MICCAI 2016 Daily - Wednesday

healing process through ultrasound, then this could be much more useful and more preferred. MICCAI Daily: When called to distinguish between soft tissue and hard tissue, how does ultrasound behave? Firat: There are quite a few characteristics of bone in ultrasound. This is actually what we base our paper on as well. You will see a lot of different reflections on different layers of soft tissue. This is actually the main drawback of bone segmentation algorithms that are currently present in the literature. They could look like bone surface, but one very distinctive case for bone is that there is a shadow under the bone surface because the ultrasound signal cannot penetrate further. We make use of this feature for example. Another thing is that, although the bone surface is extremely thin, on the ultrasound it creates a hyperechoic band which has much more than a single pixel in thickness. So we also make use of this. MICCAI Daily: What computer vision techniques did you use to provide a solution to this problem? Firat: On a very simple basis, we use machine learning approaches to actually have some probabilities for different tissues where we separated as soft tissue, bone surface, and shadow that is under the bone surface. We use a novel graphical modeling where we use machine learning approaches for the unary costs and pairwise costs would be between connected different regions or pixels. What we do is, instead of looking at this intensity image, we somehow try to encode this. Say I’m thinking about soft tissue. What is the likelihood of a particular pixel to be soft tissue? What is the likelihood of a pixel to be shadow? Then we can say, in ultrasound, bone surfaces look like this. We create a graph structure that will follow this structure. Then we optimize this. MICCAI Daily: What is the next step of your model? Firat: We didn’t focus too much on perfecting the features that we used in machine learning. This creates possibly some unnecessary computation. We actually want to do some feature selection, and come up with more accurate features that can get the same or better classification. This will help us to make it a somewhat real-time application which is what we are lacking now. Once it’s a real time application, then it can be easily carried on to the hospitals where our research will meet the patients. Presentation 17 MICCAI Daily: Wednesday