MICCAI 2016 Daily - Tuesday

MICCAI Daily: What kind of deep learning algorithm helped you in finding the optimal solution? Kevis: Usually, deep learning algorithms operate on patterns of images. There is usually heavy computational load because of redundant computations. We do it with fully convolutional networks that operate with the entire image at once so redundant computations are not needed there. It’s super fast. We also take advantage of pre- trained network which is a nice a treat from deep learning. We can take a pre-trained network on image net, and we just train for object detection. We can take all of these features learned on that and apply it to retinal images. There is a transferred learning, in a sense, of one field to the other, which is another powerful thing about deep learning. MICCAI Daily: What are the next steps in your model? Kevis: We would have to put some 3d information in the segmentation because when a surgeon operates in a retinal image, he wants the whole model. Ideally, we would have a 3d model of the eye, the vessels, and have all of the landmarks annotated and detected in real time. But this is in the future. MICCAI Daily: What did you find easy when working with retinal images, and what did you find more complicated? Kevis: The tricky part is, working with a vessel for example, they require very elongated structures that need to be segmented accurately. This is different from usual segmentation tasks. The structures are very complicated. Perhaps the easy part is that it’s always more or less in the same position. You don’t have to locate the eye. The camera always looks at the eye so that there is no detection part. MICCAI Daily: Do you find it more interesting to work on medical image processing or on general theoretical modeling? Kevis: To tell you the truth, I like working on general computer vision problems more: because of the big move with deep learning, now is the time for computer vision. There are a lot of things happening there that are not happening yet in the medical field. MICCAI Daily: Why is the medical field taking time to adopt the novelties that computer vision techniques have to offer? Kevis: Maybe it’s for a lot of reasons. There is not so much data. Usually, in the medical field, they don’t provide the code. There are more companies involved. Everything is more private because business has evolved so there is more protectiveness. That’s one of the things that we also try to change with this work, because everything is available, like code. We think sharing is important. I would also like to emphasize that we have everything online on our webpage . I would encourage people to use it freely. Presentation 7 MICCAI Daily: Tuesday