MICCAI 2016 Daily - Tuesday

MICCAI Daily: You are presenting your poster today at 16:00 (PS2-17). Can you tell us what is DRIU? Kevis: This is work for segmenting retina images. In particular, we segment both the vessels and the optical disc with deep learning methods. Our motivation was our project actually which works in robotic surgery. We want to cannulate the veins with a very tiny robot. We want something that can work in real time that a surgeon can actually use. We decided to do it with deep learning and bring speed to this field. MICCAI Daily: What is the problem today of doing it without deep learning? Kevis: Usually, there are faster methods, but they don’t perform as accurately as with deep learning. And since it needs to operate during surgery, the surgeon cannot wait for an image to be processed in minutes. He wants real time solutions. MICCAI Daily: What is the problem with current solutions? Kevis: They are usually slower. They also need pre-processing. Ours works on 10 hertz, so it’s 10 images per second. Other works take several seconds or even minutes to process a single image Also, we show in this paper that we performs better than humans in this task. MICCAI Daily: What is challenging in moving from a non-deep learning solution? Kevis: What is difficult is deep learning requires a lot of data, and we don’t have a lot of data. We need to be careful in how to train our algorithms, so that they don’t over feed. Actually, in the medical field, it’s a common case that data is not available everywhere. For a particular case, the annotations are very expensive. You have to validate every vessel. This has to be done by an expert, and we don’t have so many annotations for free. MICCAI Daily: In the practical application of your model, will it require more sophisticated machinery? Kevis: It will be expensive in the sense that it needs a GPU, which is usually expensive. Apart from that, it is common machinery. Input and Output CVPR Daily: Thursday Presentation 6 MICCAI Daily: Tuesday DRIU - Deep Retinal Image Understanding Kevis-Kokitsi Maninis