Computer Vision News - January 2019

15 ABCD Neurocognitive Prediction Challenge then is that we’ve already released data on 4,500 or so children’s neuropsychological assessments including their fluid intelligence scores to the public to analyze. Those, merged with the skull-stripped images, the T1 images, that Kilian has put together constitute the training dataset for the challenge. Those are publicly available to anybody with any application. They don’t have to be in our competition. The timing for the next data release is March 15, 2019. That’s the current time slotted for the next release of public data including the neuropsych scores. We’re soliciting people to submit their prediction algorithms before that date so that they do not have access to the test dataset, which is the training score, the neurocognitive scores, for the people who are currently not in the current version of the baseline data. That’s the structure of the challenge. Kilian: This is kind of outside the format of how, for example, MICCAI proposes challenges normally. The reason is not driven by us. The reason for this difference is basically the release schedule of data by ABCD. Wes: That’s not in our control, of course. What is the motivation behind the challenge? Wes: From my perspective, the justification for the contest is twofold, maybe threefold. The first reason why I was eager to join the challenge is because I wanted to promote the usage of the ABCD data outside of the consortium. A large chunk of how we, as a consortium, will be evaluated is how many people outside of our internal consortium members actually end up using the data in a productive way. I felt, and I’m sure Kilian felt the same way, that this would be a very nice way to really spur people to use the data. Motivation #2 is that I’m really curious to see how well we can describe the relationship between the brain and behavioral measures. Personally, I settled on the neurocognition because that’s something that clearly has a seat in the brain. There’s been a lot of research on that showing a lot of brain morphometry and function and how that’s associated with neurocognition in different ways. Susan can speak more to that in terms of a scientific perspective. Also, the fluid intelligence measure is very normally distributed, so it’s something that I think there’s a lot of evidence for. I think there’s a chance that we’ll be able to predict it with a meaningful amount of variance explained. The third reason is basically promoting methods development for modern machine learning approaches to the brain prediction problems. I’m just really curious to see how well we can do. what are the methodologies that really work well in this context? Is Challenge Computer Vision News KilianPohl is ProgramDirector at SRI International

RkJQdWJsaXNoZXIy NTc3NzU=