Computational Clinical Validation of Stage-conditioned Synthetic 3D FDG-PET for Alzheimer’s Disease Using Region-wise Metabolic Consistency

Purpose Alzheimer’s disease (AD) is a progressive neurodegenerative disorder for which early diagnosis and longitudinal monitoring face two complementary constraints: ethical and practical limits on repeated Fluorodeoxyglucose Positron Emission Tomography (FDG-PET) acquisition (radiation burden, cost, scanner availability), and structural sparsity of clinically critical subgroups (e.g., Mild Cognitive Impairment (MCI) to AD converters and fully sampled longitudinal trajectories) within existing repositories. While generative models have shown promise in addressing these targeted gaps, their evaluation has largely relied on image-level similarity metrics that provide limited insight into clinical plausibility. We present a computational clinical validation framework for synthetic FDG-PET acquisitions and demonstrate its application using PET-ADESyn, a 3D conditional generative adversarial network. Methods The model is trained on 504 volumetric FDG-PET scans from the Alzheimer’s Database Neuroimaging Initiative (ADNI) dataset to generate anatomically coherent PET volumes across the AD continuum (Normal Cognition (NC), MCI, AD). Beyond standard image fidelity metrics, namely Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), synthetic scans are evaluated through a region-wise metabolic consistency analysis based on automated FDG uptake anomaly detection, macro-region aggregation, and Bland-Altman agreement with real acquisitions. Results Results show that synthetic PET volumes preserve known disease-specific hypometabolic patterns and exhibit close agreement with real data at both regional and aggregated anatomical levels. Conclusion This study demonstrates that clinically grounded validation is essential for assessing the utility of generative PET models and provides a clear evaluation strategy aligned with clinical neuroimaging workflows.