Computer Vision News - May 2018

ISBI DAILY Saturday 21 However, he thinks there is a long way to go before they get reliable results. He points out that one of the downsides with challenges is there’s too much focus on getting that slight bit of improvement in a particular metric. A lot of different people invest an enormous amount of time and it’s very useful, but some of that time could be better spent doing broader research into how to improve the entire process. Not just getting better results on a particular metric, but the entire data collection and validation process , and really seeing if something is clinically useful. Currently, Eugene is working with someone on an approach that tries to do semi-supervised segmentation, which will be extremely useful if they can do it right, because it’s so hard to collect these segmentations. He says that you can’t just have anybody do it, it has to be a radiologist and they don’t even have enough time to do it in clinical practice, so they don’t want to spend time doing it for research. He explains: “ Semi-supervised segmentation means that we can take images that don’t have segmentations, and a few images that do, and use that to get a better model than if we used only the images that have segmentations. This particular approach here is relying on having examples of both healthy controls and sick controls, and we can use these examples right now to greatly inform a mode about what a segmentation would be like by comparing healthy and sick .” Eugene presented his work during a poster session at ISBI 2018. Eugene Vorontsov BEST OF