Blood Vessel Segmentation using Multitask U-Net

Abstract

Thick and thin vessels are significant traits for diagnostic parameter identification in fundus disorders, and accurate segmentation of fundus vessels is an important part of the diagnosis of fundus complicacies. Advancement of the computing power of deep networks extensively employed to above problem. In this paper, a multitask U-Net has been proposed to address the highly imbalance of vessels to background ratio. Separate treatment has been employed to extract thick and thin vessels using multitask U-Net followed by a confusion network used for combining thick and thin vessel to get final vessel tree. Binary cross entropy and focal loss function has been employed to train the decoders for thin and thick vessel extraction. The proposed algorithm has been evaluated on DRIVE and CHASE DB1datasets and achieve accuracy of 0.976 and 0.969 respectively.