Page 260 - Computational Retinal Image Analysis

258 CHAPTER 13 Drusen and macular degeneration

containing many types of drusen and images without drusen or other lesions. The
authors achieved a correct detection rate of 87%. Smith et al. [58] aimed to segment
drusen by developing an interactive semi-automated method. The authors used 12
color fundus photographs of patients with AMD and drusen, sequentially raised the
gray scale values of all structures within defined elliptical boundaries until a uni-
form background was obtained, and segmented the drusen in the central and middle
subfields (at 1 and 3 mm diameters) by uniform thresholding. They compared their
results with two expert graders in terms of area achieving good results.
The authors of [59] propose a three-phase method for segmenting the drusen
so that they can be quantified in terms of their sizes, area and number. The authors
pre-processing the images to remove artifacts such as non-uniform illumination and
noise and improve the quality of the image using homomorphic and Gaussian filter-
ing. The authors also took the common approach of selecting only the green channel.
Drusen candidate edge detection was performed by determining the gradient, sup-
pressing pixels that were not local maxima and thresholding. Hysteresis was used to
track the edges to remove breaks caused by weak points. Boundaries of the drusen
were then extracted using iterative edge thinning and end-point recovery and label-
ing, and edge linking. The method was tested on images containing drusen with 37
selected from the Structured Analysis of Retina (STARE) dataset [60, 61] and 11
from the Automated Retinal Image Analysis (ARIA) [62, 63] dataset. The reported
results for drusen detection demonstrated good accuracy/sensitivity/specificity of
96.2%/89.8%/99.0% on STARE and ARIA. The authors further quantified the size of
detected drusen as small, intermediate and large with an accuracy of 88.5%, 98.6%,
and 88.37%, respectively.
Mora et al. [64] also aimed to measure drusen with illumination correction and
gradient-based path labeling. Testing on 22 images, they achieved sensitivity/speci-
ficity of 68%/96%. Kumari and Mittal [65] used Otsu thresholding, morphological
operations and pixel-wise feature extraction to find the boundaries of drusen and
measure the area, achieving sensitivity/specificity/accuracy of 95%/97%/96.3%.
Köse et al. [66] proposed an inverse segmentation method to identify healthy and
unhealthy areas in the AMD fundus image. Initially, the optic disc and macula are
located using the Sobel filter and relative locations of these two structures. Further,
region-growing algorithm is initialized around macula to segment out healthy and
unhealthy regions. Next, the inverse of segmented image is taken to generate the
unhealthy segmentation result. Their method achieved a segmentation accuracy of
more than 90%.

3.3 Quantifying drusen area and distinguishing drusen type

Distinguishing the type of drusen present is an important task in early and intermedi-
ate AMD detection. Van Grinsven et al. [67] aimed to develop a method of detecting
drusen and quantifying their location, area, and size to assist the computer aided
diagnosis of non-advanced AMD. The authors developed a machine learning system
was developed to automatically detect and quantify drusen on 407 randomly selected