AAI_2025_Capstone_Chronicles_Combined

An unexpected finding was the substantial performance difference between ACGAN

and StyleGAN models, particularly evident in the VGG results. This difference likely stems

from several factors, including architectural variations between the GAN models, the

quality and detail of the generated images, and StyleGAN's ability to produce more

anatomically accurate and visually realistic synthetic images. StyleGAN’s sophisticated

generator architecture appears better equipped to capture fine details, textures, and subtle

features crucial for accurate pneumonia detection, thereby significantly enhancing the

effectiveness of the synthetic data.

Conclusion

In summary, the experiments clearly demonstrate that GAN-generated synthetic images

significantly enhance model performance beyond what is achievable with real images

alone. This aligns with existing literature, suggesting that synthetic augmentation effectively

addresses limitations caused by data scarcity and imbalance. StyleGAN demonstrated

superior performance compared to ACGAN, primarily due to its ability to generate higher

quality images with greater anatomical accuracy and finer details. While the custom

ACGAN was effective in generating synthetic data, its slightly lower realism and detail

limited its impact. Despite StyleGAN’s higher computational demands and longer training

times, the substantial improvements in model performance justify these trade-offs, making

StyleGAN a preferred choice for medical image augmentation tasks requiring high fidelity

and realism.

118