MRI-based dementia classifiers showed systematic performance gaps across racial and ethnic groups. Fairness-aware adaptation, especially RegAlign, reduced those gaps most consistently.