Comparative analysis of multiple deep learning models with mitigation-driven approaches for enhanced Alzheimer's...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Alzheimer's disease (AD) diagnosis via structural MRI is challenging.
• Deep learning (DL) offers automated dementia detection potential.
• Lack of comprehensive DL model comparisons for AD diagnosis.

Purpose of the Study:

• To systematically compare ten DL architectures for AD diagnosis using MRI.
• To evaluate a novel 2D coronal slicing methodology for MRI analysis.
• To assess the impact of class imbalance mitigation on diagnostic performance.

Main Methods:

• Analysis of 14,983 T1-weighted MRI scans from 1346 patients using a 2D coronal-10 slicing method.
• Systematic comparison of ten DL architectures, including CNNs, Vision Transformers, and Capsule Networks.
• Application of class imbalance mitigation strategies (SMOTE, cost-sensitive learning, focal loss).

Main Results:

• ECAResNet269 with class imbalance mitigation achieved 74% balanced accuracy.
• Traditional CNNs outperformed Vision Transformers and Capsule Networks, which failed classification.
• The 2D grid method retained 96% diagnostic information and offered 4.2x faster processing than 3D approaches.

Conclusions:

• Traditional CNNs are most effective for neuroimaging classification tasks.
• ECAResNet269 demonstrates clinically relevant performance for dementia screening.
• The 2D grid methodology balances diagnostic accuracy and computational efficiency for clinical deployment.