AAI_2025_Capstone_Chronicles_Combined

and converted into a compatible format using a provided script. Images were saved as

uncompressed PNGs to ensure lossless quality.

Training was conducted using a batch size of 32 and a gamma value of 6.6. StyleGAN3

training proceeds in "ticks," with each tick representing the generation and processing of

1,000 images. Model snapshots, including generator, discriminator, and optimizer states,

were saved every 10 ticks. Internally, the generator uses a non-saturating logistic loss and

path-length regularization, while the discriminator uses a logistic loss with R1

regularization (Karras et al., 2020). Overall model evaluation includes Frechet Inception

Distance, Precision and Recall, and Perceptual Path Length, but for this study, Generator

Loss and Discriminator Loss were used as primary performance measures. Regularization

terms, such as path-length penalty (Loss/pl_penalty) and R1 penalty (Loss/r1_penalty),

were also referenced to assess model stability and constraint.

The training process was repeated separately for each image category, resulting in one

StyleGAN model for NORMAL images and another for PNEUMONIA images. After training,

1,000 synthetic images from each model were generated for use in the CNN classification

task.

Results

The following section outlines two aspects of the results: image generation using

GANs and CNN model performance. The first part describes how well the GAN models

converged to generate believable images. The second part evaluates the CNN

111