Synthesis of Amyloid Images Using a Generative Adversarial Network from 2-Dimensional 18 F-FDG Images and Evaluation for Clinical Use

The use of amyloid PET to assess patient suitability of disease-modifying drugs for Alzheimer disease is increasing. This study aimed to synthesize amyloid PET images from 18F-FDG PET images using a generative artificial intelligence algorithm to reduce unnecessary amyloid PET scans. Methods: A 2-dimensional pix2pix algorithm was used. The algorithm was evaluated across 4 domains: image quality, voxel values, contrast between white and gray matter, and diagnostic performance for detecting the presence or absence of β-amyloid (Aβ) deposition. Pairs of 18F-FDG PET and amyloid PET images from 55 Aβ-negative and -positive cases were evaluated. A 6-fold cross-validation was conducted. Results: Synthetic images were visually consistent, producing plausible negative and positive patterns while preserving continuity in the sagittal plane. Voxel values of the synthetic images showed a significant linear relationship with the real images. The contrast correlated well with the real images, and the differences between the negative and positive cases were significant as well as those in the real images. The performance of the positive or negative 2-class classifier exceeded 85% for the synthetic images. Conclusion: The synthetic images successfully captured features of Aβ deposition, and evaluation with a 2-class classifier achieved an acceptable accuracy of 85%. These results suggest that amyloid images can potentially be generated from 18F-FDG PET images for use in clinical practice.