Region-Based Segmentation Of Capillary Density In Optical Coherence Tomography Angiography

Microvascular changes are one of the early symptoms of retinal diseases. Recently developed optical coherence tomography angiography (OCTA) technology allows visualization and analysis of the retinal microvascular network in a non-invasive way. However, automated analysis of microvascular changes in OCTA is not a trivial task. Current approaches often attempt to directly segment the microvasculature. These approaches generally have problems in cases of poor image quality and limited visibility of the vasculature. Evaluating the quality of the results is also challenging because of the difficulty of manually tracing the microvasculature, especially in cases of low image quality or with images with a larger field of view. In this work, we develop an automated deep-learning approach to assign each pixel within human OCTA en-face images the probability of belonging to a microvascular density region of each of the following categories: avascular, hypovascular, and capillary-dense. The AUCs (area under the receiver operating characteristic curves) were 0.99 (avascular), 0.93 (hypovascular), and 0.97 (capillary-dense) for segmenting each of the categories. The results show very good performance and enables global and region-based quantitative estimates of microvascular density even in relatively low-quality en-face images.