M.S. AAI Capstone Chronicles 2024

CNN Lung Disease Classification

15

refinements techniques using model optimization tools were needed.

Keras Tuner was then used in combination with RandomSearch to optimize the structure and hyperparameters of a CNN model. The model-building function adjusted key hyperparameters such as the number of units in the convolutional layers, the type of activation functions, the dropout rates, and the learning rate. Our objective was to minimize validation loss with early stopping, and model checkpointing to prevent overfitting and to save the best-performing model, respectively. The search was conducted over 15 trials, each comprising a maximum of 10 epochs.

The optimal hyperparameters determined by Keras Tuner were as follows:

● 64 units in the first convolutional layer ● 3 convolutional layers ● 128 units in the dense layer ● A learning rate of 0.001

Using these optimal hyperparameters, the model was trained and evaluated. The model achieved a test accuracy of 47.65% and a test loss of 0.552. Additionally, the

173