AAI_2025_Capstone_Chronicles_Combined

CNN Model Performance

The CNN model trained solely on real images exhibited strong baseline

performance, achieving an accuracy of 71% and an F1 score of 65%. Model training

demonstrated clear convergence, with decreasing validation loss and increasing accuracy

over epochs. Although validation curves displayed significant spikes, suggesting some

overfitting, overall learning remained effective. However, the lower F1 score and recall

highlight limitations due to dataset imbalance and data scarcity.

Balancing the dataset by under-sampling Pneumonia images produced a more

balanced model, improving metrics to an accuracy of 76% and F1 score of 72%. Despite

persistent validation curve spikes and signs of overfitting, the improvements confirm the

value of addressing class imbalance to enhance CNN classification performance.

Synthetic data augmentation showed even greater promise. The CNN trained on

ACGAN-generated images outperformed models trained on real images, achieving 80%

accuracy, 80% F1 score, 82% precision, and 80% recall. Training and validation curves also

smoothed out, with faster convergence and reduced overfitting. These results highlight the

effectiveness of GAN-generated images in enhancing model performance, especially in

data-limited scenarios.

CNN models trained on StyleGAN-generated images achieved the highest overall

performance, with accuracy, F1 score, precision, and recall each reaching 86%. These

115