An Efficient Segmentation of Retinal Blood Vessels using Singular Value Decomposition and Morphological Operator

The extensive study on retinal fundus images has become an essential part in medical domain to detect pathologies including diabetic retinopathy, cataract, glaucoma, macular degeneration,etc.which are the major causes of blindness. Automatic extraction of tree-shaped and unique retinal vascular structure from retinal fundus images is most exigent task and when achieved successfully, becomes a perfect tool helping ophthalmologists to follow appropriate diagnostic measures. In this work, a novel scheme to segment retinal tree-like vascular structure from retinal images is proposed using the Singular Value Decomposition’s left singular vector matrix of the weighted l*a*b* color model of the input image. The left singular vector matrix which captures the relevant and useful features helps in effective conversion of the input RGB image to gray image. Next, the converted gray image is contrast enhanced using CLAHE method which enhances the tree-shaped vasculature of the retinal blood vessel structure giving a rich contrast gray image. Further processing is carried out normalizing the contrast enhanced gray image by removing the image’s background using a mean filter by which blood vessels become brighter. Later, the result of the difference between gray image and normalized filtered image is keyed-in as a constraint to perform ISODATA thresholding which globally segments the foreground vasculature from the image’s background then followed by conversion of the resultant image into binary image upon which morphological opened operation is applied to take away small and falsely segmented portions producing accurate segmentation. This new technique got tested upon images contained in DRIVE and STARE databases and a performance metric called “area covered” is also calculated in addition with common metrics for sampled input image. This novel approach is empirically proven and has attaineda segmentation accuracy of 97.48%. Keywords—Singular value decomposition; left singular vector matrix; feature extraction; average filter; ISODAT Athresholding; morphological operators; stare anddrive databases