CBAV-Loss: Crossover and Branch Losses for Artery-Vein Segmentation in OCTA Images

Obvious errors still exist in the segmentation of artery-vein (AV) in retinal optical coherence tomography angiography (OCTA) images, especially near crossover and branch points. It is believed that these errors occur because the existed segmentation method cannot effectively identify the crossover and branch points of AV. In this study, we proposed a Crossover Loss and a Branch Loss (CBAV-Loss), which are two novel structure-preserving loss functions. By restricting the crossover and branch points of arteries and veins, the segmentation accuracy can be improved by correcting the segmentation errors near the crossover and branch points. The experimental results on a manually annotated AV dataset with 400 OCT and OCTA cubes demonstrate that the crossover and branch losses can effectively reduce errors for AV segmentation and preserve vascular connectivity to a certain extent.