Retinal vessel segmentation using modified symmetrical local threshold

Retinal vessel segmentation is important for the identification of many diseases including glaucoma, hypertensive retinopathy, diabetes, and hypertension. Moreover, retinal vessel diameter is associated with cardiovascular mortality. Accurate detection of blood vessels improves the detection of exudates in color fundus images, as well as detection of the retinal nerve, optic disc, or fovea. A retinal vessel is a darker stripe on a lighter background. Thus, the objective is very similar to the lane detection task for intelligent vehicles. A lane on a road is a light stripe on a darker background (i.e. asphalt). For lane detection, the symmetrical local threshold (SLT) is found to be the most robust feature extractor among the tested algorithms in the road marking (ROMA) dataset. Unfortunately, the SLT cannot be applied directly for retinal vessel segmentation. The SLT is a 1D filter and is designed for detecting vertical or close to vertical light stripes with predictable width. In this paper, the SLT is modified to detect dark stripes and four kernels, instead of one, are designed to detect both vertical and horizontal features of a retinal vessel with variable thickness. The proposed algorithm is tested using the High Resolution Fundus (HRF) image database and the accuracy is estimated to be 95.53%. Furthermore, when tested with the Digital Retinal Images for Vessel Segmentation (DRIVE) database, the accuracy is estimated to be 93.69%.