Page 261 - Computational Retinal Image Analysis

3 Automatic analysis of drusen and amd-related pathologies 259

a European multi-center database of fundus photographs which either had no AMD
or moderate AMD. The results were evaluated by comparison with two expert grad-
ers. The system was demonstrated to approach the performance of the human observ-
ers in detecting drusen and the estimated area showed excellent agreement, achieving
sensitivity/specificity of 85/96. The authors also computed an automatic AMD
risk assessment, achieving area under the receiver operating characteristic curves
(AUROCs) of 94.8% and 95.4%.
Bhuiyan et al. [68] aimed to quantify drusen area and classify them as intermedi-
ate or soft drusen using a region growing technique. The authors used local intensity
distribution, adaptive intensity thresholding and edge information to detect potential
drusen areas. 50 images with various types of drusen were used to test detection and
12 of these were selected to evaluate the segmentation. The images were annotated
by an expert grader. The method achieved 100% accuracy for detection and accura-
cies of 79.6% and 82.1% for classifying intermediate and soft drusen respectively.

3.3.1 Texture-based methods
Texture has commonly been used for distinguishing drusen from the background.
Parvathi and Devi [69] presented two methods for detection and counting of drusen,
exploiting their morphological characteristics such as texture and their 3D profiles.
In the first method, the authors attempt to characterize the drusen by texture and
use a multi-channel filtering technique; in the second, they characterize drusen by
its topographic profile and use a curvature-based detection method. The authors of
[70] developed a spatially adaptive algorithm for drusen detection based on GLCM
based textural features. The accuracy of the classifier was improved using OD lo-
calization and blood vessels with morphological operators. The authors evaluated
performance against a hand-labeled ground truth, achieving an accuracy of 98.05%
on 120 samples. Lee et al. [71] presented a method for learning non-homogenous
textures, mimicking the idea of selective learning, performing probabilistic boosting
and structural similarity clustering for fast selective learning. They applied their idea
to drusen segmentation.
Garnier et al. [72] presented a preliminary study for AMD detection from color
fundus photographs using a multiresolution texture analysis and wavelet decomposi-
tion. To avoid dimensionality problems, the authors use Linear Discriminant Analysis
for feature dimension reduction, followed by image classification. They tested their
method using a dataset of 45 images (23 healthy, 22 diseased). Significantly, they
used images from different cameras and of varying quality, achieving a recognition
rate of 93.3%, with sensitivity/specificity of 91.3%/95.5%.
3.4 Other imaging modalities
3.4.1 Angiography
While color fundus has been the dominant modality for investigating drusen, other
technologies have also been explored. The authors of [73] presented a 3-stage ap-
proach to segmenting drusen in retinal angiographic images taken in 1983 and 1988
respectively in order to track changes. Their method involved an optimal partitioning