Page 258 - Computational Retinal Image Analysis

256 CHAPTER 13 Drusen and macular degeneration

post-processing step, the authors remove blood vessels using a Canny edge detector
and morphological processing. They evaluated their approach using one image from
a healthy patient and six images from six AMD-affected patients. Each was trained
on 600 vectors; anomalies were found for the AMD patients while none was found
for the healthy patient.
Using a larger dataset, Cheng et al. [44] aimed to detect drusen, and hence AMD,
using biologically inspired features. They first detect the optic disc and macula in order
to zoom in on the macula for feature extraction. These extracted features are classified
using support vector machines (SVM). Tested on 350 images, the results demonstrate
the effectiveness for drusen detection with a sensitivity/specificity of 86.3%/91.9%.
Following on from this, a considerable amount of work was published in 2013–14
on drusen detection using Machine Learning techniques, particularly focused on
SVM. Many of these approaches followed a framework of pre-processing, using a
filter bank to extract features, and classification via SVM. Akram et al. [45] pre-
sented a method for drusen detection in retinal color images. After pre-processing
for contrast enhancement, they used a filter bank to extract potential drusen regions
and eliminated false positives relating to similarity with the optic disc. An SVM ap-
proach was then used to classify these regions as drusen or not. The authors evaluate
the system’s performance on the STARE database, achieving accuracy/sensitivity/
specificity of 97%/95%/98.4%. Raza et al. [46] proposed a hybrid classifier tech-
nique for drusen detection from fundus images by using a Gabor kernel based filter
bank and eliminating spurious regions, which may be confused with drusen. Their
system represented each region with a number of features and then applied hybrid
classifier as an ensemble of Naive Bayes and SVM to classify the regions as either
drusen or non-drusen. The proposed system was evaluated on the STARE database,
achieving sensitivity/specificity/accuracy of 97%/99%/98%. Waseem et al. [47] pre-
sented a method for the detection of soft and hard drusen. After some pre-processing,
the proposed method computed color and Gabor filter-based features for each pixel
to classify and extract all possible drusen pixels. Connected component labeling was
used to remove the suspicious pixels from the drusen region. Finally, the optic disk
was removed to avoid false positives. Performance was evaluated on STARE, achiev-
ing 96% accuracy for drusen detection.
Zheng et al. [48] developed an automated drusen detection system, aiming to au-
tomatically assess the risk of developing AMD. The authors develop a learning-based
system incorporating both optimal color descriptors and robust multiscale local im-
age descriptors. Their multi-step system included a considerable pre-processing step
involving denoising, generation of the retinal mask, correcting illumination and color
transfer. This was followed by feature selection in a pixel-wise way using Adaboost
[49] and in a region-based way using LS-SVM [50]. The authors evaluated their sys-
tem with color fundus photographs from two AMD clinical studies [51, 52] against
manually drawn segmentations. Using a leave-one-out strategy, they achieved a mean
accuracy of 80% compared to a trained grader on [51] and mean accuracies of 86%
and 83% on [52] compared to an ophthalmologist and trained grader respectively.
Each of these outperformed the well-known STARE [53] and HALT [36] studies.