Computer Vision News - March 2020

3 Summary Orcun Goksel 29 registration problem. it is more about f inding objects in each image frame and sequence. It can draw more on generic computer Tracking is a di fferent problem that vision techniques because of the level of activity in the f ield – sel f-driving cars, for example. Computers can easi ly integrate di fferent methods with l ittle or no effort. Registration, however, is a new thing, and the medical imaging community has to do a lot of it from scratch. How does al l this progress and advancement in deep learning translate at a patient level? In Orcun’s opinion, things are getting better, and with retrospective studies and al l the chal lenges that are happening now, it ’s relatively easy for even a smal l company or research group to access large datasets and prove their worth, whereas in the past only larger groups could make a meaningful in Orcun’s mind, there is a better solution. “How do people learn?” he asks. “A radiologist, when they ’re studying, doesn’t see a mi l l ion examples. They see a few images where things are normal , they see some counter examples, and then they have a learning process. The way we currently train the deep learning networks is not yet able to repl icate that process. I think we should look at human learning more and try to draw inspiration from that.” What about other appl ications in the f ield of computer-assisted intervention , l ike tracking and navigation, for example? He says that surgical guidance and navigation – intraoperatively al igning models to pre-operative images or planning – presents a interactive ultrasound simulation for sonographer training by Oliver Mattausch interactive ultrasound simulation for sonographer training by Maxim Makhinya