Computer Vision News - March 2020

2 Summary AI in Medical Imaging 8 problem. It wi l l take a U-Net and apply it to any problem to make a new contribution or paper. “There are lots of shortcuts,” he tel ls us. “But I think slowly this type of low-hanging fruit has started to disappear because you kind of run out of it. There is a transformation happening. One has to adapt these techniques and address other problems around. For example, when we don’t have enough data and we don’t have enough annotations, or when we have problems l ike image reconstruction or registration. That requires us to go back to the drawing board and rethink.” On that note, he thinks it ’s time we considered how we can learn things from less data. There are large image databases avai lable now, but in terms of annotations, a doctor or a radiologist has to sit and mark the images, and these can be hard to obtain. It takes a lot of expert time and there are associated costs. What i f we didn’t need hundreds of thousands or even mi l l ions of annotated images? The f ield is moving in that direction, but it ’s not there yet, and Orcun Goksel is Assistant Professor at ETH Zurich. He leads the Computer-Assisted Appl ications in Medicine (CAiM) research group and was an area chair at MICCAI 2019. His main areas of interest are computer-assisted interventions and medical image analysis, particularly image reconstruction, and his group do a lot of work on ultrasound imaging and appl ications. He speaks to us about progress and innovation in his field. Orcun begins by tel l ing us that deep learning has been a game changer in many f ields. It started with image classi f ication, which takes an image and classi f ies it according to what ’s in it. Medical imaging extends this to segmentation and pixel -wise classi f ication, which looks at an image at a pixel level and is able to identi fy, for example, what organ it is. These are now common tasks, and with suff icient data and annotations, deep learning can usual ly solve them. However, there has been less success to date with other tasks such as image registration and image reconstruction. He points out that a lot of the work out there wi l l use any avai lable method and data to solve a