Retinal Vascular Segmentation Network with Connectivity Guidance

The morphology of the retinal vasculature is closely related to a variety of ophthalmic diseases. Despite significant progresses have been achieved in retinal vessel segmentation with the development of deep learning, several challenging problems still persist. Specifically, some components present in the retina (such as optic discs or lesions) may interfere with or cover the vessels, and current segmentation models rarely focus on the connectivity of the vessels. In this paper, we propose an efficient network to conquer this issue. Firstly, we propose a semantic enhancement module, to insert into the bottom layer of the UNet encoder section for guiding the network to enhance features representation. Secondly, we introduce a skeleton fitting module that maximizes the preservation of the vascular topology using skeleton prior and contrast loss. Finally, we perform a morphological complementation operation on the final output of the decoder by means of a voting mechanism. The designed segmentation network has been evaluated on publicly available datasets and experimental results demonstrate the effectiveness of the method.