Exploring PET and MRI brain networks across healthy controls, mild cognitive impairment, and Alzheimer’s disease

This study aimed to characterize population-level and individual-level metabolic and radiomics-derived structural brain networks across healthy controls (HC), mild cognitive impairment (MCI), and Alzheimer’s disease (AD) using multimodal PET and MRI. A total of 651 participants with both [ 18 F]FDG PET and MRI were retrospectively included, with an external cohort of 45 participants used for validation. Thirty-two AD-relevant regions of interest were retained as network nodes and characterized using PET standardized uptake value ratios (SUVRs) and a selected MRI radiomics descriptor. Population-level networks were constructed using probability density function-based similarity with Jensen–Shannon divergence, and individualized perturbation networks were derived relative to an HC reference. Network-derived features were evaluated for prediction. Both SUVR-based and radiomics-based networks showed distinct patterns in MCI and AD compared with HC. Fused network features improved discrimination compared with either modality alone, achieving one-vs-rest AUCs of 0.894 [0.868–0.917] for HC, 0.888 [0.862–0.912] for AD, and 0.814 [0.783–0.844] for MCI. These findings indicate graded network disorganization across the HC–MCI–AD spectrum and support multimodal prediction.