A HYBRID CNN-TRANSFORMER FRAMEWORK FOR MULTI-CLASS ALZHEIMER’S DISEASE CLASSIFICATION USING MRI: A ROBUST AND INTERPRETABLE APPROACH

Abstract

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder for which timely and reliable assessment is important. Structural magnetic resonance imaging (MRI) provides useful anatomical information, but manual review is time-consuming and subject to reader variability. This study presents a hybrid CNN-Transformer framework for four-class classification of 2D MRI slices. A ResNet-18 backbone extracts local anatomical features, while a Transformer encoder models global relationships among the resulting spatial tokens. A feature-fusion layer combines the pooled CNN representation with the Transformer’s derived class-token embedding before prediction. The model was evaluated on a Kaggle OASIS-derived dataset containing NonDemented, Very Mild Demented, Mild Demented, and Moderate Demented classes. Because subject identities were unavailable, the result is interpreted within the stated slice-level protocol rather than as evidence of subject-level clinical generalization. The model achieved 99.81% validation accuracy. Class-wise precision and recall were high, and Grad-CAM visualizations were used to inspect spatial evidence associated with predictions. These findings support the value of combining local and global representations for MRI-slice classification. However, the slice-level split may yield optimistic estimates because slices from the same subject can occur in both subsets. Subject-level and external validation are therefore required before conclusions can be drawn about patient-level generalization or clinical use.