ICCV Daily 2019 - Tuesday

find very challenging. Inside of that I found my own personal challenge that is creating interpretable models for physicians. My goal is to build up a bridge between the two so that if the physicians want to check if there is something going on in the neural network, he can. That's why I would say that I don't know if I'm dedicating my career, but I dohave the intention todoa lot of research in the interpretability of computer vision models, of machinery models, especially for medical imaging. What would be the main patient benefit of your work? One thing is that when the patient is given a decision like, you have cancer at this stage. It's good to get him involved and explain to the patient why the prediction is something like that. This is something even physicians cannot do. Very often, they first have an intuition of what's in the image they are looking at. They have these intuitions then they go and they try to explain why. Of course, they have to write a report so they have to explain what it is that they are looking at in the image. But it's very intuitive, and it's very subjective. In fact, if you look at the disagreement rates between different doctors and different physicians, they do have a high disagreement. If they are non- experts, it's even higher. What about among the same doctor having disagreements with himself at different times? I think it's less. The thing is that these images are huge, and very often also the diagnosis depends on what region they focus on. They generally have a big scan, and then they zoom in on something. Then they look and focus their attention on something. Then they write the report. Different doctors look at different areas. Disagreement is high if you look at different areas, but it's difficult that a pathologist would change his mind. That's also why the disagreement is very high because they might have harsh discussions on that level. The good thing about the neural network is that it introduces objectivity in the process because, first of all, it highlights regions of interest. 14 Women in C.Vision DA I L Y