This study presents AlzFormer, a novel deep learning framework utilizing spatiotemporal self-attention to classify Alzheimer’s disease (AD), mild cognitive impairment (MCI), and cognitively normal (CN) individuals from structural MRI scans. By modeling MRI volumes as sequential slice-based inputs and fine-tuning a pre-trained TimeSformer model, AlzFormer achieved 94% accuracy and high class-wise F1-scores, while attention map analyses highlighted clinically relevant brain regions, demonstrating both robust performance and interpretability in multiclass AD diagnosis.