Impact of Increased Centerline Weight on the Joint Segmentation and Classification of Arteries and Veins in Color Fundus Images

Abstract

The analysis of the retinal vasculature represents a fundamental step in the diagnosis of multiple diseases, both ophthalmic and systemic. A comprehensive analysis includes the segmentation of vessels as well as their classification into arteries and veins. So far, multiple deep learning-based approaches have emerged that perform both tasks jointly. Currently, the state-of-the-art works are based on fully convolutional neural networks. In these works, the joint problem is approached either as a semantic segmentation task (SST) or as a set of multiple segmentation subtasks (MSS). These subtasks usually target arteries, veins and vessels. Unlike the SST approach, the MSS approach gives raise to complete segmentation maps. To address the low performance in the segmentation of small vessels, a state-of-the-art work proposes a SST approach that uses custom weight masks to increase the importance of vessel centerline pixels. In this work, we study the impact of increasing the importance of the central pixels of the blood vessels in training deep neural networks following the MSS approach. The experiments conducted in a public dataset demonstrate that increasing the weight of vessel centerlines improves the segmentation of small vessels, but worsens the overall segmentation. Thus, increasing the centerline weight is actually a relevant trade-off to be taken into account.