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

Area of Science:

• Artificial Intelligence in Medicine
• Neuroscience
• Medical Imaging Analysis

Background:

• Deep learning (DL) models are increasingly used for Alzheimer's disease (AD) stage classification.
• These models offer potential for improved diagnostic accuracy and earlier intervention in AD.
• However, significant barriers hinder their translation into clinical practice.

Purpose of the Study:

• To conduct a scoping review of recent research on DL for AD classification.
• To evaluate current performance benchmarks and identify methodological limitations.
• To highlight barriers to the clinical translation of DL models in AD.

Main Methods:

• A scoping review methodology was employed, analyzing 18 peer-reviewed studies published between 2018 and 2024.
• Data extracted included dataset sources, preprocessing, model architectures (CNNs, TL), performance metrics, and translational factors.
• Studies were synthesized to compare performance and identify limitations.

Main Results:

• DL models, particularly transfer learning (TL) and custom Convolutional Neural Networks (CNNs), frequently report accuracies above 90%.
• Performance is highly sensitive to task framing and dataset characteristics, with limited generalizability observed.
• A critical translational gap exists, with only one study performing external validation; cost-effectiveness was rarely reported.

Conclusions:

• DL demonstrates high accuracy for AD classification but lacks robustness and generalizability.
• Clinical translation is impeded by the need for external validation, standardized evaluation, and cost-effectiveness data.
• Future advancements require clinically interpretable, workflow-integrated models with transparent financial cost reporting.