M.S. AAI Capstone Chronicles 2024

CNN Lung Disease Classification

18

require lateral view which limits this model's reliability for practical clinical application. Our model's performance, while encouraging, underscores the need for further refinement to meet stringent clinical requirements. The AUC-ROC scores, particularly for 'Infiltration', demonstrate the model's ability to discriminate between classes. However, the overall accuracy indicates that misclassifications are still frequent. In a clinical context, this could lead to both false positives and false negatives, which could have significant implications for patient care if not properly managed. Unexpected Results The most notable unexpected result was the significant performance boost achieved by both the ensemble model and VGG16 pretrained model, compared to individual models. This finding aligns with recent research by Zhang et al. (2020), which suggests that ensemble methods can capture complementary features from different architectures, leading to more robust predictions. This unexpected synergy between the VGG16, custom-CNN, and Keras Tuner models highlights the potential of combining transfer learning (through VGG16) with custom-optimized architectures. It suggests that future work in medical image classification could benefit from exploring diverse ensemble combinations rather than focusing solely on optimizing individual models. Future Work and Deployment Further work could include a multi-faceted approach encompassing model optimization, data enhancement, and deployment. For model optimization, we aim to explore more advanced CNN architectures such as ResNet, which have demonstrated

176