Automatic localization of retinal neovascularization in optical coherence tomography angiography images using feature maps analysis

Retinal neovascularization (RNV) is a pathological condition characterized by the abnormal growth of new blood vessels within the retina. Detecting RNV can be achieved through optical coherence tomography angiography (OCTA), an advanced imaging technology capable of visualizing blood vessels under retinal layers. Identifying RNVs within OCTA images is particularly challenging due to varying patterns, unfixed sizes, and being on a vascular network background. Due to domain newness and significant challenges, there are a few studied on methods to identify the RNV location in OCTA. In this thesis, we studied how to localize the RNV location from features maps: vessel density, bifurcation points, and average vessel thinness per area. We considered regions of interest (ROI) by thresholding a weighted linear combination of those feature maps. We selected the final RNV localization from the maximum value of the combined map within the most considerable ROI. We compared this approach with variants of three used feature sets and two combination methods. The weighted linear combination of all feature maps achieved the highest localization accuracy of 69.05% on the 43 OCTA images with RNV. It demonstrates superior performance to other variants.