Computer Vision News - September 2022

38 Audience Award WBIR 2022 contrast, and it is likely to give you different results when you run different algorithms . “ We want to predict this uncertainty , so we have a method which says, for instance, we have an uncertainty of three millimeters on the vector field in this region, ” Andreas describes. “ Then we want to use that uncertainty inside our treatment. In my conference talk, I focused on the fact that it’s one thing to predict the uncertainty of a vector field, which is already super difficult, but you also want to be able to translate it into something we can use in the clinic , which we can plan upon, and which can identify where we need manual intervention. For example, if there is a region with an uncertainty of one centimeter, it would flag that it would be best if a doctor checks what’s happening. That would trigger manual intervention in this five-minute window. ” Andreas demonstrated to the WBIR audience that they could accumulate the dose, not just for one deformation, but they could take the uncertainty into account so that they have it on the accumulated doses. With such an application-oriented presentation, does he think this is what earned him the audience prize? they will give you two different outcomes. “ Let’s say we run one algorithm we know works relatively well, ” Andreas proposes. “ RayStation, for instance, has a rather good deformable image registration algorithm. Then we want to know, how certain are we about the solution? ” In some regions, where there is a lot of image contrast –around the spinal cord, for example – there is not much deformation. Running five different algorithms will give you the same result, and you can be pretty certain about the solution. In other regions, where there is not as much contrast on a CT scan, such as inside the lung or beneath the diaphragm, much deformation combines with not so much