Alzheimer’s Disease is a complex and debilitating neurogenerative disorder that necessitates accurate diagnosis and classification. This study investigates the efficacy of three deep learning models - EfficientFormer-L1, EfficientNet B3, and DenseNet-169 - in classifying AD into four distinct categories. To optimize the performance of these models, we compared the AdamW and Lion optimizers, and explored the impact of three data balancing strategies: class weighting, data augmentation, and Synthetic Minority Over-Sampling Technique (SMOTE). The DenseNet-169 model paired with the AdamW optimizer achieved the highest accuracy of 98.87% when class weights were employed. The augmented dataset favored the EfficientFormer-L1 model, which reached an accuracy of 99.53% when used with the Lion optimizer. The SMOTE balancing technique revealed that EfficientFormer-L1 model to be robust, achieving accuracies of 99.69% and 99.61% with the AdamW and Lion optimizers, respectively. These findings suggest that the Lion optimizer is particularly well-suited for transformer-based models, such as EfficientFormer-L1.