Page 89 - Computational Retinal Image Analysis

3 Retinal imaging databases 81

• OD detection to aid in the detection of other landmarks—As previously
mentioned, the OD and fovea share a geometric relationship that can be utilized
to limit the fovea search area once the OD has been found. The OD is also
the entrance point of the main retinal vessels onto the retina. Starting a search
around the found OD can help to initialize the detection of these vessels [3].
• OD detection/segmentation for disease detection—To determine a glaucoma
suspect from fundus photography, features in and around the OD are used as
biomarkers. These include disc elongation, vascular nasal shift, cup-to-disc
ratio, disc damage likelihood scale and damage to the neuro-retinal rim [4]. In
diabetic retinopathy (DR), neovascularization at the disc is important sign of
proliferation of the disease [5]. Also, analysis of the OD boundary can aid in the
detection of OD swelling in papilledema [6].
• OD and Fovea detection/segmentation to delineate areas of interest—
One OD diameter is a commonly used measurement to determines areas of
interest for finding disease or making measurements. In DR, hard exudates
within one disc diameter of the fovea center are used as markers for clinically
significant macular edema [7]. In hypertensive retinopathy, arteriovenous
ratio measurements are usually taken between one to two disc radii from the
boundary of the OD [8].
• OD detection for preprocessing—Masking the OD out of the image can
be helpful to future processing steps as it removes a bright region with sharp
contrast that could be falsely detected as part of the vasculature or as pathology.
The OD region may also be completely saturated which can affect other
processing steps, such as image normalization or background correction.

3 Retinal imaging databases

Over the years, multiple databases have been put together and made available to
researchers to enable them to test their developed algorithms. These databases also
allow different methods to be tested against one another to objectively measure per-
formance. Each database has a different focus that is not necessarily landmark de-
tection, but generally provide OD location ground truth and real-world images with
different amounts of disease and levels of image quality. The databases listed below
are those that are generally accepted and tested against in the literature:

DRIVE: Digital Retinal Images for Vessel Extraction—Although the DRIVE
dataset (2004) was created to compare vessel segmentation algorithms, it
has been used to benchmark OD detection algorithms for a long time and is
included for this reason. The dataset consists 40 images with little or no DR.
Images were acquired on a Canon CR5 non-mydriatic camera at 45° field
of view and 768 × 584 pixel resolution [9]. https://www.isi.uu.nl/Research/
Databases/DRIVE/
Messidor—The Messidor database (2008) was made publicly available in 2008
and consists of 1200 retinal images (800 with dilation, 400 without). All images