M.S. AAI Capstone Chronicles 2024

CNN Lung Disease Classification

14

After trying different techniques to improve our model, we used L2 regularization to reduce overfitting. The validation loss and accuracy both plateaued and it was evident the model was overly constrained. We then tried using a weaker form of L2 regularization, shown below. This approach resulted in a slight improvement. The validation loss decreased and was improving at generalizing to new data, and the accuracy stayed within a range that indicated better performance than previous models. However, these changes were not enough to significantly improve the model's overall performance. This indicated that we needed to try other methods in addition to dropout and regularization.

We then introduced additional layers to increase model complexity and applied dropout regularization. These changes aimed to improve generalization by forcing the model to develop more robust features, thus increasing performance with reduced overfitting. The updated model showed improved validation accuracy for nearly all epochs and reached a test accuracy of 45.3%, and a test loss of 0.56. Overall, our iterative approach with CNNs built from scratch was promising, however further

172