Extraction F the Optic Djsk Boundary in Digital Fundus Images

The.methods of active contours (“snakes”) and level sets were applied to images of the retina in order to locate the outer boundary of the optic disk. A gradient-vector-flow based acuve contour was used as it performed well over a large range of initial conditions. Images were pre-processed to lessen the influence of blood vessels on boundary detecuon. Both active wntours and level set methods accurately located the wxrect boundary; level set methods were computationally more intensive. I. INTRODUcrION The optic disk is a significant anatomical landmark in the retina. Vanous ophthalmic pathologies, especially glaucoma, are manifest by changes in the shape, pallor, or depth of the optic disk region. Accurate identification of the outer boundary of the optic disk may allow ophthalmologists to quantitatively assess changes in the optic disk over time. This paper considers two techniques for locating the outer boundary of the optic disk using color digital images of the retina, a problem which has not previously been addressed. II. THEORY Active contours (“snakes”) have been widely used in the detection of closed contours. lhese are. energy minimizing contours guided by external (image-derived) and internal (contour-derived) forces. Initial formulations of active contours suffered from a need for good initialization, and an inability to move into small concavities. A gradient vector flow (GVF) based snake was introduced to address these l i tat ions [2]. In this formulation, a more general external force is defined which gives a directional field that accounts for boundary proximity, but with a larger range of attraction. This decreases the sensitivity to initial conditions. An alternative approach to boundary detection is offered by level-set theory [3]. In this, a desired propagating boundary is considered as the zero level set of a higher dimensional function Y(x,t). Ihe goal is to produce an equation for the evolving function Y(x,t) which contains the embedded motion of the houndaq as the evolving zero level-set of Y(x,t). An evolution equation for Y(x,z) can be wrimn, with an explicit ‘‘speed” function which controls the time rate of evolution. If the speed on the propagating interface drops close to zero, a convergent boundary is found. A speed function inversely proportional to the image gradient encourages the level set to propagate to strong-edged boundaries in the image. The narrow-band extension is used to increase computational speed [3]. Figure 1: (a) Optic disk boundary located by active contour method. (h) Same boundary using level sets. m. RESULTS Both methods were applied to sample images ( ~ 9 , image size = 285 x 400 pixels). Conversion from RGB to luminance was carried ant prior to boundary detection. Initial experiments showed that the strong edges due to blood vessels crossing the optic disk prevented correct boundary identification. Preprocessing approaches were developed to minimize the confounding effect of these vessels, with the aim of removing the vessel structure from the optic disk. In one technique, candidate vessel pixels were identified as local luminance minima, and replaced with “background” optic disk pixels, if certain conditions were satisfied. In the s a n d approach, morphological operations (dilation, erosion, and maximization) were carried out using a 5~5structuring element. Both techniques produced an image on which vessel structures had been largely removed. ’Ihe optic disk boundary was then located on the processed image using both the GWbased active contour and level-set algorithms. Sample detected boundaries are shown in Figure 1. VI. DISCUSSION AND CONCLUSION Both techniques detected boundaries which were considered accurate by two clinical ophthalmologists. A disadvantage of the level set method over active contours was its high computational cost. Methods are currently b e i g investigated for improving performance, including the generation of an hierarchical adaptive mesh structure. RFZERENCES 111 M. Kass, A. Witkin, D. Tuzopoulos “Snakes: Active Contour Models”, fnt. 1. of Cornpurer Viswn, vol. I, pp 321-331, 1988. [Z] C. Xu. 1. L. Phce, “Snakes. shapes, and gradient vector flow,” IEEE Trans. on I m g e Processing, vol. 7, pp.3559-369.1998. [31 I.A. Serbian, ‘level Set Methods”, Cambridge Monogmphr on Applied and CompvarwnaI Matkemtics, Cambridge, 1996. 0-7803-5674-8/99/$10.00