Interpreting Attention Visualization

[fennievdg]

Hi Sybil authors!

Could you elaborate more on how to interpret the attention visualization?

• Is th red the image attention and blue the volume attention?
• If so, how are these image and volume attentions related? What about the model's attention does each explain?
• The blue masks seem to be binary. Does this mean that there is non-zero attention on those regions, a maximum value of attention, or is it based on some other threshold?
• If there are red heatmaps on different sections of the scan, does Sybil find those sections equally important? Or does it give more attention to one of those sections?
• Sometimes, the blue masks seem to extend to sections outside of where the red heatmaps are, or blue mask pieces are seen far away from red heatmaps. What does this mean?

Thank you!