Optical coherence tomography (OCT) is revolutionizing the way we assess eye complications such as diabetic retinopathy (DR) and age-related macular degeneration (AMD). With its ability to provide layer-by-layer information on the retina, OCT enables the early detection of abnormalities emerging underneath the retinal surface. The latest advancement in this field, OCT angiography (OCTA), takes this to the next level by providing detailed vascular information without requiring dye injections. One of the most significant indicators of DR and AMD is neovascularization, the abnormal growth of unhealthy vessels. In this work, the techniques and algorithms used for the automatic detection, classification, and segmentation of neovascularization in OCTA images are explored. From image processing to machine learning and deep learning, works related to automated image analysis of neovascularization are summarized from different points of view. The problems and future work of each method are also discussed.
Retinal neovascularization (RNV) is a pathological ocular disorder marked by the development of abnormal blood vessels within the retinal tissue, leading to potential complications such as hemorrhage and vision loss. It can be visualized using Optical Coherence Tomography Angiography (OCTA), a specialized imaging tool that allows for the detailed analysis of retinal blood vessels. However, identifying RNVs within OCTA images poses challenges including variability in patterns and sizes, overlapping vascular networks, complex morphology, noise and artifacts, low contrast, depth variability, and interpatient variability. This study addressed this challenge by extracting vessel density, bifurcation points, and vessel thinness features to pinpoint regions of interest. Our systematic approach led to selecting potential regions using a support vector machine (SVM). Notably, our method achieved a localization precision of 90.33% in two datasets comprising 69 en face OCTA images, demonstrating a superior F1-score than the state-of-the-art methods, which used VNet and the weighted combination of feature maps by 44.19% and 57.46%, respectively.