Background: Early prediction of conversion from mild cognitive impairment (MCI) to Alzheimer's disease (AD) is important for initiating timely interventions and selecting candidates for disease-modifying therapies.Structural MRI (sMRI) provides non-invasive biomarkers of neurodegeneration, and deep learning (DL) models have recently shown promise for automated risk stratification.However, performance across architectures, data inputs, and validation strategies remains highly variable.Methods: This systematic review and meta-analysis, conducted according to PRISMA 2020 guidelines, evaluated DL models trained on sMRI to predict conversion from MCI to AD within a 36-month follow-up.Five databases (PubMed, Embase, Scopus, and IEEE Xplore) were searched up to May 2025.Studies reporting sensitivity, specificity, or AUC were included.Quality was assessed using the METRICS tool.Pooled diagnostic performance was estimated using a bivariate random-effects model, and subgroup analyses examined the impact of architecture, dimensionality, training strategy, and input type.Results: Forty-two studies met inclusion criteria.Most employed convolutional neural networks (CNNs), while a growing subset used transformer-based or hybrid models.Pooled sensitivity and specificity were 0.76 (95% CI 0.72-0.79)and 0.79 (95% CI 0.76-0.82).No significant performance difference was found between 2D vs 3D models, regional vs whole-brain inputs, or transformers vs. CNNs.Studies using transfer learning showed marginally improved accuracy and lower heterogeneity compared with those trained from scratch.Conclusions: Deep learning models trained on structural MRI demonstrate promising diagnostic performance in predicting MCI-to-AD conversion under research conditions.However, the heavy reliance on ADNI-derived datasets and limited external validation across most studies preclude definitive conclusions about clinical utility.Transformer-based and transfer-learning approaches may offer incremental advantages, but methodological heterogeneity and insufficient real-world validation underscore the need for standardized frameworks, prospective multi-site studies, and transparent reporting before clinical translation can be considered.
Kiani et al. (Wed,) studied this question.