Page 257 - Computational Retinal Image Analysis

3 Automatic analysis of drusen and amd-related pathologies 255

3.1.1 Characterization, classification and quantification of drusen
The distinction of different types of detected lesion is an important task for disease
diagnosis. Quellec et al. [39] proposed a framework for not only detecting but also
characterizing target lesions. Their method relies on a feature space derived from ref-
erence image samples of target lesions. These were obtained from both expert- and
data-driven approaches. The authors derived filters using Factor analysis to classify
the image samples. This approach was tested for detecting microaneurysms on im-
ages from 2739 patients including 67 with referable diabetic retinopathy (DR), and
for differentiating drusen from Stargardt’s disease lesions on a set of 300 manually
detected drusen and 300 manually detected flecks. A key benefit of this method is
the speed, taking less than 1 s on a standard PC. Comparable performance was ob-
tained for both the expert- and data-driven approaches, indicating that annotation of
a limited number of lesions suffices for building a detection system for any type of
lesion in retinal images. This is significant for cases where no expert-knowledge is
available.
Deepak et al. [40] were interested in anomaly detection as the first step in medi-
cal image evaluation for diagnosis, followed by disease-specific anomaly evaluation.
Motivated by the idea that experts primarily focus on abnormal structures during vi-
sual examination, they aimed to model this behavior in automated image analysis by
visual saliency. The authors propose a framework for detecting abnormalities using
visual saliency computation for sparse representation of the image data, preserving
the essential features of a normal image. They evaluate their approach for bright le-
sion detection and classification in retinal fundus images on five publicly available
datasets, achieving accuracies from 93% to 96% for lesion discrimination, demon-
strating the potential of visual saliency in automated abnormality detection.
An important goal for the diagnosis of diseases such as AMD and DR is quan-
tification of the found anomalies or pathologies. In the case of drusen detection for
AMD diagnosis, counting the drusen is an aim that is achievable given successful
detection. In 2004, Barakat et al. [41] developed a method for automatically quanti-
fying drusen for clinical trials using region of interest (ROI) Selection and intensity
based thresholding. Over 20 test images, they achieved a sensitivity/specificity of
86%/96%.

3.1.2 Machine learning based approaches
Beyond saliency, clustering and intensity-based thresholding approaches, machine
learning has been a popular basis for designing methods of drusen detection in retinal
Fundus images, due to the impressive results achieved in other fields and the speed
at which results can be obtained.
In order to find anomalies in the imaging of AMD patients vs healthy controls,
Freund et al. [42] proposed a two-step approach following pre-processing of the
data by selecting the green channel and using intensity based equalization. In the
first step, using multiscale analysis to build a feature vector of the image response
to filtering at different scales. This was followed by a kernel-based anomaly detec-
tion approach based on a Support Vector Data Description [43] test statistic. As a