Computer Vision News - April 2023

42 Computational Medicine group, Usher, Edinburgh are trying to get better at it. ” The device can compute features such as vessel density or the area free from vessels, known as the vascular area in the center of these images. However, changes in the vasculature can be observed in multiple diseases, and these basic features are insufficient for distinguishing between different conditions. “ We’re working on developing and implementing more and more metrics characterizing vessel morphology and structure, including tortuosity and caliber, and the size and shape of the intercapillary spaces, ” Ylenia reveals. “ Also, looking at the topology matrix and the repeated patterns in the network, specifically exploring sub-graphs in the vasculature to see how often we see a similar structure, for example. ” The idea is to use those measurements to discriminate a group of participants and establish a baseline control group for observing changes in specific diseases. The Computational Medicine Group has of structural and functional features to characterize the entire retinal landscape. “ The vasculature already has this network structure, so we’re using the graph theory concept of edges and nodes to reconstruct it, ” she explains. “ Using this structure, we can then compute features such as bifurcation points, and there could be nodes of degree three, for example, or we can measure the length of one edge, which would be the segment of one vessel between two bifurcation points. ” Ylenia tried different methodologies for segmenting the vasculature , surveying vessel enhancement methods and neural network architectures, and choosing the U-Net architecture because it worked best with her specific device. “ There are different OCTA devices commercially available, so there’s not one method that fits them all, ” she points out. “ Every year at MICCAI, more and more people are interested in the segmentation of the vasculature coming from these OCTA images, so other neural networks