Alzheimer Disease (AD) is a progressive neurodegenerative disease that causes memory loss, cognitive impairment and behavioral deterioration. The aim of this research is to create a very precise and explainable deep learning model that would help diagnose early AD based on structural MRI scans. In order to do this, this research has proposed a multi-stage diagnostic pipeline which include six key components, namely: Generative Adversarial Networks (GANs) to generate synthetic data and normalize intensity, Mask-RCNN to segment brain regions, Graph Convolutional Networks (GCNs) to extract spatial and structural features, a hybrid feature selection method that will include Crayfish Optimization Algorithm (COA) and American Zebra Optimization (AZO), a Siamese DenseNet Alzheimer Detection (SDAD) model which fuse DenseNet. The SDAD model proposed was tested on structural MRI data and demonstrated a classification accuracy of 98.42 percent which is higher than that of individual baseline models in precision, recall, specificity and overall diagnostic performance. The proposed study is robust, interpretable, and effective deep learning model, which combines several state-of-the-art methods to detect AD. The findings show that deep feature extraction, optimization, and explainability is used to build reliable and clinically trustworthy automated Alzheimer diagnosis. © © 2025. Published by Elsevier Ltd.