Deep Learning-based Synthesis of Amyloid PET Images from Structural MRI in Alzheimer's Disease.
Authors
Affiliations (5)
Affiliations (5)
- Russel Morgan Department of Radiology and Radiological Sciences, School of Medicine, Johns Hopkins University, Baltimore, MD, USA.
- Department of Electrical and Computer Engineering, Whiting School of Engineering, Johns Hopkins University, Baltimore, MD, USA.
- Department of Psychiatry and Behavioral Sciences, School of Medicine, Johns Hopkins University, Baltimore, MD, USA.
- Russel Morgan Department of Radiology and Radiological Sciences, School of Medicine, Johns Hopkins University, Baltimore, MD, USA. [email protected].
- Division of MR Research, Department of Radiology, Johns Hopkins University, 600 N. Wolfe Street, Park 306G, Baltimore, MD, 21287, USA. [email protected].
Abstract
Amyloid-beta (Aβ) plaques in Alzheimer's disease (AD) are commonly imaged with specific PET radiotracers. To overcome the cost and availability limitations of PET, developing a more accessible diagnostic tool is essential. This study aimed to develop a deep-learning model that can synthesize Aβ-PET images from widely available MRI scans. The study utilized 431 subjects with both structural MRI and Aβ-PET scans from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database, including 213 cognitively normal (CN), 180 mild cognitive impairment (MCI), and 38 AD dementia cases. A specialized Vector Quantized Generative Adversarial Network (VQGAN) framework has been trained to generate Aβ-PET images from MRI features, utilizing datasets from patients across various stages of AD spectrum. A paired t-test was used to compare synthetic and real Aβ-PET scans, while an ANOVA was performed to evaluate standardized uptake value ratio differences among the CN, MCI, and AD dementia groups. The synthesized Aβ-PET images closely resemble real Aβ-PET scans in terms of regional Aβ distribution, accurately capturing disease-stage characteristics across the AD spectrum. In three key cortical brain regions (frontal cortex, lateral temporal lobe, and posterior cingulate cortex and precuneus), the synthetic images have successfully replicated disease-related Aβ trends, showing statistically significant group differences (p < 0.05). Quantitative evaluation has confirmed the superiority of VQGAN over other models (lower MAE, higher PSNR/SSIM), demonstrating minimized errors, reduced noise, and better structural fidelity in synthesized images. This technology potentially offers a cost-effective, non-invasive alternative method for AD diagnosis and staging.