Computer Vision News - October 2018

three categories, in the context of lesion segmentation these are: lesion; non-lesion; and a new, uncertain class. In her work, what she does is evaluate the performance on remaining certain predictions and sees if the performance is greater than that of when we do no uncertainty classification. If this is the case, then these uncertainty measures are indeed quantifying incorrect predictions, which is what we want to see, so we can use and extend the use of these methods. Tanya laughs nervously when we ask her how it feels to have an oral at MICCAI with her first paper: “ Stressful! ” However, she quickly adds: “ But it’s a huge opportunity and I don’t make light of it. What it means is that I’ve been preparing a lot . I have been taking the support of my colleagues, and especially with Tal on how we can communicate this work in a way that can be easily understood, so that someone who is working in a different domain can see that and make those connections. ” Tal adds: “ I think the importance of the work is that deep learning has made a lot of progress and has been integrated into this field, but in the context of, for example, multiple sclerosis, we’ve found that sometimes there are false positives and false negatives, particularly with the small lesions and around the boundaries of lesions. Rather than give a clinician a deterministic answer to say, ‘I’m sure there’s a lesion here,’ when it could be wrong, I think it’s important now that we can quantify that and give it to the clinician and say, ‘I think there’s a lesion here, I’m not sure.’ Then have the clinician be able to then look at it and say, ‘Okay, I’d better look a little bit further. ’” Now that Tanya has finished her Masters, she is working in industry for a company called Imagia . Imagia is an AI start-up based out of Montreal, doing research and engineering . It is focused on performing early cancer detection and detecting features from imaging that can be used for prediction and early diagnosis tools . “ A recent method from computer vision says that we can apply dropout at test time and get multiple predictions. ” Tanya Nair 35 Monday