Implementation of the Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) Method to Address Class Imbalance in Alzheimer’s Disease Magnetic Resonance Imaging (MRI) Datasets

Class imbalance in medical imaging datasets often leads to biased machine learning models, particularly in Alzheimer’s disease (AD) diagnosis using MRI. This study proposes the use of Wasserstein Generative Adversarial Networks with Gradient Penalty (WGAN-GP) to mitigate class imbalance in AD MRI datasets. Realistic MRI images were synthesized for underrepresented AD stages, and the quality of the generated data was quantitatively validatedusing the Fréchet Inception Distance (FID), with the lowest FID score recorded at 31.84, indicating a high degree of realism and diversity. The synthetic images were used to augment a dataset of 6,400 T1-weighted scans for training four Convolutional Neural Network (CNN) architectures: ResNet-50, AlexNet, VGG-16, and VGG-19. Results demonstrated statistically significant improvements in balanced accuracy across all models (p < 0.01 for all comparisons). The AlexNet + WGAN-GP combination achieved the highest accuracy of 98.54%, representing a mean improvement of 4.76% (95% CI: 2.45% to 6.98%) over its baseline. Significant gains were also observed for ResNet-50, VGG-16, and VGG-19. These enhancements were consistent across multiple evaluation metrics, including precision, recall, F1-score, and AUC. These findings confirm that WGAN-GP is a highly effective and statistically validated strategy for boosting the diagnostic accuracy of CNN models in Alzheimer's disease classification