Edge Detection of the Optic Disc in Retinal Images Based on Identification of a Round Shape Thanapong Chaichana*, Sarat Yoowattana*, Zhonghua Sun†, Supan Tangjitkusolmun*, Supot Sookpotharom+, and Manas Sangworasil* *

Faculty of Engineering, Department of Electronics, King Mongkut’s Institute of Technology Ladkrabang 3 Moo 2, Chalongkrung Road, Ladkrabang, Bangkok 10520, Thailand. † Discipline of Medical Imaging, Department of Imaging and Applied Physics, Curtin University of Technology GPO Box, U1987, Perth, Western Australia 6845, Australia. + Electrical Engineering Department, School of Engineering, Bangkok University, Pathumtani 12120, Thailand. E-mail: {kcthanap,s0060402,ksamanas,ktsupan}@kmitl.ac.th, [email protected], [email protected]

Abstract— This paper presents a novel method for identification of the position of the optic disc in retinal images. The method is based on the preliminary detection of the main edge detection of retinal image. The segmentation optic disc is estimated as a circular area. We searched for areas of optic disc using Hough transform which detected several straight lines and approximated them as a circular line. The position of optic disc was identified as the center of the circle. An evaluation of the proposed procedure was performed using a set of 40 images from the DRIVE database, containing images from both healthy (33) and diseased (7) subjects. The optic disc position was correctly identified in 39 out of 40 images and the efficiency was 97.5 percent. The results illustrate that the proposed method can be used for pathological analysis of retinal images. Keywords: Edge detection, Hough transforms, retinal images, optic disc (a)

I.

INTRODUCTION

The optic nerve (ON) conducts impulse signals from retina to the brain. ON sits in the anterior region of the eye ball and it is close to the macula. The optic disc (OD) is the distal end of ON which is adjacent to the eye ball. Fig. 1 (a) shows an anatomical sketch of the eye ball and its structure, while Fig. 1 (b) demonstrates the locations of OD and macula in a retinal image. The position of OD in retinal image is essential for diagnosis of retinal abnormality. Example cases of OD diseases in retinal images are glaucoma disc, optic atrophy, disc drusen, high myopia, optic disc pit, etc. In these cases, abnormalities can be seen in OD area of retinal images. Identification of OD position in the retinal image has been performed in previous studies. Hoover and Goldbaum [1] proposed an algorithm based on a novel algorithm or fuzzy convergence. They identified positions of ON as the main point in retinal blood vessels network. The main point is defined as the location where several retinal blood vessels converge (Fig. 1 (b)). They determined the location of OD after normalizing the brightness of the images. The position of OD was found at the brightest spot in the retinal image.

Macula Optic Disc

Main point (b) Fig. 1. The structure of retina: (a) Sagittal section of the eye, (b) The retinal image showing the optic disc and the macula.

978-1-4244-2336-1/08/$25.00 © 2008 IEEE 670

2008 International Symposium on Communications and Information Technologies (ISCIT 2008)

Foracchia et al. [2] proposed an algorithm based on model of geometrical direction of blood vessels system. They identified position of OD as the center of cross sectional parabola. Huiqi and Chutatape [3] presented an automatic method for identifying the position of OD. They separated pixels with high brightness from the original retinal image and applied the principal component analysis (PCA) to determine the position of OD. Zheng et al. [4] presented a method for analysis of retinal image for classification of retinas with diabetic retinopathy from healthy retinas. They identified fovea and OD using Hough transform technique and extraction of blood vessels in retinal images using Gaussian filter. Thongnuch and Uyyanonvara [5] proposed an automatic method to identify the position of OD from low intensity infant retinal image. They reduced the dimensions of Hough transform to two and used its histogram for estimation of OD radius. Sinthanayothin et al. [6] proposed an automatic computer technology to analyze main components of retinal images. They applied adaptive local contrast enhancement technique to retinal images, and identified the position of OD by variance measurement between the intensity of each pixel and intensities of adjacent pixels. Blood vessel segmentation was analyzed by neural network multilayer perception model, and the identification of fovea was performed by matching correlation technique of dark pixels near OD. Finally, Corona et al. [7] proposed digital stereo image analyzer for generating automated three-dimensional measures of optic disc deformation in glaucoma. They used the size of OD area for analysis of glaucoma disc. This study proposes a novel searching technique to identify the position of OD from retinal images based on edge detection of retinal images. The position of OD can be identified by Hough transform technique which detects several straight lines and approximated them as a circular line. The position of OD was identified as the center of the circle. II.

METHODOLOGY

The searching technique used in this study consists of four steps. Firstly, we converted color retinal images to gray scale images. Secondly, we generated edge images and modified edge detection to keep edges with high brightness. Thirdly, we searched several straight lines and approximated them as a circle line using Hough transform technique. Finally, we identified the position of OD as the centre of circle. Fig. 2 summaries the steps involved in our method. A.

Conversion to Gray Scale The color retinal images are composed of red channel, green channel, and blue channel. We created histogram [8] of all channels as shown in Fig. 3 (a), Fig. 3 (b), and Fig. 3 (c), respectively. In the histogram of red channel, the majority of the pixels had high luminosity values. In contrast, pixels in

Color retinal image

Gray scale image

Edge detection image

Searching circle line

Identification of OD Fig. 2. The flowchart of proposed method.

the blue channel were mostly in the low luminosity area (dark). In both red and blue channels, OD area had low resolution. We selected the green channel for conversion to gray scale image because luminosity values of the pixels were found to be in the median region B.

Application of Edge Detection

The color retinal image in Fig. 1 (b) was converted to gray scale image as described in Section II-A. The algorithms for edge detection used in this study were Laplacian of Gaussian, canny edge detector, robots operator, and Sobel operator. Results of edge detection using various algorithms in previous studies are shown in Fig. 4. Results from Laplacian of Gaussian algorithm in Fig. 4 (a) appear to have higher edge details than those from other methods. However, it is not used for searching OD because several straight lines cannot be estimated as a circular line around OD area. Fig. 4 (b) illustrates average edge details, and it is used for searching OD because edges surrounding OD area can be estimated as a circular line. Finally, Fig. 4 (c) and Fig. 4 (d) are bare edge details for searching OD. The proposed method in this study uses modified canny edge detector [9-10] for searching OD. The procedure for OD search in this paper is as follows: Step 1 Remove Noise from the gray scale retinal image by Gaussian filter. We created the 13×13 kernel and the standard derivative used was 1-4 as shown in (1).  

 

(1)

is a 2 dimensional kernel; and are column where and row respectively; and is a standard derivative.

2008 International Symposium on Communications and Information Technologies (ISCIT 2008)

671

(b)

(a)

(c)

Fig. 3. The histogram and the images: (a) red channel, (b) green channel, (c) blue channel.

Step 2 Compute the gradient magnitude using (2) and the orientation of edge image using (3).  





(2)

 

(3)

is the gradient amplitude;  and where derivatives in direction and respectively; and the gradient angle.

 

are first is

Step 3 Compute the non-maxima suppression (NMS) using (4). If a point lies on the edge, it will have a high pixel value and has same direction. Other points that do not lie on the edge have zero pixel value. (4)

Step 5 Use double threshold algorithm for detecting edges of the image. Assign two thresholds⎯T1 = 0.21, a (high threshold value), and T2 = 0.20 (low threshold value) Pixel values over T1 are considered as edges of the image, whereas pixel values between T1 and T2 are considered pixels value close to the edges of the image. Other pixels are considered as zero valued pixels. Finally, the pixels less than T2 are zero valued pixels. In this method, the difference between thresholds T1 and T2 is 0.01 which would generate a minimum edge image with high brightness as shown in Fig. 4 (b). C.

Detection of Circle Line by Estimating Straight Line Straight lines found around OD area can be interconnected to form a circular line. We detected them by using a technique based on Hough transform [11]. The algorithm for detection of a circular line is shown below.

is the output image of NMS.

where

Step 4 Compute the adjust gamma of image using (5). The mainly parameter adapts gamma level into brightly and shadowy image. We assign the gamma value as = 2/3 for optimal brightness the monochrome retinal image is found to be the image with best edge on OD. Then mapping brightly pixels is performed between NMS image and gamma image for generation of the following binary image.



672

where  is the output image of adjust gamma; is the monochrome retinal image from the resultant image in section II-A;  and !"# are maximum and minimum pixels in monochrome retinal image;

 

Step 1 Initialization of the minimum radius and the maximum radius of the circle line is searched. In this study, we configured the minimum and maximum radials to be 15 and 100, respectively. Step 2 Compute position of straight line following as (6). While pixels of many lines are collected in Hough space or . coordinated

 

( ) where is the distance between the origin point to are row position intersection point on a straight line. and column position, respectively, while is the angle between and

2008 International Symposium on Communications and Information Technologies (ISCIT 2008)

Steep 3 Com mputing poosition of circle linee followinng as (7) consiists of cirrcle equattion and parametriic equatioon. Then assign ning radiu us value and confiigured thee cooordinate system m into * + cooordinated system. In n computeed coord dinated sy ystem whicch increassed angle starting aat to 360 . The piixel value on coordinatte will be pplus 1, if found d repeated dly ccoordinatee in 360 steeps. &

,

-

/

'

2 3

where is radiius; . c pointt of the circle; positiion respecctively.

1

(a)

(b)

(c)

(d)

are thhe coordin nates of thhe center are row position and column

Steep 4 We identified the cennter pointt of the circle is s the high bright pix xel value in $% () by searching coord dinate systtem. The eexample reesults as sh hown in F Fig.5. III.

EXPPERIMENTTAL RESULLTS

Th his study demonstra d ated detection of th he OD position in colorr retinal im mages. W We used thhe retinall images database from Digital Retinal R Im mages for Vessel V Ex xtraction ((DRIVE) [11]. The retin nal imagees are com mposed off 40 coloor retinal images in RGB B color ssystem, wiith dimen nsions of 5565×584 pixels, and reco orded in T Tagged Im mage Form mat (TIF). IIn the 40 retinaal images used, 33 iimages weere of heaalthy retinaas, and 7 images were from off diseased retinass. The searching c algorrithms weere implemented on a computer running Wind dows XP 64-bit opperating system, s In ntel Coree 2 Duo proceessor at 1.8 1 GHz. We deveeloped thee algorithm using MAT TLAB® R2 2007a verrsion 7.4. The execu ution perfformance of ou ur method is shown in Fig. 6. Processin ng of 34th image is very time-conssuming as binary im mage has many m straight lines t biggesst OD cirrcle, thus it is difficult to seearch the and the circlee line. On the other hand, proocessing of the 21st image is the qu uickest beecause few w straight lines l are in n the binarry image. Proceessing of the t 3rd imaage was unnsuccessfu ful due to vvery low brigh htness at the t OD poosition. Frrom 40 reetinal imaages, our method could identify the OD positionss correctlyy in 39 images. The evaluation of the peerformancee of the pproposed method can be computedd by (8).

F From (8), the perfoormance of the prop posed metthod was 97.5 percent as shownn in Tablle 1. OD D detectioon errors occurrred in rettinal imagges wheree the edgees of the iimage in OD area a had few detaills, due too low brig ghtness at the OD positiion. The example results from f heallthy and diseased retinaa images are a shown in Fig. 7.

Fig. 4. Results of 16 1 th image eddge detectionn using: (a) Laplacian L off Gaussian, (b) caanny edge deetector, (c) roobets operatoor, and (d) Sobel S operatoor. Veryy bright

(a) Very bright V

V Very bright

(b) Veery bright

(c)

(d)

Fig. 5.. Results of identified i higgh bright of centre c circle:: (a) the 16th image, (b) thee 3rd image, (c) the 34th im mage, and (dd) the 21st im mage. TABLE I PERFORMAN NCE A Algorithms

Performaance (%)

H Hoover [1]

89.00

Foracchia [2]

97.50

Ouur Method

97.50

2008 International Symposium on Communications and Information Technologies (ISCIT 2008)

673

IV. CONCLUSION Our proposed method has a high efficiency rate for identification of OD positions in color retinal images. The method is successful with control of modified canny edge detection technique. The results illustrated that the proposed method could be used to assist analysis of pathological retinal images. ACKNOWLEDGMENT The authors would thank the Image Sciences Institute for DRIVE database available on internet. Fig. 6. The time consumption of proposed method.

(a)

(b)

(c)

(d)

(e)

(f)

Fig. 7. The example results: correctly identified healthy retinas, (a) the 16th image, (b) the 1st image, (d) the 21st image, and (e) the 39th image; correctly identified diseased retina, (c) the 34th image; shows false identification of healthy retina, (f) the 3rd image.

674

REFERENCES [1] Adam Hoover and Michael Goldbaum, “Locating the Optic Nerve in a Retinal Image Using the Fuzzy Convergence of the Blood Vessels,” IEEE Transactions on Medical Imaging, vol. 22, no. 8, pp. 951-958, 2003. [2] M. Foracchia, E. Grisan, and A. Ruggeri, “Detection of Optic Disc in Retinal Images by Means of a Geometrical Model of Vessel Structure,” IEEE Transactions on Medical Imaging, vol. 23, no. 10, pp. 1189 – 1195, 2004. [3] Huiqi Li and O. Chutatape, “Automated Feature Extraction in Color Retinal Images by a Model Based Approach,” IEEE Transactions on Biomedical Engineering, vol. 51, no. 2, pp. 246 – 254, 2004. [4] Zheng Liu, O. Chutatape, S. M. Krishnan, “Automatic Image Analysis of Fundus Photograph,” Proc. IEEE Conf. on Engineering in Medicine and Biology society, vol. 2, pp. 524 – 525, 1997. [5] V. Thongnuch and B. Uyyanonvara, Automatic Detection of Optic Disc from Fundus Images of ROP Infant Using 2D Circular Hough Transform. Proc. ISBME Conf. on International Symposium on Biomedical Engineering, pp. 328 – 330, 2006. [6] C. Sinthanayothin, J. F. Boyce, H. L Cook, T. H Williamson, “Automated Localisation of the Optic Disc, Fovea, and Retinal Blood Vessels from Digital Colour Fundus Images,” British Journal of Ophthalmology, vol. 83, pp. 231 – 238, 1999. [7] E. Corona, S. Mitra, M. Wilson, T. Krile, Y. H. Kwon, P. Soliz, “Digital stereo image analyzer for generating automated 3-D measures of optic disc deformation in glaucoma,” IEEE Transactions on Medical Imaging, vol. 21, no. 10, pp. 1244 – 1253, 2002. [8] M. Jankowski, Mathematica Digital Image Processing. Wolfram Research, 2007. [9] R. C. Gonzalez, R. E. Woods, and S.L. Eddins, Digital Image Processing Using MATLAB. Pearson Prentice Hall, 2004. [10] J. Canny, “A Computational Approach to Edge Detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 8, pp. 679 – 698, 1986. [11] M. Nixon and A. Aguado, Feature Extraction and Image Processing. Newnes, 2002. [12] http://www.isi.uu.nl/Research/Databases/DRIVE

2008 International Symposium on Communications and Information Technologies (ISCIT 2008)