Ret2Ret: Retinal Blood Vessel Extraction via Improved Pix2Pix Image Translation

High prevalence of retinal diseases is a global concern, but the recent advances in artificial intelligence has brought in new hopes of devising automated tools for detecting and monitoring critical retinal diseases even at their onsets. An accurate segmentation of the blood vessel structure from the retinal fundus image is a basic prerequisite for subsequent disease diagnosis in any computerised retinal disease diagnosis system. Although the said research area has been well studied in the past few decades, there are still scopes for improvement particularly on pathological images, since recent research has continually expanded the list of diseases that have early markers in retinal images. In this paper, we exploit the power of deep generative adversarial networks (GAN) in extraction of retinal vessels by proposing Ret2Ret which is a modified Pix2Pix GAN model for image to image (retina-to-retina) translation. The generator module of the backbone Pix2Pix model has been re-designed into a light-weight architecture having fewer parameters. In addition, we have introduced the bi-FPN architecture in the generator for an accurate extraction of thin vessels. Results show that our proposed Ret2Ret method outperforms a number of competing recent approaches on public benchmark databases like DRIVE and CHASE DB1.