![](bg2.png)
is processed to enhance the vascular structure and ex-
tract linear segments. The processed results from im-
ages taken before and after the bypass surgeryare then
compared (via a logical AND operation) to identify
the differences. However, to our knowledge, there are
no reports in the literature of any technique to detect
the cause of PDR namely the presence of the CNP re-
gions anywhere in the retina. Detecting and segment-
ing CNPs is the focus of this paper.
The clinical procedure to detect CNPs is a visual
scan of an FFA image. In order to estimate the amount
of area damaged, the scan is generally done on the
composite image of the retina obtained after suitable
mosaicing of several retinal segments. Such a pro-
cedure suffers from several drawbacks: the variable
skills and subjectivity of the observer, which also de-
pend on the quality of the images; a lack of precise un-
derstanding of the area of retina affected which helps
in deciding the nature and extent of laser treatment.
Automated image analysis techniques can be used to
address these issues but there are several challenges in
devising solutions for CNP segmentation. FFAs suf-
fer from non-uniform illumination due to the eye ge-
ometry, imaging conditions and presence of other me-
dia opacity such as cataract. Inter-patient and intra-
patient variablility is also possible. The former is due
to different pupil dilations and the latter is due to the
time of image capture after injection of fluorescein
dye. Another compounding factor is that the mean
grey level of CNPs as well as their shape and size are
variable, with the size ranging from very small to very
large (from 100 to 55000 pixels). Often, the bound-
aries of CNPs are not well defined because of an in-
homogeneous textured background. Thus, the only
visually distinguishing characteristic of a CNP is that
it is relatively darker than its surround.
In this paper, we propose a novel method to ex-
tract and quantify regions of CNP based on modeling
CNPs as valleys in the image surface. The algorithm
for CNP segmentation is developed and its details are
presented in the next section. Section 3 provides im-
plementation details and illustrative test results of the
algorithm. Finally, some discussions and conclusions
are presented in the last section.
2 VALLEY BASED CNP
SEGMENTATION
2.1 Modelling CNP Regions
As discussed earlier, CNP occurs when the capillary
network in a region of the human retina stops func-
tioning and does not supply blood to the correspond-
ing areas. In FFAs, regions receiving normal blood
supply appear as bright white regions since they carry
a fluorescent dye and regions lacking in blood (due
to abnormal supply of blood) appear as dark regions.
Hence, regions of CNP appear as dull/dark lesions
bounded by healthy vasculature.
A sample FFA image and an enlarged view of a
CNP region and its surroundings is shown in Fig. 1.
Also, included in this figure is the surface plot of the
corresponding CNP region from which we can ob-
serve that the prominent vessels, the healthy capillary
network and the CNP have very different topographic
charactersitics: While the major vessel appears as a
ridge, the CNP appears as a valley with the healthy
capillary network appearing as a plateau in the image.
Hence, one can conclude that CNPs can be modelled
as valleys. Watershed-based solution to valley detec-
tion (for example, (Gauch, 1999)) is possible, how-
ever, these result in oversegmentation or in the case
of marker-based versions, require additional informa-
tion. In the case of CNP detection, since the size of a
CNP and the nature of its surround can be highly vari-
able, obtaining such markers can be quite challenging.
A better alternative is to identify the trough (lowest
point on a curve) and use it to segment a CNP. Hence,
we have taken a different approach to the problem
and propose a technique that detects trough points and
collates them across scales. We next present the de-
tails of our proposed algorithm for CNP segmentation
comprising several steps.
2.2 CNP Detection Algorithm
The proposed CNP detection algorithm consists of
these stages: Firstly, illumination correction (IC) is
done to minimise the background intensity variation
followed by denoising to eliminate noise that is fre-
quently found in FFAs. Next, valley detection is per-
formed to locate the seed points in the CNP regions
which are used to extract the candidate CNP regions
using a region growing algorithm. Finally, threshold-
ing is done to reject false positives among the detected
candidates. The processing in each of these stages are
described next.
2.2.1 Illumination Correction
Nonuniform illumination is a problem in retinal
colour images as well as angiograms. A camera-
model based solution for illumination correction in
angiograms, obtained with non-confocal imaging, is
given in (Cree et al., 1999) which assumes a macula-
centric view of the retina. Our images are not neces-
sarily macula-centric and are obtained from a laser-
AUTOMATIC SEGMENTATION OF CAPILLARY NON-PERFUSION IN RETINAL ANGIOGRAMS
171