Research On Blood Vessel Segmentation Methods For Fundus Image And Its Implementation

Abstract:Fundus image vascular analysis is the most important basis to diagnose eye diseases and other systemic diseases such as diabetes and hypertension. With the rapid growth of fundus image data, the diagnosis depending on artificial observation and experience is low efficient and strong subjective. Therefore, using computer to detect and segment the vascular network in fundus image has vital clinical significance.Large number of existing segmentation algorithms have achieved a relatively good performance, however, they do not work well for pathological retinal images and there are still many bottleneck problems such as the lost of small vessels, vascular intersections and bifurcations. In addition, automatic blood vessel segmentation methods require manual labeled standard images to verify the performance, but manual labeling task is very laborious. In order to solve these problems, this paper based on color fundus image studies the automatic and interactive retinal blood vessel segmentation method. The research work of this paper is carried out in the following two aspects:In view of the poor robustness of existing automatic retinal blood vessel segmentation methods, this paper proposes a new automatic retinal blood vessel segmentation method based on the divergence of gradient vector field and the morphological bot-hat transform. Firstly, the vessel centerlines with no noise and lesions are extracted by using the divergence of the normalized gradient vector field. Secondly, the main vessels are segmented by a sequence of bot-hat operators with different scales and directions. The vessel pixels too far from the centerlines will be removed as noise or lesions. Finally, a repair procedure is performed to regain the pixels at the positions of vascular intersections and bifurcations through comparing the number of connected regions before and after denoising.Considering the dual contradiction between automatic segment-ation methods and manual labeling task, this paper proposes an interactive retinal blood segmentation method by a modified region growing algorithm based on equi-gradient distance. The method allows user to label blood vessel and background in some region in the image. Then the method calculates the equi-gradient distance map based on the labels and set the labeled pixels as seed points to implement bi-directional region growing in the equi-gradient distance map. Subsequently, an edge optimization process based on centerline self-adaption and neighborhood self-adaption is executed for the segmented result by region growing and the final segmentation result is obtained.This paper adopts the publicly available DRIVE database and STARE database to test the proposed methods. Experiments show that the automatic blood vessel segmentation method based on divergence and bot-hat transform can obtain higher accuracy, sensitivity and specificity. The method is superior to other existing algorithms and has a good robustness for the pathological retinal images. Experiments also show that the interactive blood vessel segmentation method based on equi-gradient distance can meet the real-time requirement and can obtain the accuracy which is consistent with the manual segmentation result. In addition, the interactive segmentation method increases the labeling efficiency significantly.