For prompt and accurate determination of dementia stages, it is extremely important for clinical decision making and patient care planning. Conventionally, radiologists determine and interpret brain MRI scans manually—this process is often long and, to some extent, subjective. Accordingly, we present an automated and interpretable dementia stage classifier, based on a variant of Swin Transformer architecture, trained on the OASIS dataset. In addition to classifying dementia severity into four levels (cognitively normal, very mild, mild, and moderate) using the Clinical Dementia Rating (CDR); we demonstrate additional clinical trust by presenting attention maps that support the rationale on how the model makes its decisions. Our modified Swin Transformer architecture is trained and validated on OASIS dataset using a series of data augmentation methods to help promote generalization. Further, we considered various robust training techniques such as early stopping, gradient clipping and label smoothing to promote stability and accuracy. We put together a custom attention visualization module to help clinicians interpret the model's areas of focus. Experimental results indicate high classification accuracy for all dementia severity levels, and classification stability when the model is evaluated with noisy imaging, which suggests strong concept feasibility for application to the real clinical world.