AAI_2025_Capstone_Chronicles_Combined

results strongly validate StyleGAN’s ability to generate realistic and diverse synthetic

medical images that substantially improve CNN training for pneumonia detection.

Further experimentation with a fine-tuned VGG-16 model provided additional

insights. The VGG model trained on real images displayed smoother training curves but still

showed overfitting, as indicated by gaps between training and validation performance. This

model achieved 78% accuracy, 72% F1 score, 87% precision, and 71% recall. Balancing

the real image dataset slightly improved performance to 82% accuracy, 78% F1 score, 88%

precision, and 76% recall, though with more pronounced curve spikes and declining

validation accuracy during training.

The VGG model trained on ACGAN-generated images demonstrated improved

performance, achieving 81% accuracy, 81% F1 score, 82% precision, and 84% recall, with

spiky but closely aligned training and validation curves. Most notably, the VGG model

trained on StyleGAN-generated images achieved the highest overall performance across all

models, reaching 92% accuracy, 91% F1 score, 92% precision, and 92% recall. Training and

validation curves for this model indicated stable convergence and minimal overfitting,

particularly when trained on synthetic StyleGAN datasets.

116