance scans of patients with temporal lobe epilepsy (n = 68; mean age 37 ± 12 years), people with mild cognitive impairment and controls (total n = 85: 53 with mild cognitive impairment, 32 cognitively unimpaired; aged 50–85 years), we developed FlowGAN, a deep learning model that turns arterial spin labeling and structural MR images into metabolic-like images. Using cross-validation and a held-out test set, we tested how closely synthetic images matched real ones in image quality, regional values, and distinguishing patients from controls. Here we show that synthetic FDG-PET images resemble real metabolic images, showing higher structural similarity, signal-to-noise ratio, and cross-correlation than ASL blood-flow images (all p < 0.001). Region by region, synthetic FDG-PET tracks real FDG-PET more closely than ASL blood-flow, for asymmetries in epilepsy (31 of 36 regions) and uptake in mild cognitive impairment (33 of 36 regions). Synthetic and real images agree on which regions best separate mild cognitive impairment from controls (correlation 0.91, p < 0.001), but agree less in lateralizing temporal lobe epilepsy (0.28, p = 0.094), though they outperform ASL in 39–55% of regions. FlowGAN turns radiation-free MR scans into metabolic images that reproduce the reduced-metabolism patterns of epilepsy and mild cognitive impairment and could be repeated when a tracer scan is impractical. Physicians often use a brain scan, called an FDG-PET scan, that shows how the brain uses glucose. FDG-PET can help find where epileptic seizures start, and also diagnose memory diseases such as Alzheimer’s disease. The scan needs an injection of radioactive glucose, so it cannot be repeated often. We built an artificial intelligence model called FlowGAN that creates PET-like images from an ordinary magnetic resonance (MRI) scan, which uses no radiation. We tested FlowGAN in patients with epilepsy and in people with cognitive impairment. The images generated looked like real PET scans and showed the same disease patterns, especially in memory disease. With this approach, a radiation-free, repeatable MRI scan could therefore show much of what a PET scan shows, helping more patients. Lucas, Vadali, et al. develop a deep learning framework for synthesizing glucose-metabolism imaging with positron emission tomography (FDG-PET) images from MRI as a radiation-free alternative. They tested their results in epilepsy and mild cognitive impairment, finding that synthetic images approximate real FDG-PET in these neurological disorders.