ENGRAP: an explainable ai application for mri-based staging of Alzheimer’s disease

Abstract

Staging Alzheimer’s disease (AD) from brain magnetic resonance imaging (MRI) is challenging because inter-class differences are subtle and clinical interpretation requires transparent decision support. This study presents ENGRAP, a hybrid deep architecture that combines a ResNet-50 feature extractor, a capsule layer, and a lightweight Transformer encoder with a classification token. The model performs late fusion to predict four stages (Non, Very Mild, Mild, Moderate). Training uses focal loss with class weighting, and evaluation relies on macro-averaged metrics to ensure equitable assessment across all diagnostic categories. On a validation subset (n = 2,800) of the Alzheimer’s Disease Neuroimaging Initiative dataset, ENGRAP attains 0.988 accuracy and 0.988 macro-F1. The confusion pattern indicates that most residual errors occur between adjacent stages, while Very Mild Dementia is classified without error. Training dynamics show rapid convergence and stable performance across epochs. To support interpretability, gradient-weighted class activation mapping and Signed RISE highlight disease-relevant regions in the images, and a comparative evaluation against LIME, SHAP, and Score-CAM situates the proposed approach within the broader landscape of explainable AI methods. Overall, the results suggest that ENGRAP provides competitive accuracy together with transparent visual explanations, making it a promising component for research-oriented screening and decision-support pipelines that require reliable multi-class staging from routine MRI.