AAI_2025_Capstone_Chronicles_Combined

improves at distinguishing real from synthetic images. This adversarial process results in high

quality synthetic images that complement the original training data. A loss function ensures

both models improve over time and prevents one from consistently outperforming the other.

The production of synthetic medical images using GANs offers multiple benefits, from

data augmentation to improved data privacy (Arora & Arora, 2022). GANs can increase dataset

sizes when real data availability is limited. Auxiliary Classifier GANs (ACGANs), which generate

synthetic data for specific classes, can address class imbalance issues by augmenting minority

classes. Additionally, using synthetic data helps alleviate ethical concerns around patient

privacy, as GANs can generate realistic datasets without directly attributing data to any

individual patient.

Synthetic data generated by GANs has been observed to improve CNN classification

accuracy (Frid-Adar et al., 2018). Synthetic images can be created from scratch using deep

convolutional GANs or through style transfer techniques leveraging pretrained networks such as

StyleGAN (Fetty et al., 2020). Although both methods are effective, transfer learning produces

higher-quality and more realistic images in a resource-efficient manner, as the model has

already been extensively pretrained (Che Azemin et al., 2024).

Other Methods to Augment and Analyze Medical Datasets

In the rapidly evolving field of AI model development, even GANs are increasingly

being replaced by diffusion-based models. Diffusion models generate high-quality images

by progressively and deterministically denoising an image surface. Unlike GANs, diffusion

103