❓ Why CNN Models? Why ResNet-50 over DenseNet-121?

We chose CNN models as they have proven effective in medical image classification, especially where anatomical structures need to be spatially interpreted.

Among the tested architectures, ResNet-50 provided a good balance between model complexity, training time, and interpretability, which was critical for our Grad-CAM–based explainability.

Although DenseNet-121 showed slightly higher accuracy, ResNet-50 was chosen for final Grad-CAM visualization because:

• Its skip connections make attention maps more localized and interpretable.
• It had more stable convergence during cross-validation.
• And it is widely adopted in clinical imaging research, supporting reproducibility.

❓No augmentation. Why?

We deliberately did not apply data augmentation in this experiment because:

• The dataset is anatomically sensitive. Flipping, rotation, or shifting could disrupt spatial orientation (e.g., left vs. right cheek, superior vs. inferior).
• Our primary goal was model interpretability, not just accuracy.
• We focused on inherent spatial patterns, as learned from unaltered, real patient data.

In future work, domain-specific augmentation strategies could be explored—such as intensity variation or speckle simulation.

❓Why only Grad-CAM?

We used Grad-CAM because:

• It’s a widely accepted, model-agnostic method for visualizing CNN attention.
• It allows direct mapping of class-discriminative regions, which we can compare with anatomical landmarks like the Buccinator muscle or Zygomaticus region.