MICCAI 2021 Daily – Wednesday

27 DAILY MICCAI Wednesday will tell you something is wrong. Don’t continue using this image because it looks like something happened. That is how we help networks be more robust to whatever they consume. We get into a field and say, “ You shouldn't be using this to classify because it looks odd. ” That’s one AI-related solution. On the other hand, we had a paper at MICCAI last year. That was on fairness of datasets and classifiers in dermatology across skin tones, so how the network performs when darker skin tones show up in the dataset. This work was done here at the lab by a team of interns coming from Carnegie Mellon Rwanda, in Kigali. We worked across labs at IBM with Kush Varshney and the team in Yorktown Heights and here in Nairobi. Basically, what we discovered is that most of the datasets used in dermatology and machine learning are heavily skewed to lighter skin tones populations. The samples that we have for darker skin tones are very small. The network has a very tiny portion of datasets to learn how skin lesions look on darker skin tones. It is not a solution yet, but we are in the process of understanding representations of the skin tones for these deep learning applications. Has some of the work that you’ve done in healthcare translated into the real world? That’s a tough question. I wouldn’t say that anything has gone straight to users, particularly because we are talking about machine learning in healthcare. This is a serious process, and we are just at the beginning. When I try other populations, maybe it will not work. I would say that we are on the research stage, at the moment. It’s a long process until you get it to the end-user. It’s a long and necessary process to get all the ethical and medical approvals. It’s better to have a robust model ready for different communities, rather than just allow any model for end- users. You don’t know if you are impacting those users negatively or positively. But people are waiting! [ both laugh ] We are trying to solve real problems, but we have a gap between the moment you are ready and when the market is ready. Doesn’t this feel frustrating for Celia Cintas “I didn’t have any preconceptions before I moved here!”