AAI_2025_Capstone_Chronicles_Combined

simplicity of the VGG architecture also facilitates straightforward hyperparameter tuning for this

classification task. An Auxiliary Classifier GAN (ACGAN) was chosen to associate class labels with

generated images, a requirement for training the CNN model. StyleGAN was selected as the pre

trained GAN model due to its ability to produce high-quality images that can be adjusted for

specific problem sets. The architecture of each model is outlined in the following sections, with

further detail provided on design choices.

Custom Convolutional Neural Network

A CNN was employed for the binary classification task of distinguishing between

NORMAL and PNEUMONIA chest X-ray images. The architecture consisted of two convolutional

blocks, each comprising a convolutional layer followed by batch normalization, ReLU activation,

and max-pooling. All convolutional layers used a 3×3 kernel with "same" padding to preserve

spatial dimensions, while 2×2 max-pooling progressively reduced spatial resolution to extract

increasingly abstract features.

Following the convolutional blocks, a flattening operation converted the multi

dimensional feature maps into a one-dimensional vector for classification. This was passed

through a fully connected dense layer, followed by batch normalization and dropout. The final

output layer consisted of a single neuron with a sigmoid activation function, appropriate for

binary classification.

Hyperparameters were tuned using a RandomSearch strategy via Keras Tuner, exploring

variations in the number of convolutional filters, dense layer units, dropout rates, and learning

105