Page 91 - Computational Retinal Image Analysis

5 Optic disc and fovea detection 83

5 Optic disc and fovea detection

There have been many algorithms developed over the last 30 years for automatically
detecting the optic disc and fovea in retinal fundus photography. The reasons for do-
ing so have been stated above and this is generally a first image processing step in a
pipeline of algorithms, whether they be for automatic or semiautomatic systems. In
semiautomatic systems, where a human can intervene after certain processing steps
and make changes, the consequences of incorrectly identifying the OD and fovea are
minor. It would simply take a little extra time for the user to make the correction and
continue. In automatic systems, the wrongly detected OD or fovea can affect the ac-
curacy of future algorithms, possibly impacting disease diagnosis.
Algorithms for the detection of the optic disc vary widely, but the approaches
can generally be narrowed into subsets based on the type of information they use to
make their OD prediction. These subsets include (1) detecting the OD as the brightest
region in the image, (2) Detecting the OD as the convergence of the retinal vascu-
lature, (3) Detecting the OD through template matching and (4) Detecting the OD
through supervised methods. The next section will give examples of methods de-
veloped using information from these subsets and show the ways researchers have
combined this information to make robust algorithms that can detect the OD in the
presence of disease, low image quality or even when the OD is not present in the
image. The methods highlighted in this chapter were chosen to cover a large period
of time (20 years) and show how popular early methods were built upon to address
inherent shortcomings. To find in-depth literature reviews on this topic, please see
[18, 19]. There are fewer methods available for fovea detection. A priori knowledge,
such as eye side and field of view, can constrain the problem and make detection
much easier. When available in the following OD methods, fovea detection is also
highlighted.

5.1 Automated localization of the optic disc, fovea, and retinal
blood vessels from digital color fundus images (Sinthanayothin
et al., 1999 [20])
This method was chosen as the starting point for this review because it is one of
the earliest papers in this area and is subsequently very highly cited. The authors
developed methods to detect all the retinal landmarks and shows how fairly simple
methods were able to detect both the OD and fovea with high accuracy. First, the
images were preprocessed to enhance local contrast. The images were converted
from the RGB color space to the intensity hue saturation space. This separates the
intensity and color information so that the contrast can be adjusted without affecting
perceived color. A variance image is then created by calculating the mean variance
within 81 × 81 pixels ROI’s within the main image, with the maximum value being
the found OD. The reasoning being that the bright OD combined with the dark ves-
sels and OD border will provide the maximum variance within a sub-image roughly