Segmentation of Blood Vessels from Digital Fundus Images

Ocular fundus image assessment has been extensively used by ophthalmologists for diagnosing vascular and non vascular pathology. Examining the retinal blood vessel network may reveal arteriosclerosis, diabetes, hypertension, cardiovascular disease and stroke [12]. Furthermore, the segmentation of the vessel network is the most suitable representation for the retinal image registration since vascular tree does not change except in a few diseases and includes adequate information for the identification of some anchor points. In addition, vessel tree can also be used as a land mark feature for image-guided laser treatment of choroidal neovascularization. Therefore reliable methods for segmentation of blood vessels in fundus images are needed. The methods used for blood vessel segmentation discussed in Chapter-2 can work well to segment the major parts of the blood vessels. However, the major challenges confronting the vessel segmentation methods which are shown in Fig. 3.1are:  Segmentation of the thinner blood vessels as the image contrast is normally low around thin blood vessels;  The presence of pathologies as they may be mis-enhanced and mis-detected as vessels. In order to solve these problems, Histogram Matched Local Relative Entropy (HMLRE) method is developed to segment blood vessels in fundus images. For efficient detection of vasculature, high contrast between vessel network and the fundus background is desired while there must be low contrast between the fundus background and retinal pathologies [93]. In red channel of a RGB colour image, the gray levels are spread over a wider range compared to green channel. Therefore the contrast between bright pathologies and fundus background is less in red channel. Thus in HMLRE method, red channel " s intensity information is used in pre-processing of colour fundus images. The histogram of green channel is modified by employing the histogram of red channel (of the same fundus image) to Fig.3.1. Challenges in Extraction of Retinal Vasculature having Severe Pathology from STARE Database. Arrows drawn on the image in black dashed lines show lesions and the boundary of the optic disc., Arrows drawn in white lines highlight narrow blood vessels in low contrast regions. obtain a new image in preprocessing. The contrast of vasculature against the background of the preprocessed image is improved by using matched filter. The local relative entropy thresholding with histogram compression and translation is employed to discriminate blood vessel segments from the background in the matched filter response. The misclassified pixels are then removed by using label …