Computer Vision News - February 2017

One can wonder: when you look at a standard retina image, segmentation between the black area at the top and the bright line at the bottom looks straightforward. So how can this be difficult, from an algorithmic point of view? This remark is quite true for images of a healthy eye; but when you try to do segmentation on pathological cases, the basic assumptions are not respected anymore. First, the upper layers, where you expect to see the highest gradient in the image, will have an unexpected look and the largest gradients may be found in other areas. That is also true for the brightest line, which you expect to see at the bottom of the retina, though you may find brighter lines on the top of it, like in the Outer Nuclear Layer (ONL) . Also reflexions and hotspots in the retina might look like the RPE. There may be also some shadowing, due to blockage of blood vessels, to interfere with the grey level of the lower part. These situations call for more sophisticated tools, which are mainly found in graph theory, like Graph Cuts , as well as Dijkstra and Bellman-Ford . These algorithms, in addition to snakes from active contours , balance between internal and external information: the latter is information coming from the image itself (like brightness and gradient); the former involves requirements of smooth boundaries between areas. Optimization methods are thus necessary to ensure the right balance when doing both at the same time. RSIP Vision ’s algorithms identify and segment the layers of the fundus in the images, producing an accurate measurement for each of the 9 areas of the retina. Blind tests showed our retina thickness measurement software to be more stable and accurate than competing software. This software for retinal thickness measurement provides a non-invasive, accurate and efficient method of assessing retinal health. Such a tool can help ophthalmologist determine the course of action in light of macular disease finding. Read on our website about this and other ophthalmology projects by RSIP Vision . 21 Computer Vision News Project “When you try to do segmentation on pathological cases, the basic assumptions are not respected anymore” Project “A non-invasive, accurate and efficient method of assessing retinal health”