Detecting the Optic Disc in Retinal Images

by means of a Geometrical Model of Vessel Network

Alfredo Ruggeri, Marco Foracchia, Enrico Grisan

Abstract

We present here a new method to identify the position of the optic disc in retinal fundus images. The method is based on the preliminary detection of the
main retinal vessels by means of a vessel tracking procedure. All retinal vessels originate from the optic disc and then follow a parabolic course
towards retinal edges. Thus, a geometrical parametric model was proposed to describe the direction of these vessels and two of the model parameters
are just the coordinates of the optic disc center. Using samples of vessels directions (extracted from fundus images by the tracking procedure) as
experimental data, model parameters were identified by means of a simulated annealing optimization technique. These estimated values provide the
coordinates of the center of optic disc. A Matlab prototype implementing this method was applied to a set of 40 images of both normal and pathological
subjects. In all these images, the optic disc position was correctly identified, even in rather difficult pathological situations. An extensive validation on a
set of 81 images (STARE project data set) is currently in progress to assess the robustness of the proposed technique.
only describes the general course of main vessels. Moving away from the optic
disc, these vessels bifurcate, and branch vessels tend to diverge from main
vessel direction: vessels inside the parabolas bend towards the center of the
image, while those outside bend towards the external edges of the image (Fig.
2).

Introduction

The optic disc appears as a

round region usually brighter
than the surrounding (Fig. 1).

Locating its position is

crucial,
especially
in
automatic analysis:
important image landmark
affected by many retinal
pathologies
might easily be confounded
with exudative lesions
caliber of the main vein at
its edge used as a length
reference (125 m).

The vessel direction provided by a complete model should thus assign a direction
tangent to the parabolas to points belonging to locus , and an increasingly
diverging direction, as described above, to points outside .

Results and discussion

Matlab prototype realized and tested on 40 fundus images (605x700 pixels, 45
field of view) with vessels already tracked, from normal and pathological subjects
(STARE project [10]).
The method was able to correctly position the optic disc in all 40 images.
Four examples of pathological images shown in Fig 4. Despite their problems, the
technique was able to correctly detect the position of the optic disc in all of them.

Complete model for vessel direction at any point (x,y) of the image (Fig. 3):

2 ax

mod ( x, y ) = arctan sgn( x ) sgn( y )

Fig. 1. Retinal fundus images with vessels
and optic disc (yellowish circle on the righthand side).

Many techniques proposed to detect the optic disc, based e.g. on its relatively
high brightness (see e.g. [1-4]). They often fail on pathological images.
Other techniques exploit the information provided by the vessel network, i.e.,
the fact that all main retinal vessels originate from the optic disc [5].
Our method is also based on the detection of vessel network, but with
additional robustness from the a priori knowledge included in the model of
vessels direction. Using a geometrical description of the vascular structure, it
derives the coordinates of the origin of main retinal vessels, which is inside the
optic disc and very near its center (Fig. 1).

+ d ( x, y; )

where d is a term modelling

the
increasingly
diverging
direction when moving away
from the parabolas; (a,) are
model parameters, and vector
includes the optic disc center
coordinates (xOD,yOD).

Model identification
Minimization of the residuals sum
of squares RSS is performed
with respect to model parameters
(a,) (minus sign indicates a
direction difference):

Fig. 3. Complete model of vessels direction.

RSS = i ( xi , yi ) mod ( xi , yi ; )

Methods
Vessel tracking algorithms extract the vascular structures from fundus images (e.g.
[6-8]). They can provide, for each detected point in a vessel centerline, its position
(xi, yi) and vessel direction (xi, yi).

Model of retinal vessels direction

Optimized values of parameters xOD and yOD represent the best positioning of
the optic disc according to the model fit on the available samples (xi,yi) of the
vessels direction.
Minimization of RSS with classical gradient-based techniques is rather critical,
since RSS has many local minima. Fig. 4 represents, e.g. , a plot of RSS as a
function of parameters xOD and yOD only.
Simulated Annealing (SA)
optimization procedure [9]
was used, a probabilistic
technique with the capability
of moving out of local
minima.

The course of the main

vessels originating from the
optic disc can be modeled as
two parabolas, having a
common origin inside the optic
disc (Fig. 2).

= {( x, y ) : ay 2 = x }

Fig. 2. Parabolic model of main vessels course.

Acknowledgments
This work was supported in part by a research grant from Nidek Technologies, Italy.
The authors wish to thank Prof. Adam Hoover, Clemson University, SC (USA), for
having kindly provided images and tracking data from the STARE project data set.
MF is now with M2 Scientific Computing, Italy.

Bibliography
[1] C. Sinthanayothin et al., Br. J. Ophthalmol., Vol. 83, pp. 902-910, 1999.
[2] H. Li, O. Chutatape, Proc. IEEE-ICIP, vol. 2, pp. 837-840, 2001.
[3] M. Lalonde et al., IEEE Trans. Med. Imag., vol. 20, no. 11, pp. 1193-1200, 2001.
[4] A. Osareh et al., Proc 16th Internat. Conference on Pattern Recognition, vol. 1, pp. 743-746, 2002.
[5] A. Hoover, M. Goldbaum, IEEE Trans. Med. Imag., vol. 22, no. 8, pp. 951-958, 2003.
[6] Y. Tolias, S. Panas, IEEE Trans. Med. Imag., vol. 17, no. 2, pp. 263-273, 1998.
[7] A. Hoover et al. IEEE Trans. Med. Imag., vol. 19, No. 3, pp. 203-210, 2000.
[8] M. Foracchia et al., in EMBEC'02 - IFMBE Proc. Series, Wien, vol. 3, pp. 1558-59, 2002.
[9] N. Metropolis et al., Journal of Chemical Physics, vol. 21, pp. 1087-1092, 1953.
[10] http://www.ces.clemson.edu/~ahoover/stare.

SA was run six times for

each image, starting from
different initial points in the
parameter space, and the
lowest value reached was
assumed as the absolute
minimum.

This can be described by the

geometrical locus

Fig. 4. Some pathological images with estimated parabolas (blue line) and optic disc
position (white cross).

Fig. 4. Plot of RSS values as a function of

model parameters (xOD, yOD).

Department of Information Engineering - University of Padova

Via Gradenigo 6/a - 35131 Padova, ITALY
alfredo.ruggeri@unipd.it - enrigri@dei.unipd.it