M.S. AAI Capstone Chronicles 2024

CNN Lung Disease Classification

9

In the context of neural architecture search (NAS) and hyperparameter optimization, utilizing a combination of Keras Tuner and random search is an effective strategy for refining convolutional neural networks (Cetiner & Metlek, 2023). Keras Tuner automates the exploration of neural network architectures and hyperparameters, not only streamlining the model development process but also leading to quicker optimization and performance improvements. By leveraging random search, this approach provides a more efficient alternative to exhaustive methods like grid search, as it samples randomly within a predefined hyperparameter space and enables the discovery of near-optimal configurations with fewer trials (Cetiner & Metlek, 2023). This is particularly advantageous when dealing with complex tasks such as multi-label classification, where the search space can be vast and computationally demanding. Experimental Methods Three methods were explored in our classification project, including an iterative CNN built from scratch, a CNN optimized through Neural Architecture Search (NAS) using Keras Tuner, and a transfer learning approach using the VGG16 model. We will evaluate the effectiveness of these models in diagnosing different lung conditions and to understand their strengths and limitations.

Custom CNN Models:

The initial custom CNN was designed with an architecture consisting of three convolutional layers followed by max-pooling operations. A flatten layer was then used to transform multi-dimensional feature maps into a one-dimensional vector, followed by

167