Information Technology Reference
In-Depth Information
Genetic algorithms (GA) have also been used eciently as a search method to
identify the papillary contour [1,7]. Other recent approaches have used different
types of techniques like, for example, mathematical morphology [8], image filters
[9] or the Hough Transform [10].
Basically, our method uses a Laplacian pyramid of the eye fundus image to
obtain a set of interest points (IPs) in each pyramid level. Then a two-phase
genetic algorithm is used to achieve a progressive solution to the ONH contour.
The method here presented is a variation of the method described in [1]. The
main difference between both methods lies in how to get the set of IPs. On the one
hand, in the proposed method, the IPs are obtained using a Laplacian pyramid.
On the other hand, in [1], they are obtained using a domain-knowledge-based
method.
The article is organized as follows: Section 2 describes the methodology of the
proposed method. In section 3, the obtained results are evaluated and compared
with other methods. Finally, section 4 presents the conclusions and future work.
2 Method Description
The aim of this work is to locate and segment the ONH in eye fundus color
photographs through the process shows in the block diagram of the figure 1.
First, in order to reduce the processing computational cost, the process begins by
automatically extracting a subwindow from the original image. This subwindow
is approximately centered at a point of the papillary area. Then a Laplacian
pyramid is applied to this image subwindow and the result is an image multi-
scale representation where, in each pyramid level, a set of IPs is obtained. These
IPs correspond to image points of high-medium frequency, i.e., border points.
We use a two-phase GA: GA2+GA1. In both phases, the goal is to search for an
ellipse containing the maximum number of IPs in an offset of its perimeter. GA-2
is applied to IPs from level-2 and GA-1 is applied to ones belonging to level 1.
GA-2 is run first and, when it completes its execution, the best solutions obtained
are used as part of the GA-1 initial population. Finally, after running GA-1, we
select the best papillary contour from the final population as the solution of our
problem.
2.1 Extracting Subwindow by Gaussian Pyramid
The figure 2 shows the operator block diagram used for extracting an image
subwindow from the eye fundus original image, in order to reduce the noise (dis-
tracting structures) and the processing computational cost. This operator starts
applying a Gaussian pyramid of N-levels to the original image (600x400 pixels).
The idea is to smooth the image intensity levels in order to eliminate the image
high frequencies. In each pyramid level, the intensity of pixels belonging to the
retina, blood vessels and papilla are smoothed in different intensity intervals.
The smoothing process is more pronounced as the pyramid level increases. At
level-N, the region of the brightest pixels (RBP) will correspond to pixels be-
longing to the papilla. Here we use the property that the papilla pixels have an
Search WWH ::
Custom Search