M.S. AAI Capstone Chronicles 2024

CNN Lung Disease Classification

16

AUC-ROC scores showed decent discriminatory power, particularly for ‘Infiltration’. The macro and micro average AUC-ROC scores were 0.7227 and 0.7357, respectively.

The pretrained VGG16 model was then employed to enhance classification performance. Known for its robustness in image classification tasks, VGG16 was adapted to accept grayscale input by adding a layer to convert single-channel images to three channels. This model served as a feature extractor with additional layers appended for final classification. Training over 20 epochs with early stopping and checkpointing ensured optimal performance, resulting in a test accuracy of 49.75% and a test loss of 0.52. The classification report showed improvements in precision and F1 scores. AUC-ROC scores were higher across all classes compared to the Keras Tuner optimized model. Lastly, we leveraged the strengths of our custom-built CNN, the CNN optimized using Keras Tuner, and the VGG16 models by employing a weighted average ensemble approach. This method combined predictions from the best-performing models, resulting in a slight improvement in overall performance metrics. The ensemble achieved an average AUC-ROC of 0.76, matching the performance of the VGG16 model, and demonstrated a balanced improvement across precision, recall, and F1 scores. Notably, the ensemble approach yielded the highest F1 score and AUC-ROC, particularly for ‘Effusion.’

174