Detection of Retinal Vascular Bifurcation and Crossover Points in Optical Coherence Tomography Angiography Images Based on CenterNet

Optical coherence tomography angiography (OCTA) is a non-invasive imaging technique developed in recent years and has been used in ophthalmology to assist clinical diagnosis and treatment. Detecting the retinal vascular bifurcation and crossover points (feature points) in OCTA images is helpful for disease prediction, image registration and some other biomedical applications. In this paper, we construct an OCTA dataset with manually annotated vascular bifurcation and crossover points. In order to detect and classify these feature points, we first propose a method based on CenterNet, which adds attention gates (AGs) to the skip connection of the Stacked Hourglass Network. AGs can highlight valuable features in the input image to improve detection performance. Moreover, since we focus more on the coordinates of vascular feature points, we modify the traditional average precision (AP) and mean average precision (mAP) by calculating the Euclidean distance between two points rather than the intersection over union (IOU) of two bounding boxes. Experiments indicate that our method can achieve 80.81% AP for bifurcation points, 85.86% AP for crossover points and 83.34% mAP.