AAI_2025_Capstone_Chronicles_Combined

rates. The search was conducted over 25 trials using Kermany’s pneumonia dataset. The best

performing model used 64 filters in both convolutional layers, 64 neurons in the dense layer,

and a dropout rate of 0.3. L2 regularization (λ = 0.001) was applied to reduce overfitting. Adam

was chosen as the optimizer, with a learning rate of 0.0001 yielding the most stable and

accurate results when minimizing binary cross-entropy loss.

Training followed a structured methodology to ensure robustness. Image preprocessing

was performed using Keras’s ImageDataGenerator, applying real-time data augmentations such

as normalization, shear transformations, horizontal flipping, and brightness adjustments. All

images were resized to 128×128 pixels and converted to grayscale to standardize inputs and

minimize variations due to lighting or color differences. Data augmentation techniques helped

prevent overfitting and encouraged the model to learn actual features rather than memorize

feature locations.

The model was trained three times, corresponding to each data source: Kaggle training

data, synthetics generated via custom ACGAN, and synthetics produced through StyleGAN. For

each synthetic dataset, 1,000 images representing pneumonia and normal states were

aggregated, then divided into training and validation sets with an 80/20 split. Actual Kaggle test

images served as the test dataset for all models.

Leveraging Pre-Trained VGG16 Model

To improve validation performance metrics, transfer learning was tested using the

VGG-16 model due to its versatility and proven success in similar applications. The model

106