Weight Averaging Impact on the Uncertainty of Retinal Artery-Venous Segmentation

Abstract

By examining the vessel structure of the eye through retinal imaging, a variety of abnormalities can be identified. Owing to this, retinal images have an important role in the diagnosis of ocular diseases. The possibility of performing computer aided artery-vein segmentation has been the focus of several studies during the recent years and deep neural networks have become the most popular tool used in artery-vein segmentation. In this work, a Bayesian deep neural network is used for artery-vein segmentation. Two algorithms, that is, stochastic weight averaging and stochastic weight averaging Gaussian are studied to improve the performance of the neural network. The experiments, conducted on the RITE and DRIVE data sets, and results are provided along side uncertainty quantification analysis. Based on the experiments, weight averaging techniques improve the performance of the network.