A hybrid CNN–BiLSTM network with additive attention for automated four-stage classification of Alzheimer's disease from brain MRI

Abstract

AD is a progressive and irreversible neurodegenerative disorder. Because the structural changes that accompany the earliest stages of the disease are subtle and easily overlooked, manual interpretation of MRI is time-consuming and prone to inter-observer disagreement. This study presents an end-to-end deep learning framework that fuses spatial and sequential reasoning to classify brain MRI into four clinically meaningful stages: Person with no dementia, Person with very mild dementia, Person with mild dementia, and Person with moderate dementia. An additive attention mechanism subsequently weights the most informative positions before a compact dense head produces the final decision. Class imbalance is addressed by capped oversampling of real minority slices combined with residual cost-sensitive weighting, applied to the training fold only; pixel-space SMOTE is deliberately avoided because interpolating between two unregistered MRI slices produces images without an anatomical referent. Exact-duplicate screening is applied prior to partitioning, and a quantitative near-duplicate audit is reported for the fixed folds. Trained on 6,400 grayscale MRI slices under a stratified 70/15/15 protocol, the proposed CNN–BiLSTM–Attention model attains a test accuracy of 95.94%, a balanced accuracy of 96.08%, an MCC of 0.9334, a Cohen's kappa of 0.9325 and a macro-averaged one-vs.-rest AUC of 0.9954, on a slice-level held-out test fold. Per-class analysis shows strong sensitivity and specificity for the three better-represented stages; the extremely rare Person with moderate dementia class (nine test images) shows a favorable point estimate but is too small to support a claim of robust detection on its own. Attention maps and Grad-CAM overlays localize the evidence the network uses, and an error-structure analysis shows that no misclassification spans more than two stages. Because the corpus carries no subject identifiers, patient-grouped partitioning is not possible and the reported accuracies should be interpreted as an upper bound attainable under a slice-level split; this constraint is quantified explicitly and motivates the external-validation study outlined in the manuscript. Overall, the results highlight the methodological value of coupling convolutional feature extraction with bidirectional sequence modeling and additive attention under a leakage-aware, imbalance-aware evaluation protocol for multi-stage AD staging.