A new method for segmentation of retinal blood vessels using morphological image processing technique Roya Aramesh Faculty of Computer and Information Technology Engineering,Qazvin Branch,Islamic Azad University Qazvin, Iran Karim Faez Department of Electrical Engineering, Amirkabir University of Technology, Tehran, Iran Abstractct: Retinopathy caused by complications of diabetes, which can eventually lead to blindness. The lack of oxygen in the retina causes fragile, new, blood vessels to grow along the retina and in the clear, gel-like vitreous humour that fills the inside of the eye. For this reason it is important to extract retinal blood vessels. This paper introduces a new method for the detection of retinal blood vessels in retinal fundus images. The pre-processing technique was applied to input image in order to reduce the image noise and then, the image was divided into 16 smaller blocks. Afterward, a threshold was obtained for each block using maximum and minimum points of image histogram. Eventually, line detector filters and mathematical morphology was applied to the image and optimum results were obtained. Consequently, the proposed method showed an accuracy of 0.9480 with 0.7840 and 0.9826 sensitivity and specificity, respectively on the DRIVE database and better results have been achieved in comparison with existing methods. Keywords- line detector filters; blood vessels; image processing; Image blocking; Morphology operation I. INTRODUCTION Retinal vessel segmentation is important for the detection of eye diseases and plays an important role in automatic retinal disease screening systems. Automatic detection and analysis of the vasculature can assist in the implementation of screening programs for vessel diameter measurement in relation with diagnosis of hypertension [1], and computerassisted laser surgery [2]. Segmentation retinal anatomical structures are the first step in any automatic retina analysis system [3]. Detection of large vessels is relatively easy due to their strong contrast against background in the images but detection of small vessels is much more difficult due to their low contrast in the images. The centerlines are extracted by using the first order derivative of a Gaussian filter in four orientations and then evaluation of derivative signs and average derivative values is performed. The combination of morphological filters and crosscurvature evaluation to segment vessel-like patterns is employed by Zana et al. [5,6]. Mathematical morphology exploits the fact that the vessels are linear, connected and their curvature is a cross-curvature evaluation was performed to identify the structures in a retinal image whose curvature is linearly coherent. The present method is based on the maximum and minimum points of image histogram which to find detect suitable threshold for retinal thick and thin vessels. Several researches have been directed to find automatic detection algorithms from retinal images for the better diagnosis of retinal diseases. Unfortunately, fundus images may be blurred which affects the performance of automatic detection algorithms. The purpose of this paper is to design an efficient new algorithm that can exactly extract the retinal blood vessels from background, more accurate and easy for detecting of retinal diseases. This work is composed of five sections. Proposed method is described in section 2 including step by step methodology. Experimental results are given in section 3 and followed by conclusion in section 4. Fraz et al. [4] have proposed a unique combination of vessel centerlines detection and morphological bit plane slicing to extract the blood vessel tree from the retinal images. www.ijascse.org Page 1
II. A. Proposed method diagram PROPOSED METHOD The algorithm consists of four main steps: Pre-processing of the image, blocking image and calculating the maximum and minimum points, image filtering and then morphology operations on binary image. Flow diagram for our vessel segmentation system is shown Figure 1. Input RGB image B. Data base The template is used to format your paper and style the text. All margins, column widths, line spaces, and text fonts are prescribed; please do not alter them. You may note peculiarities. For example, the head margin in this template measures proportionately more than is customary. This measurement and others are deliberate, using specifications that anticipate your paper as one part of the entire proceedings, and not as an independent document. Please do not revise any of the current designations. C. Preprocessing Calculating maximum point Pre-Processing Image blocking Filtering Morphology operation Calculating minimum point Fig. 1 Flow proposed method diagram Output binary image I Gray Redchannel* 0.4 Greenchannel* 0.999 (1) Then, using a filter reduces out the noise. Following formula is used to filter Average: 1 f (x, y) I ( s, t) 20 ( s, t) S Gray (2) xy, D. Image blocking Input image size is 768 584 pixels, the gray image is divided into 16 blocks (I 1,..,I 16 ), and then a histogram is plotted for each block. Finally, using the following relations Suitable threshold for each block is determined. b 1 b 2 b 3= unique I i,i = 1,..,16 (3) H = histc I,b (4) i 1 Max histogram = Max(N(b 1,H)) (5) Minhistogram Min( N( b, H)) (6) 1 / Threshold Max Min Max (7) E. Filtering histogram histogram histogram In the filtering stage, line detector filters are used. These filters can be used to find one-cell-thick vertical, horizontal, or angled (135-degrees or 45-degrees) lines in an image [8]. Notice that line-finding is a similar application to edgedetection. Values in the input image are replaced with the average value of all valid cells within the kernel. This is also the procedure when the neighborhood around a grid cell extends beyond the edge of the grid. Filters can be multiplied by the optimal threshold to filter the image to reveal more details. After applying filters to optimize image blocks, the result is a normalized image block. 192 146 (8) ( I _ optimize 0) I _ optimizecd c1 d1 cd Then, averaging is the each normalize block image (I_normalized). F. Morphplogy operation Mathematical morphology is a nonlinear tool in retinal images analysis which has revealed itself as a very effective method for retinal diseases detection. Morphological opening operation is applied to smooth the background image and to highlight the retinal blood vessels. The closing operation produces objects that are not completely surrounded by retinal vessels, like things at branching points. B creates a flat, disk-shaped structuring element, where 1 specifies the radius. I _ normalized B ( I _ normalized B)! B (9) 8 I _ normalized Bi (( I _ normalized B)! Bi) B (10) i i1 www.ijascse.org Page 2
Top-hat filtering computes the morphological opening of the normalized image and then subtracts the result from the gray block image. For reduce noise and extract vessels more details Designed following four filters. These filters are properly extracted thin vessels. 1 1 1 1 1 1 2 2 2 1 1 1 1 1 1 1 2 2 2 1 the dimension of images in DRIVE database was 768 584. The experiments on 40 retinal images showed that the average performance of proposed method is 0.7840 in true positive rate (TPR) and 0.0174 in false positive rate (FPR). TPR is ratio between the number of pixels on our proposed method result images that also appear on the ground truth, and the total pixel number on the thinned ground truth. FPR is ratio between the number of vessel pixels on our proposed methods results but not on the ground truth, and the number of pixels on the background of the thinned ground truth. A RGB block and the output block image after operations different are shown in Figure 3. As shown in this figure, thin vessels and vessels are identified for each block. 1 2 2 2 1 1 1 1 1 1 1 2 2 2 1 1 1 1 1 1 Fig. 2 Four filters designed to reduce noise and extraction of thin vessels The most basic morphological operations are dilation and erosion with a structure element [9]. The dilation is extending the image boundary or the object of the image itself by adding pixels based on the structuring element design. The dilation operation was applied to a binary block image ( I_normalize i ). i=1,..,16 if the input image is I and the structuring element B, then[10] (a) I B { I _ normalize ( B) I } (11) i I _ normalize i Finally, after morphological operation, 16 binary block image obtained and the blocks connected together. Then the erosion operator applied on the image. The erosion operator creates erosion on all 1 values in the image. The structural element can be 1 or zero values. The center of structural element was placed on each image pixel and erosion operator is applied on a pixel based on structural element value. The image erosion of I, and structural element of B, are defined as follows: I! B { w ( B) I, w I} (12) This algorithm is to detect and properly diagnosed all main vessels, even the tiny vessels are also correctly diagnosed. (b) III. RESULT AND DISCUSSION The performance of presented algorithm is evaluated by applying it to DRIVE fundus image databases. The images of DRIVE database are mostly taken from healthy people, (c) (d) www.ijascse.org Page 3
(e) (f) (l) (g) (h) (i) (j) (m) (k) (n) Fig.3 (a)original image (b)green channel (c)preprocessing operation (d)a RGB block image (e)normalized block image (f),(g),(h),(i) image filtering (j),(k)morphological operation (l)connected blocks (m)proposed method result (n)manual segment www.ijascse.org Page 4
True positive rate Jan. 31 Image block histograms are shown in Figure 4. This figure is used to obtain maximum and minimum values of block histograms and the maximum and minimum values are used to define suitable threshold. Performance of presented algorithm was also evaluated with receiver operating characteristic (ROC) curve. These trends were generated by measuring the true and the false rate on all DRIVE database images. The ROC curve for the DRIVE database shown in Fig 5. False positive rate Fig5. ROC curve for the DRIVE database Table I shows the comparison of performance of this work with the other approaches in terms of average accuracy, sensitivity, specificity and FPR. TABLE I COMPARISON OF PERFORMANCE BETWEEN THE RECENT STUDIESAND THIS WORK Method Average accuracy Sensitivity Specificit y FPR Fraz et al. [4] Zana et al. [5],[6] This work 0.9430 0.7152 0.9769 0.0231 0.9377 0.6971 - - 0.9480 0.784 0.9826 0.0174 Fig4. Block histograms According to this method, the sensitivity, average accuracy and specificity in the diagnosis are 0.7840 and 0.9480 and 0.9826 respectively. IV. CONCLUSION A new algorithm for vessel extraction in retinal color fundus images was presented in this work. This approach is effective in medical and biomedical applications as automated retinal image analyses system. In this paper, the two methods of histogram maximum and minimum points and mathematical morphology were combined and used to www.ijascse.org Page 5
characterize the retinal blood vessels. The proposed method is applied for a database of 40 images and an accuracy of 0.9480 with 0.7840 and 0.9826 sensitivity and specificity, respectively on the DRIVE database obtained. The results of proposed method were compared to those obtained from existing methods and better performance has been achieved. In morphological operation needs two principal components, the maximum and minimum points of the histogram for these two principle component have been used. This method provides clear blood vessels of retinal images, suitable for detection retinal pathologies. For further study an efficient classification process can be considered for every image point in order to minimize or even eliminated some of misdetections of retina blood vessels. REFERENCES [1] J. Lowell, A. Hunter, D. Steel, A. Basu, R. Ryder, R.L. Kennedy, Measurement of retinal vessel widths from fundus images based on 2- D modeling, IEEE Transactions on Medical Imaging 23 (2004) 1196 1204. [2] J.J. Kanski, Clinical Ophthalmology, 6th ed., Elsevier Health Sciences, London, UK, 2007. [3] M. Foracchia, E. Grisan, and A. Ruggeri, "Detection of Optic Disc in Retinal Images by Means of a Geometrical Model of Vessel Structure," IEEE Trans. Medical Imaging, vol. 23, pp. 1189 1195, Oct. 2004. [4] M.M. Fraz, S.A. Barman, P. Remagnino, A. Hoppe, A. Basit, B. Uyyanonvara, A.R. Rudnicka, C.G. Owen, An approach to localize the retinal blood vessels using bit planes and centerline detection, Computer Methods and Programs in Biomedicine, http://dx.doi.org/10.1016/j.cmpb.2011.08.009, in press. [5] F. Zana, J.C. Klein, A multimodal registration algorithm of eye fundus images using vessels detection and Hough transform, IEEE Transactions on Medical Imaging 18 (1999) 419 428. [6] F. Zana, J.C. Klein, Segmentation of vessel-like patterns using mathematical morphology and curvature evaluation, IEEE Transactions on Image Processing 10 (2001) 1010 1019. [7] M. Niemeijer, J.J. Staal, B.v. Ginneken, M. Loog, M.D. Abramoff, DRIVE: digital retinal images for vessel extraction, http://www.isi.uu.nl/research/databases/drive, 2004. [8] Fraz, M.M., et al., Blood vessel segmentation methodologies in retinal images A survey. Computer methods and programs in biomedicine, 2012. 108(1): p. 407-433. [9] R. Gonzalez and R. Woods 1992 Digital Image Processing. Adison Wesley. [10] W. K. Pratt, Digital Image Processing, 3rd ed. New York: Wiley,2001. www.ijascse.org Page 6