Interpretable Deep Learning for Alzheimer Diseases Classification: Integrating XAI for Trustworthy AI-Assisted Diagnosis

Alzheimer’s Disease (AD) moves progressively through the brain leading to neurodegeneration in millions of individuals across the world. The proper identification of AD subtypes with Normal Cognition (NC), Mild Cognitive Impairment (MCI), Moderate Alzheimer’s (MAD), and Severe Alzheimer’s (SAD) remains vital for early healthcare decisions regarding diagnosis and therapy. Traditional deep learning models including both Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) demonstrate restricted classification abilities because they are designed to extract local features alone or process global dependencies. We have designed a hybrid CNN-ViT framework that joins CNN spatial understanding with ViT global context to handle enhanced medical diagnosis classification. The implementation of SHAP Grad-Cam and LIME as XAI techniques allowed healthcare professionals to interpret the model's decision-making process for clinical use. The implemented hybrid CNN-ViT model performed better than standard CNN and ViT methods along with ResNet-50 and EfficientNet-B4 variants by reaching 92.5% accuracy along with 91.2% F1-score and 94.0% AUC-ROC. The model demonstrated excellent subtype discrimination according to ROC-AUC curves in addition to its low confusion matrix analysis misclassification rates. Results from ablation studies proved that the elimination of CNN ViT or XAI components substantially deteriorated model performance which demonstrates why a mixed and intelligible approach is required. The CNN-ViT and XAI model achieves both superior classification results and trustworthy operation characteristics according to the research findings thus proving suitable for clinical use. The research delivers an important solution for closing the clinical application distance between AI diagnostics systems by delivering a robust framework for diagnosing Alzheimer's disease.