The Detection of the Optic Disk Motivation

The optic disk (or papilla) is one of the main features of the retina, its detection is essential for a system of automatic analysis of retinal images; it is the prerequisite for other segmentation algorithms (exudates, macula).

In the context of diagnosis of the glaucoma, the detection of and measures on the optic disk may also be of great importance. Hence, an algorithm of automatic detection of the optic disk is required. Properties

The optic disk is the entrance of the optic nerve and the vessels into the retina. It is situated on the nasal side of the macula and it does not contain any photoreceptor: It is also called the blind spot. In color fundus photographs, the optic disk appears as a big bright spot of circular or elliptical shape, interrupted by the outgoing vessels. Its size varies from patient to patient, but its diameter is always comprised between 40 and 60 pixels in 640 x 480 images. The optic disk is characterized by a strong contrast between outgoing vessels and the bright color of the optic disk itself.

Unfortunately, this description is not valuable for all images: Sometimes, the contours are not clearly visible, the color tends more to a pale white, and there may be other regions in the image which are as bright or even brighter than the optic disk (due to nonuniform illumination or the presence of exudates). State of the Art

In [13], the authors localize the optic disk using the high contrast between the papilla and the outgoing vessels. This method fails if there are exudates in the image.

In [14], the authors use an area threshold for localization of the papilla, the Hough transform for the detection of its contours. The Hough transform is also used by [15]. The main problems that have been stated are low contrast and the case where its shape does not correspond to a circle (for example, if the optic disk is situated on the border of the image).

In [16], the authors use a template matching approach for the localization of the optic disk. The problem with this approach is the size variability of the papilla between different images and the presence of large accumulations of exudates. The Algorithm

The presented algorithm can be subdivided into two parts: the localization and the detection of the contours of the optic disk. First versions of this algorithm have been presented in [17,18].

Localization: As the optic disk belongs to the brightest parts of the image, the idea to apply an area threshold in order to find at least a part of the optic disk may work well, if there do not exist large accumulations of exudates or other bright regions. The atrophy in Fig. 7.21(a), for example, corresponds to a yellow spot and its size and shape are comparable to the one of the optic disk. Before we can apply a threshold to the image, it is therefore necessary to remove these bright features. This can be done using the vascular tree we have already detected: As the optic disk is the entrance of the vessels into the retina,

(a) The luminance channel of a retinal image containing an atrophy

(b) The morphological reconstruction using the vascular tree

Figure 7.21: The elimination of bright features.

(a) The luminance channel of a retinal image containing an atrophy

(b) The morphological reconstruction using the vascular tree

Figure 7.21: The elimination of bright features.

it must be connected to a dilated version of the main branches of the vascular tree:

It is recommended not to use the complete vascular tree V, but only the main branches that can be extracted easily by applying a stronger contrast criterion in the algorithm presented in section 7.5.1.

The effect of this filtering is shown in Fig. 7.21: The atrophy present in the image (a) is removed in (b), the optic disk stays nearly entirely unchanged by the reconstruction. Using the methods presented in [14,17,18], the localization algorithm would have failed in this case.

Now, we can assume that the optic disk belongs to the brightest elements of the image, and the application of an area threshold should give a part of the optic disk:

L1 = T[a>imax](l1) with a such that #L 1 > K (7.22)

L1 normally contains more than one connected component: A part of the optic disk, some noise, and eventually other bright features connected to the vascular tree. The latter ones are normally exudates of small size. Hence, it is sufficient to choose the connected component with the largest surface to obtain a part of the optic disk:

The center of the (only) connected component of L can be seen as the approximative center c of the optic disk and is used for the detection of the contours described in the following paragraph.

Detection of the contours: The contours of the optic disk appear under the best contrast in the red channel fr of the color image. Unfortunately, the red channel is sometimes saturated and cannot be used. In this case, we propose to work on the luminance channel fl. The first step is to determine if the red channel is saturated or not. Let c be the approximative center determined in the localization step of the algorithm, f a subimage of the red channel centered in c, and tmax(fr) the maximal gray-level value within this subimage. We define the

(a) The luminance channel

(b) The biggest particle of the threshed image

(c) The distance function of the particle

"Si iJ

Was this article helpful?

0 0

Post a comment