Abstract

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder and the most common cause of dementia worldwide. Early and accurate staging of AD is critical for patient management, yet traditional diagnostic techniques face challenges such as inter-observer variability, limited interpretability, and difficulty in distinguishing early stages. This paper introduces a Modified Wavelet Convolutional Neural Network (MW-CNN) that integrates both spatial and frequency-domain features for robust multi-class classification of AD stages. Using the OASIS MRI dataset with 80,000 brain slices from 461 participants, the model classifies subjects into four categories based on the Clinical Dementia Rating (CDR): Non-Demented, Very Mild Demented, Mild Demented, and Demented. Unlike earlier wavelet-based models, the proposed MW-CNN emphasizes multi-class staging and incorporates Grad-CAM visualizations for interpretability. Experimental results demonstrate a classification accuracy of 98% and a micro-average ROC-AUC of 0.9995, outperforming EfficientNetV2B1 and standard CNNs.