Page 256 - Computational Retinal Image Analysis

254 CHAPTER 13 Drusen and macular degeneration

Much of the work in this area to date has been in the detection and assessment of
drusen, which appear as yellow deposits in the retina when viewed via fundus pho-
tography. Drusen vary in size, number and confluence degree, and are a dominant
characteristic of AMD. The early detection of drusen is useful in diagnosing AMD,
and fundamental to many potential systems for automatically providing a diagnosis
of AMD. There are several common, related aims in the computer assisted analysis
of drusen: (i) identification of the presence of drusen in fundus photography and lo-
calization, (ii) quantification/counting of drusen, (iii) segmentation and measurement
of drusen, and (iv) attempting to link such measurements to a diagnosis of AMD.

3.1 Drusen detection in retinal fundus photography
Given its prevalence in ophthalmology for visualizing drusen, fundus photogra-
phy has been the focus for the majority of work in detecting drusen and a huge
variety of techniques have been developed. Intensity filtering and thresholding
were the basis of many techniques in the earlier work, such as that of Rapantzikos
et al. [36], who in 2003 presented a novel segmentation algorithm for automati-
cally detecting and mapping drusen in retina fundus images. The authors use a
multilevel histogram equalization to enhance local intensity structures and de-
velop a novel segmentation technique called histogram-based adaptive local
thresholding (HALT). This involves extracting the useful information from an im-
age without being affected by the presence of other structures and is followed by
median filtering. The algorithm was tested on 23 fundus images, including images
from eight pairs of eyes, achieving an impressive overall sensitivity/specificity of
98.85%/99.32%.
In 2004, Mora et al. [37] developed a methodology for automatic unsupervised
detection and modeling of Drusen spots for evaluating therapy effectiveness. A
gradient-based image segmentation technique is firstly used to detect Drusen spots
and facilitate estimation of their strength. For each detected spot, the authors opti-
mize a 3D model with a Gaussian function using Levenberg-Marquardt nonlinear
least-squares. The drusen is then shown using a contour of the Gaussian functions
at a pre-defined amplitude percentage. The results were demonstrated on only two
images. Later, in 2009, Remeseiro et al. [38] proposed a method of automatically
detecting drusen from retinal images using a top-down strategy, Gaussian template
matching and thresholding. They achieved sensitivity/specificity of 83%/87% calcu-
lated over four images.
It is important to note that, while these works have achieved good quality results,
the algorithms were tested on only a limited number of images. In order to demon-
strate robustness and real applicability, it is increasingly more important to develop
and evaluate medical image analysis algorithms on larger datasets, including multi-
ple stages of disease and types of pathologies, including healthy cases. For example,
the more recent works of Quellec et al. [39] and Deepak et al. [40] have evaluated
their models on multiple patients and multiple datasets, showing pathologies of dif-
ferent diseases and types of lesion.