Abstract
Alzheimer's disease classification using MRI scans is challenging due to class imbalance in medical imaging. This study examines the effects of data augmentation, weight balancing, and synthetic image generation on classification accuracy. Six deep learning models—EfficientNetB4, ResNet152, VGG19, Xception, InceptionV3, and DenseNet121—were tested under three conditions: (1) no augmentation, (2) traditional augmentation, and (3) synthetic image augmentation. Models trained on imbalanced data performed poorly, while weight balancing and synthetic data improved accuracy and robustness. The best performance (97%) was achieved with EfficientNetB4 and Xception using synthetic images, highlighting synthetic data as a promising solution for class imbalance in MRI-based Alzheimer's classification.