Improvement and restoration of retinal images are vital for clinical applications, from abnormality classification through segmentation to automated medical diagnosis. The major problem of retina restoration is estimating the image regions obscured by unwanted features, of which the most significant culprit is the blood vessel network. The challenge lies in the unavailability of true, unobstructed images. The commonly used methods apply masked filtering, dictionary-based approaches, or rely on an innate ability of machine learning models to deal with blood vessels. To solve the blind blood vessel inpainting problem, we propose a convolutional network architecture with multi-level self-attention capable of learning both segmentation and inpainting of blood vessels in retinal images. Furthermore, we introduce an efficient training method for the inpainting task with unknown ground truth. Our focus is on the optic nerve head region, which is essential in fundus analysis and glaucoma diagnosis. Our approach surpasses the state-of-the-art methods in blood vessel inpainting on the examined data while being trainable on personal computers. We examine the accuracy of vessel segmentation and the quality of inpainted images produced by our approach. The results show a statistically significant increase in segmentation accuracy of traditional methods after inpainting. In conclusion, we present a reliable vessel removal method applicable as a crucial first step in retinal segmentation, in the shape and color analysis of separated retinal vessels and background, in blood vessel detection, or in generating clear retinal background for generative methods.
