Stratification of Alzheimer’s disease (AD) progression effectively is crucial for the efficient management of patients. This study compares cutting-edge deep learning models for multi-class classification of AD severity based on MRI scans. A balanced dataset of 10,240 images from four classes—No Impairment, Very Mild Impairment, Mild Impairment, and Moderate Impairment—was used. Pre-trained convolutional neural networks (CNNs), i.e., DenseNet201, MobileNetV2, and DenseNet169, were first considered as baselines with transfer learning, achieving accuracy between 0.81 and 0.83. To optimize performance, heavy hyperparameter tuning and optimization were carried out for the most promising models. This yielded considerable gains, with MobileNetV2 rising to 0.90 accuracy upon tuning. More significantly, the tuned DenseNet169 topped them all, with 0.96 accuracy and 0.95 precision and recall. These results highlight the point that diligent tuning can push even smaller models like DenseNet169 to state-of-the-art performance in distinguishing subtle stages of AD. In addition to accuracy, the study emphasizes interpretability. Explainable AI techniques, such as LIME and Saliency Maps, were integrated to ensure clinical reliability. Visualizations confirmed that the models employ meaningful neuroanatomical features for prediction and therefore provide both excellent diagnostic performance and clinically meaningful insights to guide clinical decision-making.