AAI_2025_Capstone_Chronicles_Combined

Figure 7. StyleGAN fine-tuning statistics: Generator loss (blue), Discriminator loss (orange), real

image loss (brown and purple), R1 penalty (red), and P1 penalty (green).

Overall, the training metrics demonstrate that the StyleGAN3 model achieved

balanced and convergent training dynamics. The generator steadily improved image

synthesis quality, while the discriminator maintained sufficient pressure to drive

meaningful updates. No signs of mode collapse or divergence were observed, and the loss

patterns were consistent with successful GAN training. The fine-tuned models for both

image classes produced high-quality, realistic synthetic X-rays. These images were

assessed visually and judged to be indistinguishable from real images by laypersons. While

no quantitative metrics were used, the images were deemed suitable for augmenting

downstream CNN classification tasks based on their perceived realism and anatomical

consistency. These findings support the effectiveness of StyleGAN3 as a transfer learning

approach for medical image synthesis in data-constrained environments.

114