AI-assisted PET radiomics improves Alzheimer's diagnosis over conventional methods in meta-analysis

ObjectiveAI-radiomics can analyze radiological images more thoroughly and quickly than the human eye. This study aims to compare the diagnostic efficacy of AI-assisted PET radiomics for Alzheimer’s disease (AD) with conventional PET diagnosis through a systematic review and bivariate meta-analysis performing indirect, study-level benchmarking versus conventional PET.MethodsPubMed, Embase, and Web of Science were searched through April 11, 2025, for human diagnostic accuracy studies for AI-assisted PET radiomics. Two reviewers extracted data per PRISMA guidelines, risk and bias were appraised using QUADAS-AI. Effect sizes were synthesized via a bivariate random-effects model with HSROC. Prespecified strata contrasted with AD vs. healthy controls (HC), AD vs. mild cognitive impairment (MCI), and tracer class. The analyses were conducted based on bivariate random-effects model realized using R and Stata.ResultsNine studies (25 2 × 2 tables; n = 5,765) were included. A strong correlation between sensitivity and specificity signaled substantial study heterogeneity. This heterogeneity was further illustrated by the dispersion of the HSROC prediction region. In AD vs. HC, proteinopathy PET yielded SE 0.89, SP 0.91, and AUC 0.96. In comparison, the 18F-FDG PET demonstrated near-parity (SE 0.92, SP 0.92 AUC 0.94), suggesting limited incremental value. In AD vs. MCI, current data suggested a trend toward improved performance with proteinopathy PET relative to 18F-FDG PET (SE 0.94, SP 0.95, AUC 0.96 vs. AUC 0.84). These results underscore the potential of proteinopathy PET in facilitating early diagnostic evaluations, necessitating further validation. In contrast to conventional benchmarks, the AD vs. MCI demonstrated notably higher diagnostic metrics (AUC 0.96; LR + 19.64; LR − 0.06; conventional amyloid-PET specificity approximately 0.49), while the gains in AD vs. HC were negligible (ΔAUC +0.02). Sensitivity analyses confirmed that primary estimates were not influenced by a single study.ConclusionAI-radiomics on proteinopathy PET shows promising potential for distinguishing AD from MCI, yet only marginal benefits comparing AD to HC. However, given the heterogeneity of the data, the risk of bias, and the limited external validation, there is a need to prioritize multi-site validation, standardized reporting, and prospective decision-impact studies.Systematic review registrationhttps://www.crd.york.ac.uk/PROSPERO/view/CRD420251029823, identifier, PROSPERO (CRD420251029823).