ADNET: Dual-Stream Alzheimer's Diagnosis Network with Integrated Grad-CAM Explainability

The increasing prevalence of Alzheimer's disease (AD) requires tools for early and accurate diagnosis. While deep learning models applied to Magnetic Resonance Imaging (MRI) have shown promise in automating diagnostic workflows, their clinical adoption is often hindered by a lack of interpretability, a characteristic of their "black box" nature. To address this challenge, this work introduces ADNET, a dual-stream convolutional neural network designed to classify MRI slices into five stages: Cognitively Normal (CN), Early Mild Cognitive Impairment (EMCI), Mild Cognitive Impairment (MCI), Late Mild Cognitive Impairment (LMCI), and AD. The proposed model achieves a test accuracy of 0.9755, an F1-score of 0.9755, and an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.9987. By using Grad-CAM to generate heatmaps that visualize key diagnostic brain regions, the model provides interpretable predictions, demonstrating its potential as a trustworthy clinical decision-support tool.