Blood Vessel Segmentation in Fundus Images Based on Improved Loss Function

摘要

In this paper, we improve the accuracy of segmentation of blood vessels in U-Net networks by improving the loss function. More specifically, we compare five loss functions which are suitable for image segmentation and study their effects on segmentation performance in U-Net networks. The optimal Focal + Dice loss achieves a vessel segmentation accuracy of 96.51% from the publicly available fundus image datasets DRIVE, better than original loss.