A Retinal Vessel Segmentation Method Considering Small Vessels Using Gray-voting And Gassian Mix Model

As the only microvasculature that can be directly observe noninvasive, the vessel segment of retinal fundus image is the preliminary step to the clinical diagnosis for some eye diseases. Many diseases such as hypertension, diabetes, and hardening of the arteries may cause structural change of the fundus retinal blood vessels. The segmentation of retinal blood vessels are segmented by physician manually, and the efficiency is extremely low.Many methods have been proposed to retinal vessel segmentation. However, most of the existing methods can only segment the main vessel structure, and the affection of small vessels segmentation is far to satisfaction. We present a method for vessel segmentation of the retinal fundus images. A regional gray enhance algorithm called gray-voting is employed in the small vessels extracting and two-dimensional Gabor wavelet is used to extract the main vessels. Then we fused the gray-voting result with the2D-Gabor filter result as the pre-process outcome. Gassian Mix Model is employed to obtain the vessel cluster form the pre-process outcome. Small vessels fragments are obtained by another gray-voting process, which would be using to complement the vessel cluster. At last eliminate the fragments that not belong to vessel according the shape of the fragments. Our approach shows good affection in the detection of both think and small retinal vessels.In this paper we evaluated the approach on two publicly available DRIVE (Staal et al.,2004) and STARE(Hoover et al.,2000) datasets of manual segmented results. To the DRIVE dataset, our approach obtained average ACC0.9418and SEN0.7359. And to the STARE dataset our approach obtain average ACC0.9364and SEN0.7769by using the first manual segmented result as the evaluation standard. When using the second manual segmented result as the evaluation standard, our approach obtain average ACC0.9214and SEN0.6502. The SEN of our approach is much higher than the other proposed methods, and the manual segmented result contains more small vessels is more suitable for the algorithm evaluation of our method.