MRI-to-PET synthesis via deep learning for amyloid-β quantification in Alzheimer's disease.
Authors
Affiliations (13)
Affiliations (13)
- Department of Radiology and Nuclear Medicine, Xuanwu Hospital, Capital Medical University, Beijing, China.
- Beijing Key Laboratory of Magnetic Resonance Imaging and Brain Informatics, Beijing, China.
- Key Laboratory of Neurodegenerative Diseases, Ministry of Education, Beijing, China.
- Department of Radiology and Nuclear Medicine, Xuanwu Hospital, Capital Medical University, Beijing, China. [email protected].
- Beijing Key Laboratory of Magnetic Resonance Imaging and Brain Informatics, Beijing, China. [email protected].
- Key Laboratory of Neurodegenerative Diseases, Ministry of Education, Beijing, China. [email protected].
- Department of Nuclear Medicine, Nanjing First Hospital, Nanjing Medical University, Nanjing, China.
- Department of Radiology and Nuclear Medicine, The First Hospital of Hebei Medical University, Shijiazhuang, China.
- YIWEI Medical Technology Co., Ltd., ShenZhen, China.
- Department of Neurology, Xuanwu Hospital, Capital Medical University, Beijing, China.
- Department of Radiology and Nuclear Medicine, Xuanwu Hospital, Capital Medical University, Beijing, China. [email protected].
- Beijing Key Laboratory of Magnetic Resonance Imaging and Brain Informatics, Beijing, China. [email protected].
- Key Laboratory of Neurodegenerative Diseases, Ministry of Education, Beijing, China. [email protected].
Abstract
Amyloid-β (Aβ) PET is crucial for diagnosing and monitoring Alzheimer's disease (AD), but its high cost and radiation exposure limit its use. Deep learning techniques make it possible to generate PET from structured MRI data. In this study, we built a deep learning model to generate 3D synthetic Aβ PET images from structural MRI. The generative adversarial network with share parameters (ShareGAN) model was trained and tested with 1009 Aβ PET and paired MRI images from the Alzheimer's Disease Neuroimaging Initiative database and three tertiary hospitals in China. The 3D synthetic model operates on the whole volume rather than 2D image slices, realistically reproducing minor discrepancies between neighboring image planes. ShareGAN-based PET images were evaluated using quantitative metrics and visual assessment. Pearson correlation coefficient and Bland-Altman analyses were used to assess the correlation and concordance between synthetic and real PETs. 3D Synthetic PET images showed high similarity and correlation with real Aβ PET in external testing sets 1 and 2 in terms of structural similarity index measure (0.898, 0.899), peak signal-to-noise ratio (34.690, 34.725), mean absolute error (0.031, 0.031), and standardized uptake value ratio (R = 0.758, 0.828). The diagnostic accuracy of PET positive or negative status in external testing sets 1 and 2 was 88.5% and 89.4%, respectively. MRI-based 3D synthetic Aβ PET images can serve as a safe and cost-effective tool for Aβ status visualization, providing PET-eligible patients with Aβ PET-like imaging analysis to guide subsequent real Aβ PET scans. Question Amyloid-β (Aβ) PET limitations (high cost, radiation, limited access) hinder early Alzheimer's disease (AD) detection. Clinical practice urgently requires a suitable supplementary method for Aβ pathology assessment. Findings AI-synthesized 3D Synthetic Aβ PET from structural MRI demonstrated strong consistency with real PET and effectively triaged high-risk patients for confirmatory scans. Clinical relevance This non-invasive, cost-effective method holds the promise of enabling wider Aβ pathology screening, reduces unnecessary PET scans, and supports early intervention in resource-limited settings, while preserving diagnostic rigor for treatment decisions.