AAI_2025_Capstone_Chronicles_Combined

No signs of mode collapse or divergence were observed, and the smoothed loss

curves confirmed consistent training. Early stopping was triggered at epoch 643, indicating

diminishing returns on further training. The resulting synthetic images were manually

inspected and found to be visually coherent and class-consistent, providing valuable data

augmentation for the CNN classifier. These results support the effectiveness of a custom

ACGAN in generating realistic medical images for classification tasks under limited data

conditions.

The StyleGAN model also showed strong evidence of stable convergence and

effective adversarial learning dynamics. During early training, the generator loss exhibited a

downward trend, starting around 4.27 and declining to approximately 1.73 within the first

five ticks. Each tick represents a training milestone during which the discriminator has seen

approximately 4,000 images (both real and generated), summarizing multiple training

iterations. This reduction indicates the generator's progressive improvement at producing

realistic images capable of deceiving the discriminator. In parallel, the discriminator loss

began at 0.77 and gradually increased to 1.07 (Figure 7), reflecting the expected behavior in

a balanced GAN system where the discriminator's task becomes more challenging as the

generator improves.

113