AAI_2025_Capstone_Chronicles_Combined

classification results, comparing a custom CNN and a pre-trained VGG model using

transfer learning. Each model was tested with real and synthetic data.

GAN (ACGAN and StyleGAN) Model Performance

The ACGAN model exhibited stable and convergent training dynamics across

epochs. During training, the discriminator loss spiked early but quickly declined and

plateaued by around epoch 150. In contrast, the generator loss gradually increased before

stabilizing at a higher level around the same time. A noticeable gap remained between the

two losses, reflecting the ongoing adversarial dynamic. (Figure 6). This balance suggests

effective adversarial learning, where the generator steadily improved at producing realistic,

class-conditional X-rays, and the discriminator remained competitive without

overwhelming the generator.

Figure 6. ACGAN fine tuning statistics: Generator loss raw (blue), Generator loss smoothed (green),

Discriminator loss raw (orange), and Discriminator loss smoothed (red)

112