Generation of Synthetic [18F]FDG PET From Early-Phase Amyloid PET in Alzheimer's Disease

This is a retrospective observational study conducted at IRCCS Ospedale San Raffaele, Milan, Italy. The study evaluates the accuracy of synthetic \[18F\]FDG PET images generated using a SwinUNETR deep learning model compared to native \[18F\]FDG PET images. Study Population: Adults (≥ 50 years) who underwent amyloid PET imaging (using Florbetaben or Flutemetamol), structural MRI, and \[18F\]FDG PET due to cognitive symptoms between January 2025 and December 2025. Approximately 35 patients meeting inclusion criteria will be included. Methodology: All imaging and clinical data were collected as part of routine diagnostic care; thus, no additional procedures, interventions, or interactions with patients are required for this study. All data are fully deidentified before analysis, consistent with GDPR and institutional data protection policy. The SwinUNETR model processes volumetric images to generate synthetic FDG PET images from early-phase amyloid PET and MRI inputs. Objectives and Endpoints: Primary Objective: To quantitatively and qualitatively assess the accuracy of synthetic FDG PET images compared with native FDG PET images. Primary Endpoint: Pearson correlation coefficient and mean absolute error (MAE) of SUVR values obtained from native FDG PET and synthetic FDG PET in Alzheimer relevant areas of interest (precuneus, posterior cingulate, lateral temporal cortex, and frontal cortex). Secondary Objective: To assess visual interpretability and clinical intuitiveness of synthetic FDG PET images by expert nuclear medicine physicians. Secondary Endpoint: Inter-rater agreement (Cohen's kappa) among 2 blinded nuclear medicine physicians rating synthetic FDG scans as "clinically acceptable" or not. Ethical Considerations: Due to the retrospective and non-interventional nature of this study, no additional informed consent is required. A waiver of informed consent will be requested from the Ethics Committee. The image data will be fully anonymized in accordance with institutional policies.