AAI_2025_Capstone_Chronicles_Combined

deliberately weakened with a lower learning rate (lr_d = 0.00001). A dropout rate of 0.2 was

applied in the discriminator, and Gaussian noise (σ=0.15) was added to real and fake images to

encourage generator diversity. The generator was updated twice per discriminator update

(gen_update_steps=2) to maintain training balance, particularly during early epochs.

Binary Cross-Entropy with Logits Loss (BCEWithLogitsLoss) was used for adversarial

training, and Cross-Entropy Loss for class predictions. Label smoothing was applied, setting real

labels to 0.98 to avoid discriminator overconfidence. Convolutional layers used Kaiming normal

initialization, and linear layers used Xavier uniform initialization to promote stable training. A

learning rate scheduler with a step size of 10,000 iterations and a decay factor of 0.5 enabled

gradual model refinement.

This ACGAN implementation successfully balanced realism and controllability, producing

visually coherent synthetic X-rays and maintaining stability throughout training, avoiding

common issues such as mode collapse.

Transfer Learning with StyleGAN

The decision to employ StyleGAN for transfer learning was motivated by recent studies

demonstrating its effectiveness in generating medical images. Fetty et al. (2020) showed that

StyleGAN can successfully produce realistic magnetic resonance (MR) and computed

tomography (CT) images suitable for deep learning applications. Similarly, Che Azemin et al.

(2024) employed StyleGAN3 to generate synthetic pterygium images, capturing essential

vascular patterns and morphological features necessary for clinical evaluation.

109