W-net: A Convolutional Neural Network for Retinal Vessel Segmentation

In this paper we propose a method for retinal vessel segmentation based on a multi-stage deep convolutional neural network with short connections. The proposed method is a two-stage application of an improved U-net architecture. In the first stage, a probability score for the vascular structure presence is computed from a set of random patches taken from the image dataset. In the second stage, this probability is refined to obtain a final threshold image of the vessel structure. The main contributions of this paper are the following: (1) We propose a modification for the distribution of weights in the U-net, called here the V-net model, which is more convenient for reconstruction tasks. (2) We propose a multi-stage version of our model, called here the W-net, and we conduct extensive experimental evidence in which the W-net produces high-quality results for retinal vessel segmentation. (3) We also propose a fast operating version of the W-net, and evaluate potential improvements when modify our proposal. We evaluate the performance of our methods in various public available datasets, and compare our proposal versus other recently developed methods. The experimental results demonstrate the capabilities and potential of our proposal.