Proceedings of the International Conference on Computer and Communication Engineering 2008 May 13-15, 2008 Kuala Lumpur, Malaysia

978-1-4244-1692-9/08/$25.00 ©2008 IEEE

Optic Nerve Head Segmentation Using Genetic Active Contours

Abdul Razak Hussain Faculty of Information and Communication Technology

Universiti Teknikal Malaysia Melaka razak@utem.edu.my

Abstract As one of the leading causes of untreatable blindness worldwide, glaucoma is more likely to occur in persons of African or Asian descends as compared to Caucasians. Glaucoma is a manifestation of a group of conditions originated from different causes resulting in an increased pressure inside the eye. The pressure causes the blood vessels and retinal nerves to atrophy and leads to an eventual loss of vision. Early detection of glaucoma is important because it can minimise damage and allow for prompt and adequate treatment in avoiding blindness. The segmentation and evaluation of the optic nerve head (ONH) plays a significant role in the diagnosis of glaucoma. This paper reviews segmentation approaches using active contours (snakes) for locating and detecting the ONH in terms of the representation of the contours and the energy formulation used. It proposes an algorithm that combines active contour models and genetic algorithm (GA); coupled with the modified energy terms and the incorporation of the energy minimization procedure using GA, the proposed approach can greatly affect the performance of active contour models in medical image segmentation. Preliminary results indicates that an improved algorithm may improve the manual glaucoma screening efficiency.

I. INTRODUCTION Glaucoma is the third leading cause of blindness

worldwide after cataract and trachoma [1]. While both cataract and trachoma are treatable causes of loss of vision, glaucoma is untreatable. It is estimated that between 67 million and 107 million people will have glaucoma. Generally, persons of African or Asian descends are more likely to develop glaucoma and lose their sights compared to Caucasians. Glaucoma is one of the various eye-related conditions that require treatment. Rather than being considered as

a single disease, glaucoma should be treated as a group of conditions originated from different causes. Together, these conditions resulted in an increased pressure inside the eye that eventually causes the destruction of blood vessels and retinal nerves leading to loss of vision. Glaucoma is associated with a characteristic form of visual dysfunction and optic disc appearance. This progressive and irreversible damage to the optic nerve often results with subtle signs or even without symptoms; thus it is nicknamed the “Sneak Thief of Sight”. There are five types of glaucoma: primary open-angle glaucoma (POAG), normal-tension glaucoma (low-tension glaucoma), pigment dispersion syndrome (PDS)/pigmentary glaucoma, exfoliation syndrome (XFS) and angle-closure glaucoma [2,3,4,5]. Often, if defects were to be detected, they would be too late as significant damage to the nerve fibres may have occurred, causing a certain level of visual field loss. Therefore, as there is no cure for glaucoma, early detection of glaucoma is important because it can minimise damage and vision loss and allow for prompt and adequate treatments. One of the important tests for diagnosing glaucoma is the determination and evaluation of the optic nerve head (ONH) from retinal images. The ONH is a circular area where the optic nerve fibres converge. As glaucoma progresses, it causes the nerve fibres to atrophy and results in apparent changes in the shape of the ONH. Often, variability in the appearance of the ONH caused by image contrast and obscurity by blood vessels leads to subjective manual screening and analysis. In this paper, we outline a proposed segmentation algorithm to identify the ONH using active contour models (snakes) incorporating energy minimization procedure based on genetic algorithms (GA). This paper is organized in the following manner: Section 2 gives an overview of snakes. Section 3 reviews the approaches used in the segmentation of the

783

optic nerve head in retinal images. Section 4 provides motivations for using a segmentation algorithm using GA-based snakes. Finally, the paper concludes by outlining a proposed GA-based snakes algorithm.

II. ACTIVE CONTOURS Active contour models are splines that dynamically minimize their energy functions and in doing so, they “slither” like a snakes. Since proposed by [11], snakes have been used extensively, not only in general image processing but also in medical image processing applications to locate and outline object boundaries. The following examples indicate the use of snakes to some extent, as contributing to the assessment or diagnosis of diseases related to the cells [13], mouth [8], breast [24], heart [23] and leg [10]. The energy functionals of a common parametric snake model can be represented as:

Esnake = ∫0

1

[Einternal(v) + Eexternal(v) ] ds (1)

where v = v(s) = (x(s), y(s)), s ∈[0, 1] is the snake

contour expressed parametrically. Einternal represents the internal energy assumed

by the snake contour due to stretching and bending.

Eexternal represents the external potential energy influencing the snake configuration, derived from the image.

The user will initialize a snake by setting up the initial set of coordinates v(s). Often, this initialisation is represented as coordinates of a circle or an ellipse (for the sake of convenience) and placed close to the region of interest. The minimization of Equation (1) will result in the final contour of the snake.

A. Internal energy, Einternal

Einternal is associated with the smoothness of the contour. In [11] the internal energy is defined as: Einternal[v(s)] = Eelastic[v(s)]+ Ebending[v(s)]

Einternal[v(s)] = ( ) ( )⎥⎥⎦

⎤

⎢⎢⎣

⎡

∂∂

+∂

∂2

2

22 )()(21

ssvs

ssvs βα (2)

where

( )ssv

∂∂ is the first order derivative of the

snake contour, describing the elasticity of the contour.

2

2 )(s

sv∂

∂ is the second order derivative of the

snake contour, describing the curvature of the contour.

α(s) is a weighting parameter indicating the significance of elasticity (tension) of the contour.

β(s) is a weighting parameter indicating the significance of curvature (rigidity) of the contour.

While a higher value α(s) means higher tension as a result of the contraction of the snake contour, increasing the value of β(s) results in a less flexible snake, resisting to form corners. On the other hand, letting α(s) and β(s) to be zero means allowing discontinuities in the contour, thus resulting in a corner. We normally seek smooth contours in natural object; thus requiring the use of low values of the first and second derivatives that lead to low contour elasticity and curvature. External energy, Eexternal A commonly used edge functional is Einternal[v(s)] ( )2, yxIG ∗∇−= σγ (3)

where

γ is a weighting parameter ∇ is the gradient operator Gσ is the Gaussian convolution filter with a standard deviation σ I(x,y) is the image intensity

III. SEGMENTATION OF THE OPTIC NERVE HEAD

Although high intraocular pressure (IOP) poses a risk for glaucoma, the evaluation of the intraocular pressure alone is not diagnostic. It is important to note that while high IOP may not necessarily indicate glaucoma, low IOP also does not rule out glaucoma. Prior to the analysis, one needs to locate and segment the optic nerve head accurately from the background image. In general, the detection of the optic nerve head consists of two stages: pre-processing (localisation) to get the

784

initial region of the optic disk and segmentation to detect the boundary of the optic nerve head. The following sub-sections overview selected relevant existing work on the localisation and segmentation of the optic nerve head. Subsequently, Section 4 looks into the optimization of snakes using GA. A. Non-snakes approaches

[16] first located the candidate regions containing 180×180 pixels having the highest 2% grey level. The gradient was then calculated and the edge points were detected using the Sobel edge detector. The exact boundary of the ONH was not done; it was estimated as a circle using the least squares regression. [14] followed similar methods used by [16] except the candidate region was defined as a 200×200 pixels area.

[12] proposed a two-stage method in detecting the optic nerve head. The first stage tracked the optic disk through a multi-scale analysis (pyramidal decomposition) using a Haar-based wavelet transform. Then it searched for the contour based on the Hausdorff distance incorporating Canny edge detection. Similar method was used by [15] in detecting anatomical structures in optical fundus images.

Fundus images in [15] were pre-processed by selecting potential clusters whose pixels are the highest 1% grey level in intensity image. These candidate clusters were then disregarded if the number of pixels in that cluster is less than 100. Then, the Principal Component Analyses (PCA) was used to locate the center of the optic disk. After having calculated the eigenvectors from the training image, the distance between the new retinal image and the one specified by the eigenvectors was obtained. The minimum distance will be the point of the optic nerve head center. This method only detects a point indicating the center of the optic nerve head. This method fails to detect the centre of the ONH if the retinal image contains large number of exudates having similar (or greater) area or intensity with the ONH.

[9] experimented with multiple binary (blood) vessel segmentation to highlight vessels of varying scale. An automated method called fuzzy convergence was used to determine the origination of the blood vessel network. The method achieved 89% correct detection based on 81 images. Although this method did not specifically detect the ONH, it may be useful in locating the ONH as the central retinal artery and central retinal vein emanate from this region.

In general, these non-snakes approaches

successfully locate the ONH by identifying the centre of the disk or approximating it as a circle without having the boundary detected. These methods may be useful as an initialization stage for the snakes’ approaches. B. Snakes approaches

[19] noted that “whilst locating discontinuities boundaries using snakes in not novel, determining what pre-processing is required to make a robust algorithm is”. The images were pre-processed in three stages: image equalisation to enhance the difference between the bright nerve head and the darker surrounding retinal region, image thresholding to remove pixels that are definitely not part of the nerve head and image enhancing (edge detailing) using a pyramid edge detector. The snake was then initialised as a spline curve with the control points manually edited. The authors demonstrated that the algorithm failed to detect the boundary of the ONH in poorly exposed images noting the sensitivity of the snake algorithm to the pre-processed image.

[18] tested two different pre-processing methods: minima detection and mathematical morphological filtering, followed by a novel Gradient Vector Flow (GVF)-based snake. The grey-level mathematical morphology pre-processing allowed the snake to detect the boundary of the ONH better and least sensitive to previous initialisations. In the nine images used for the experiments, an accurate nerve head boundary was successfully detected. Nevertheless, the snake’s control points still need to be placed manually around the optic disk.

[22] improved the work by [18] through: 1). Automatically estimated the centre of the optic disc using template matching, and 2). Using a colour mathematical morphology, instead of grey-level mathematical morphology. The ONH boundary detection was done using a Gradient Vector Flow (GVF)-based snake similar to that in [18]. A success rate of 90.32 % was reported for detecting the ONH boundary in 75 images.

[21] tested two different pre-processing methods in order to provide an approximation of the ONH location: template matching and least squares regression arc estimation. Either of these methods was adequate to isolate the ONH region from the background retinal image. A Gradient Vector Flow

785

(GVF)-based snake was implemented after the colour mathematical morphology was done. A success rate of 90.7% was reported based on 16 images.

IV. GENETIC SNAKES A. Motivation

GA has been acknowledged and applied effectively in image processing and image analysis applications. Snakes also have the reputation of being widely used in computer graphics and computer vision applications. Until recently, there is no reported work on the segmentation of retinal images pertaining to glaucoma using genetic snakes. These motivate us to consider GA-based snakes for the segmentation of retinal images pertaining to glaucoma.

Various segmentation approaches involving genetic snakes have been reported. [17] used genetic snakes (with six control points) to segment image contours on a variety of images. They showed that snakes with a larger neighbourhood (5×5 search window) have a better convergence rate compared to a 3×3 search window. [20] followed similar approach as in [17] but by dividing the contour into sub contours, each having four control points. Such set-up allows large contour movement thus making it possible to find the final contour in less iteration. [6] experimented with genetic snakes in the segmentation of the Foveal Avascular Zone (FAZ) in relation to diabetic retinopathy.

In short, GA-based snakes may be advantageous over traditional methods by: 1). operating on the coding of the parameters rather than on the parameters, they are able to handle arbitrary constraints thus having a low order of complexity, 2). exploring in a discrete space, a population of points rather than a single point thus enabling parallel computation, and 3). avoiding higher order derivatives of the fitness function through usage of probabilistic rather than deterministic rules.

B. Segmentation algorithm Our proposed method consists of two stages: 1. Image pre-processing. The initial colour retinal

image is to be pre-processed to eliminate unnecessary background and to suppress the blood vessels using the colour mathematical morphology as suggested by [22].

2. ONH Boundary detection. This boundary detection technique incorporating GA-snakes follows similar approach as in [7]. We differ in

terms of the initial population and the neighborhood setting up. The initial population consists of four different groups and a radial window is selected instead of a square window.

We use polar coordinates to simplify the

implementation. A region of interest (ROI) is defined as the area bounded between Rmin and Rmax (to be defined by the users) as shown in Figure 1. The ROI is further segmented into four quadrants as designated by ophthalmologists – (I)nferior, (N)asal, (S)uperior and (T)emporal quadrants.

Figure 1 : Snakes space parameterization

The initial population of snake points is chosen randomly within this ROI in terms of radius r(s) and an equally spaced angular displacement θ(s). There are four groups of initial populations – one for each of the quadrants. The parameters to be optimized is the positions of the snake points in the image plane vN(s) = ( θN(s), rN(s) ). (refer Figure 2. Subscript indicates quadrant). Since the angles in the quadrant are multiples of one another and change from 0 to π/2, we may evolve the populations in the four quadrants in parallel. The fitness function is defined as the snake energy defined by Equation (1). Each point vN is moved to the point rN, corresponding to the minimum energy value, and the entire contour v(s) should approach and stop at the object boundary.

V. CONCLUSION AND FUTURE WORK We have presented an overview of glaucoma and

the significance of the ONH analysis in aiding the diagnosis. We reviewed previous work done in locating the ONH and detecting its boundary. Finally, we

Inferior

Superior

NasalTemporal Rmax

Rmin

r θ

region of interest

feature to be detected

786

Figure 2: As θ changes, the search space lies

Rmin Rmax

vNi-1

vN

vNi+1

rN

contour

object boundary

outlined an on-going work on a proposed segmentation approach incorporating GA-based snakes to facilitate the diagnosis of glaucoma. Among the future work planned: on-going image processing application development using Java, testing and benchmarking of the proposed algorithm using synthetic and medical images and verifying results with the medical experts.

REFERENCES [1] World Health Organisation (WHO) 1997. Fact 143.

(on-line) http://www.who.int/inf-fs/en/fact143.html (22 May 2003)

[2] The Glaucoma Association of New York (on-line) http://www.glaucoma.net/gany/faq/contents2.html (21 May 2003)

[3] The Glaucoma Foundation (on-line) http://www.glaucomafoundation.org (22 May 2003).

[4] The Glaucoma Research Foundation (on-line) http://www.glaucoma.org (22 May 2003).

[5] International Glaucoma Association (on-line) http://www.iga.org.uk (22 Feb 2004).

[6] Ballerini, L. 1999. Medical image segmentation using genetic snakes. In Bosacchi, Fogel, Bedzek (eds) Proc. SPIE: Application and Science of Neural Networks, Fuzzy Systems and Evolutionary Computation II 3812: 13-23.

[7] Ballerini, L. 2001. Genetic snakes for color images segmentation. In E.J.W. Boers et al. (eds.) EvoWorkshop 2001, Lecture Notes in Computer Science 2037:268-277. Berlin: Springer-Verlag.

[8] Hamarneh, G., Chodorowski, A. and Gustavsson, T. 2000. Active contour models: application to oral lesion

detection in color images. IEEE International Conference on Systems, Man and Cybernetics 4: 2458-2463.

[9] Hoover, A. and Goldbaum, M. 2002. Locating the optic nerve head in a retinal image using the fuzzy convergence of the blood vessels. IEEE Transaction on Medical Imaging (accepted Oct.2002, to appear).

[10] Jones, T.D. and Plassman, P. 2000. An active contour model for measuring the area of leg ulcers. IEEE Transactions on Medical Imaging 19(12): 1202-1210.

[11] Kass, M., Witkin, A. and Terzopoulos, D. 1988. Snakes: active contour models. International Journal of Computer Vision 1(4): 321-331.

[12] Lalonde, M., Beaulieu, M. and Gagnon, L. 2000. Fast and robust optic disk detection using pyramidal decomposition and Hausdorff-based template matching. Report CRIM-00/12-10, ISBN 2-89522-005-0. pp.1-19.

[13] Leymarie, F. and Levine, M.D. 1993. Tracking deformable objects in the plane using an active contour model. IEEE Transactions on Pattern Analysis and Machine Intelligence 15(6): 617-634.

[14] Li, H. and Chutatape, O. 2000. Fundus image features extraction. Proc. of the 22nd Annual EMBS Int’l Conf. pp. 3071-3073.

[15] Li, H. and Chutatape, O. 2001. Automatic location of optic disk in retinal images. Proc. of IEEE Int’l Conf. Image Processing 2001. pp.837-840.

[16] Liu, Z., Opas, C. and Krishnan, S.M. 1997. Automatic image analysis of fundus photograph. Proc. of the IEEE/EMBS 19th Int’l Conf. 2: 524-525.

[17] MacEachern, L.A., Manku, T. 1998. Genetic algorithms for active contour optimization. IEEE Int. Sym. For Circuits and Systems pp.229-832.

[18] Mendels, F., Heneghan, C. and Thiran, J-P. 1999. Identification of the optic disk boundary in retinal images using active contours. Proc. of Irish Machine Vision and Image Processing Conf., pp.103-115.

[19] Morris, D.T. and Donnison, C. 1999. Identifying the neuroretinal rim boundary using dynamic contours. Image and Vision Computing 17: 169-174.

[20] Ooi, C. and Liatsis, P. 2001. Co-evolutionary-based active contour models in tracking of moving obstacles. Proc. of Intl. Conf. on Advanced Driver Assistance Systems. pp. 58-62.

[21] Osareh, A., Mirmehdi, M. Thomas, B. and Markham, R. 2002a. Classification and localisation of diabetic-related eye disease. In A. Heyden et al. (eds.) ECCV 2002, Lecture Notes in Computer Science 2353: 502-516. Berlin: Springer-Verlag.

[22] Osareh, A., Mirmehdi, M., Thomas, B. and Markham, R. 2002b. Comparison of colour spaces for optic disc localisation in retinal images. In R. Kashturi et al. (eds.) Proc. of the 16th Int’l Conf. On Pattern Recognition, IEEE Computer Society pp. 743-746.

[23] Ranganath, S. 1995. Contour extraction from cardiac MRI studies using snakes. IEEE Transactions in Medical Imaging 14(2): 328-338.

[24] Wirth, M.A. and Stapinski, A. 2003. Segmentation of the breast region in mammograms using active contours. In T. Ebrahimi, T. Sikora (eds.) Visual Communications and Image Processing 5150: 1995-2006. Lugano, Switzerland.

787