M.S. AAI Capstone Chronicles 2024

CNN Lung Disease Classification

10

two dense layers. ReLU activation functions were employed in all layers except the final one, which used a sigmoid activation function to output probabilities for each of the four diagnostic classes: ‘No Finding,’ ‘Infiltration,’ ‘Effusion,’ and ‘Atelectasis.’ The Adam optimizer was chosen for its strong performance across deep learning tasks, and the dataset was split into training, validation, and test sets with an 80%-10%-10% ratio. The model was trained with a batch size of 18 due to computational constraints. During the development process, several iterations of this custom CNN were explored. The initial model, which did not include any regularization, was followed by versions incorporating dropout and varying degrees of L2 regularization to address overfitting. We then experimented with increasing model complexity by adding more layers and units and later chose a best-performing model for further evaluation. This approach allowed us to refine the model architecture and identify the most effective regularization techniques. To test additional CNN architectures, we employed NAS using Keras Tuner (Keras, n.d.). This tool allowed us to efficiently explore various hyperparameters and architectural choices. The search space included the number of convolutional and dense layers, units in each layer, activation functions, dropout rates, and learning rates. The NAS-optimized CNN was trained using a binary cross-entropy loss function, with validation accuracy as the primary evaluation metric. The training process involved monitoring validation loss for early stopping to prevent overfitting, with the Keras Neural Architecture Search (NAS) with Keras Tuner:

168