Scoping the Landscape of Deep Learning for Alzheimer's Disease Stage Classification: Methods, Challenges, and...

Deep learning (DL) models have been widely applied for Alzheimer's disease (AD) stage classification. This scoping review synthesizes recent research to evaluate current performance benchmarks, identify methodological limitations, and highlight translational barriers. DL has potential to augment diagnostic accuracy and accelerate early intervention in AD, but translation requires models that generalize across datasets and integrate into real-world clinical workflows. Following scoping review methodology, 18 peer-reviewed studies published between 2018 and 2024 were analyzed. We extracted dataset sources, preprocessing strategies, model architectures, performance metrics, and translational considerations. Most studies employed convolutional neural networks (CNNs) or transfer learning (TL) backbones with accuracies frequently reported above 90%. Comparative synthesis revealed that TL and custom CNNs achieved similar headline accuracies, with differences of less than one percentage point. Reported performance was highly sensitive to task framing (cross-sectional vs. progression) and dataset provenance, with curated subsets often yielding near-ceiling internal accuracies but limited generalizability. Only one study implemented true external validation, underscoring a critical translational gap. Cost-effectiveness was rarely discussed explicitly; however, several studies indicated that open datasets reduce financial barriers, while adapting pipelines for EMR, or multisite data entails substantial resource demands. DL for AD classification shows consistent high accuracy but limited robustness, with external validation and financial cost-effectiveness remaining underreported. Future progress depends on standardized evaluation protocols, explicit reporting of financial costs, and the development of clinically interpretable, workflow-integrated models.