UCLA

The advancement of generative models have enabled the creation of high-quality synthetic images, offering a cost-effective solution to augment datasets in resource-intensive fields like medical imaging. This study evaluates the efficacy of synthetic image augmentation for Alzheimer's disease classification using MRI data. By training convolutional neural networks (CNNs) and vision transformers (ViTs) on varying proportions of original and synthetic images, we assessed the impact of selective augmentation across dementia severity classes, comparing its performance to traditional and automated augmentation methods

Overall, synthetic image augmentation demonstrates significant potential, particularly for CNNs, as a complementary or standalone augmentation strategy. Selectively augmenting specific classes, such as mildly and moderately demented cases, improved CNN performance, reducing cross-entropy loss when compared to training on the original dataset alone. Synthetic augmentation often outperformed traditional augmentation, including automated methods like AutoAugment. For ViTs, performance gains were minimal, reflecting the architecture's reliance on larger datasets. These findings highlight its value in resource-constrained settings, emphasizing selective application to maximize diagnostic accuracy in medical imaging.