A hybrid deep segmentation network for fundus vessels via deep-learning framework

Abstract

High-precision segmentation of fundus vessels is a fundamental step in the diagnosis and treatment of fundus diseases, in which both thick and thin vessels are important features for symptom detection. With the rapid development of artificial intelligence, the deep convolutional neural network (DCNN) has been widely applied into image analysis of fundus vessels. Nevertheless, due to the imbalanced ratio between thick and thin vessels, the existing segmentation methods are weak in the task of microvessel extraction from fundus images. To address this problem, this paper proposes a new hybrid deep image segmentation method for fundus vessels that consists of a multitask segmentation network and a fusion network. For the proposed method, a multitask segmentation network is developed to precisely segment both thick vessels and thin vessels from fundus images separately. In addition, an effective loss function is designed to adapt to the two different vessel segmentation tasks and ultimately solve the imbalanced ratio between these two vessels. Furthermore, an improved U-net network model is proposed to serve as the basic segmentation network to ensure the segmentation performance of the multitask segmentation network. Together with these networks, a fusion network is also proposed to fuse these two kinds of blood vessels to obtain the fusion images as the final segmentation results of fundus vessels. The proposed segmentation method is validated on many different public data sets of fundus images, such as DRIVE, STARE and CHASE_DB1. Experimental results show that the proposed method obtains a better segmentation performance on fundus images and acquires a higher recall, F_1 value, and accuracy than other advanced segmentation methods.