Retinal vessel segmentation based on self-distillation and implicit neural representation

Segmenting retinal blood vessels from retinal images is a crucial step in ocular disease diagnosis. It is also one of the most important applications and research in ophthalmic image analysis. However, the contrast between the retinal vessels and background in fundus images is low. The size and shape of retinal vessels vary significantly, and the width of some small vessels is often below 10 pixels or even 1 pixel. Moreover, some blood vessels are discontinuous owing to illumination, which complicates the segmentation of retinal blood vessels. To address these problems, this paper innovatively proposes a novel retinal vessel segmentation network framework based on self-distillation and implicit neural representation, which predicts retinal vessels in two stages. First, the self-distillation method extracts the main features of retinal images using the properties of Vision Transformer (ViT) to obtain preliminary images for the blood vessel segmentation. Second, implicit neural representation improves the resolution of the original retinal image, and the details of blood vessels are enhanced through the texture enhancement module to obtain accurate results of the blood vessel segmentation. Furthermore, we adopted an improved centerline dice (clDice) loss function to constrain the topology of blood vessels. We experimented on two benchmark retinal datasets (i.e., Drive and Chase) to quantitatively and qualitatively analyze the proposed method. The results indicate that the proposed outperformed the mainstream baseline. The visual segmentation results also show that this method can segment thin blood vessels more accurately and ensure the continuity of blood vessels.