Multi-Scale Retina Vessel Segmentation in OCTA with a Vascular Connectivity Module in the Convolutional Neural Network

The segmentation of retinal blood vessels in optical coherence tomography angiography (OCTA) is of great importance for the diagnosis and treatment of various diseases such as diabetic retinopathy and dementia. Currently, UNet is one of the classical and popular networks in the segmentation field. Although significant progress has been achieved with the rapid development of UNet-based neural networks, some critical issues in retinal vessel segmentation remain unsolved. First, blood vessels in OCTA show large variations in length and width, imposing challenges in identifying the small vessels at the ends. Second, the vessels should be continuous and smooth, and the capillaries should not detach from the main vessels. Nevertheless, the current UNet-based neural networks lack the capability to preserve the shape of prior information. This study introduces a modified UNet framework for retinal vessel segmentation using OCTA images. First, multi-scale learning modules are employed to improve the ability of the network to extract multi-scale vessel objects. Then, we introduce a novel vascular connectivity module in the network to incorporate prior shape information. The proposed method id extensively evaluated on a public dataset named OCTA500, with significantly improved performance compared with the state-of-the-art methods.