A Convolutional Encoder-Decoder Architecture for Retinal Blood Vessel Segmentation in Fundus Images

A variety of retinal pathologies use fundus images to do non-invasive diagnosis through the analysis of retinal vasculatures. An Encoder-decoder architecture based on fully convolutional neural network for retinal vessel segmentation in fundus images, termed RetNet, is presented in this paper. RetNet consists of an encoder module as a contracting pathway to extract hierarchical features and a corresponding decoder module as an expansive pathway to reconstruct the full-size input. Particularly, RetNet integrates two different shortcut connections to capture more contextual and semantic information and can output more precise results without any post-processing techniques. The architecture is evaluated on the publicly accessible dataset of Digital Retinal Image for Vessel Extraction (DRIVE). Its comparisons with the ground truth and several state-of-the-art segmentation approaches including unsupervised and supervised methods show that RetNet can achieve strong performance on the limited medical dataset at a faster convergence speed.