M.S. AAI Capstone Chronicles 2024

CNN Lung Disease Classification

12

yield noticeable benefits compared to using RGB images, leading to the decision to use RGB images with a resolution of 256x256 for the final models.

In summary, each model variant offered insights into balancing complexity, regularization, and data processing in the context of multi-label classification of lung diseases. The combination of custom-built CNNs, NAS-optimized models, and transfer learning approaches provided a comprehensive evaluation of different strategies for this classification task. Results Our initial modeling of the chest X-ray classification task utilized the basic Convolutional Neural Network (CNN) without any regularization techniques. This model achieved noticeable improvements in training accuracy and reductions in loss over successive epochs. However, the validation accuracy did not mirror this trend and the validation loss began to increase after a few epochs, signaling overfitting. This discrepancy between training and validation metrics was visualized in the training and validation accuracy/loss curves, where the training loss consistently decreased while the validation loss initially decreased but then began to rise. The training accuracy improved steadily, whereas validation accuracy exhibited very minimal improvements,

170