A framework for hierarchical division of retinal vascular networks

Abstract

Human retinal vascular network plays an important role in ophthalmology diagnosis. For example, in the diagnosis of ophthalmology, the severity of the disease has a direct correlation with the lesion location, in the sense that the closer the lesion area is to the optic disk, the higher will be the severity of the disease. If a framework is able to provide the hierarchical structure of the retinal vascular network, then the severity of disease can be quantified by leveraging the hierarchical characteristics of vessels in the vicinity of the lesion location. Thus, in this paper, an executable framework is recommended for the hierarchical division of the retinal vascular networks. Specifically, a supervised method based on deep neural network is used for retinal blood vessel segmentation. A graph-based method is also applied to generate vascular trees from the segmented retinal vessels. As part of our proposed approach, we present two algorithms: the potential landmark detection algorithm (PLDA) is used to identify the bifurcations and crossings; and the adaptive hierarchical classification algorithm (AHCA) is used in the hierarchical characteristics classification of vascular bifurcations. By classifying the hierarchical characteristics of vascular bifurcation, the hierarchical characteristics of the vessel segments containing these bifurcations are identified. Thus, the hierarchical division of retinal vascular network is realized. When applied to two publicly available datasets, DRIVE and STARE, the proposed framework achieves an accurate rate of 98.99% and sensitivity rate of 92.17%.