AttentionMS-Net: An Attention-Enhanced Multi-Scale Framework for Alzheimer’s Disease Classification with Subject-Level Validation

Many MRI-based Alzheimer’s disease (AD) classification studies report near-perfect accuracy; however, these results are often inflated by data leakage caused by slice-level splitting, where correlated slices of the same subject appear in both the training and test sets. In this study, we introduce AttentionMS-Net, an attention-enhanced multi-scale deep learning architecture that combines channel-spatial attention using the Convolutional Block Attention Module (CBAM) with multi-scale feature aggregation from intermediate EfficientNet-B3 layers for binary AD classification (Non-Demented vs. Demented). Using strict subject-level 10-fold cross-validation on the OASIS dataset (347 subjects, 86,437 slices), our experiments clearly show the impact of data leakage: image-level 10-fold CV achieves about 99.9% accuracy, whereas subject-level 10-fold CV with the same model results in 80.8% accuracy—a reduction of roughly 19 percentage points. Aggregating predictions at the subject level further improves accuracy to 82.4% (AUC: 0.889), suggesting that prediction errors are mainly slice-specific rather than subject-specific. Systematic ablation reveals a complementary interaction between attention and multi-scale components, neither of which performs as well alone. Post-hoc Grad-CAM++ visualization and SHAP analysis suggest that AttentionMS-Net’s attention patterns are focused on ventricular regions—visual biomarkers indicative of overall brain atrophy rather than early hippocampal degeneration. These findings highlight the unreliability of current benchmarks and establish methodologically rigorous baselines for future AD classification research.