Grad-CAM colours where a network’s last convolutional features pushed its prediction: red is higher relative attention, blue lower. Each heatmap is scaled per image, so colours compare regions within one image, not across images.
Models A–D are the four ensemble networks’ outputs in the order the app saved them. The app does not record which architecture made which file, so the same letter on two scans may be different networks.
LIME marks superpixels that support (green) or oppose (red) the prediction. Its four runs per scan rarely overlap, so treat any single run with caution.
Blend fades the app’s saved overlay over the original; nothing is recomputed. These are research outputs, not clinical findings. Full analysis.