Background <p>The early diagnosis of Alzheimer’s disease (AD) plays a crucial role in managing the disease and improving patient outcomes. One of the challenges in detecting early AD is distinguishing patients with progressive mild cognitive impairment, which will further develop AD, from those with stable MCI. Advanced imaging analysis techniques, such as deep learning (DL) and radiomics applied to fluorodeoxyglucose positron emission tomography ([¹⁸F]FDG-PET), provide promising tools for enhancing this early detection.</p> Methods <p>In this systematic review and meta-analysis, we identified studies that utilized DL or radiomics on [¹⁸F]FDG-PET imaging to predict MCI-to-AD conversion and conducted a meta-analysis to assess the diagnostic performance of this method. Comprehensive searches were performed across four databases, including PubMed, Web of Science, Embase, and Scopus. Data extraction, quality assessment (using METRICS), and meta-analyses were conducted on eligible studies following PRISMA guidelines.</p> Results <p>Twenty-one studies met the inclusion criteria, and 17 were included in the meta-analysis. The pooled area under the ROC curve (AUC) was 0.90 (95% CI: 0.87–0.92), with sensitivity and specificity of 78% and 87%, respectively. Hand-crafted radiomics (HCR) performed slightly better than DL models, with a sensitivity of 86% and 75%, respectively. PET-based models outperformed multimodal models in sensitivity and AUC. Saliency map analyses identified key regions for predicting AD conversion, including the posterior cingulate cortex, precuneus, and lateral temporoparietal cortices.</p> Conclusions <p>DL and radiomics models using [¹⁸F]FDG-PET imaging demonstrate high diagnostic performance for predicting MCI-to-AD conversion. Despite strong results, challenges such as methodological heterogeneity, limited external validation, and lack of transparency exist, which need further investigations.</p>