Comparative analysis of multiple deep learning models with mitigation-driven approaches for enhanced Alzheimer’s disease classification

Alzheimer's disease diagnosis from structural MRI remains challenging in clinical practice. While deep learning shows promise for automated dementia detection, comprehensive comparisons of different neural network approaches are lacking. It analyzed T1-weighted MRI scans comprised 14,983 2D grid images derived from 1346 unique patients. Ten coronal brain slices spaced 2mm apart were arranged in 512 × 512-pixel grids using our 2D coronal-10 slicing sMRI methodology to preserve anatomical relationships while reducing computational demands. Ten deep learning architectures were systematically compared, including traditional CNNs, Vision Transformers, and Capsule Networks. Patient-level data splitting prevented information leakage. ECAResNet269 achieved the highest balanced accuracy (63%), with mild performance across all classes: dementia (38% sensitivity/77% specificity), MCI (72% sensitivity/66% specificity), and healthy controls (44% sensitivity/90% specificity). Class imbalance mitigation strategies substantially improved model performance, with combined SMOTE, cost-sensitive learning, and focal loss approaches achieving 74% balanced accuracy and (78% CN, 76%MCI, 69% AD) sensitivity in the ECAResNet269 model. Pretrained CNNs architectures substantially outperformed advanced methods-Vision Transformer and CapsNets showed complete classification failure. The 2D grid method retained 96% of diagnostic information compared to 3D approaches while providing 4.2 × faster processing. Traditional CNNs architectures remain most effective for medical neuroimaging classification. ECAResNet269 achieved clinically relevant performance suitable for dementia screening applications. The 2D grid methodology successfully balances diagnostic accuracy with computational efficiency, enabling deployment on standard clinical hardware.