Retinal Blood Vessel Extraction by Using Multi-resolution Matched Filtering and Directional Region Growing Segmentation

A new method to extract retinal blood vessels from a colour fundus image is described. Digital colour h d u s images are contrast enhanced in order to obtain sharp edges. The green bands are selected and transformed to correlation coefficient images by using two sets of Gaussian kernel patches of distinct scales of resolution. Blood vessels are then extracted by means of a new algorithm, directional recursive region growing segmentation or D-RRGS. The segmentation results have been compared with clinically-generated ground truth and evaluated in terms of sensitivity and specificity. The results are encouraging and will be used for fiuther application such as blood vessel diameter measurement. David usher2 , James F. Boyce Department of Physics, King's College London therefore required as a pre-processing component of an automatic diagnosislscreening system. In this study, an effective segmentation algorithm for the classification of retinal blood vessels is described. Digital fundus images are colour standardised and contrast enhanced in order to obtain sharp edges in a consistent way. The green bands of the images are then transformed to correlation coefficient images by using two sets of Gaussian kernels. A new algorithm named Directional Recursive Region Growing Segmentation or D-RRGS is applied to the correlation coefficient images to extract blood vessels. The segmentation results are compared with clinically generated vessel segmentations and assessed in terms of sensitivity and specificity of vessel identification. The results are 1. Preface encouraging and will be used in fiuther applications such as Retinal blood vessels are one of the most important blood vessel diameter measurement. components in ophthalmic diagnosis. As the network of the retinal blood vessels doesn't change very much over time, 2. Algorithm the location of bifurcation andlor end points can be used as Pre-processing feature points to identify an eye. Detecting abnormalities This process includes local contrast enhancement and green such as venous looping or beadings is critical for early band extraction of the digital fundus image. (If necessary, treatment, as they are, in most cases, indications of colour standardisation to reduce the variation of retinal potentially sight-threatening retinopathy[l]. In order to colour found in a racially heterogeneous patient sample may utilise these u se l l characteristics of retinal blood vessels, it be applied.) is very important to obtain their locations and shapes The original image is converted from the RGB colour model accurately. to HSI. The I band of the resulting image is processed by In many of the reported studies on automatic fundus image local contrast enhancement. Local contrast enhancement is a analysis and diagnosis[2][3][4] normal components within signal to noise enhancement process to create images which the image, such as blood vessels or fovea, are detected and are improved for the subsequent retinal blood vessel identified before starting abnormal component detection. detection. It emphasises the local contrast of the intensity Pathologies such micro aneurysms or haemorrhages, located values of an image so that the blood vessels are more clearly close to blood vessels, may be misclassified as blood vessels distinguished from the background. It is then converted back and removed in the pre-processing, resulting in reduced to RGB colourmodel. specificity of pathology detection and hence possible From visual observation, blood vessels generally exhibit the misdiagnosis. Accurate retinal blood vessel extraction is greatest contrast from the background in the green band and 1Address: 1 Ikegami, Owariasahi-city, Aichi, 488-8501, Japan. E-mail: [email protected] 2Address: Strand, London WC2R 2LS, UK. E-mail: [email protected] 3Address: Strand, London WC2R 2LS, UK. E-mail: [email protected] Fig.21: Pre-processing (a) Original image (multi-band) (b) Contrast enhanced image (multi-band) (c) Contrast enhanced image (green band) therefore the green band is selected from the contrast where enhanced images for further processing. M N 1 M N The stages of pre-processing for a typical fundus image are I , ) pp = -c c p( i , i ) (2.3) p~ = ,=, i=, NM ,=, i= l shown in Fig.21. Gaussian kernel patches The original range of t he correlation coefficient is In order to cope with a large of blood vessel widths, [-l , l l , however, it is to convert the range two sets of kernel patches were created; the first set of [092551 in order a 256-t0ne greyscale kernels (set A) is for medium to thick blood vessels and the image as Output. other set (set B) is for fine vessels. Each set consists of The conversion is linear and the equation used for the twelve kernel patches of different angle. creation of the intensity is The suitable two dimensional kernel K(x,y) may be Intk,y) = (thy)+ 1) x127.5 (2.5) mathematically described as M/2 XI W-M/2, N/2 < y < H-N/2 (2.6) where Int(x,y) is the intensity value for the output image and K(x,y)=-expe 1 2 4 for 1 y 1 L/2 (2.1) c(x,y) is the correlation coefficient at the coordinate (x,y) where a is the variance of the intensity profile, L is the where the centre point of the kernel patch matches. length of the segment for which the blood vessel is assumed to have a fixed orientation. In this study, o = 2.0 and L=9 are employed as suggested by Goldbaum et a1[5]. Correlation coefficient images creation The correlation coefficient images used in this paper are created by the following algorithm. Let I(m,n) ,P(m,n) be the input image and the kernel patch respectively, and WxH, MxN be the size of I(m,n) and P(m,n) respectively, then the comelation coefficient at (x,y) in the input image is given by