Page 265 - Computational Retinal Image Analysis

4 Diagnosis of AMD 263

Other data mining techniques based on spatial histogram and Dynamic Time
Warping (DTW) were previously used in [88]. The authors of [89] proposed an au-
tomated dry AMD detection system using several entropies, Higher Order Spectra
features, Fractal Dimension, and Gabor wavelet features, which were ranked using
many ranking methods to select optimum features for classification as normal or AMD.
Many classifiers were considered, including Probabilistic Neural Networks and SVMs.
Performance was evaluated using a private dataset, as well as the ARIA and STARE
datasets. The highest average classification accuracies of 90.19%, 95.07% and 95%
were achieved using SVM for the private dataset, ARIA and STARE respectively. All
these studies are quite small in terms of the number of images (patients) involved [90].
The use of FA for the diagnosis of AMD is rare, primarily because it is invasive
in nature and typically used in the later stages. However, there has been some work
to develop methods for the differentiable diagnosis of AMD, e.g. neovascularization.
With the advent of OCT, it has become indispensable in the management of AMD
and in particular the decision-making regarding anti-VEGF treatment. Venhuizen
et al. have proposed an unsupervised approach [91] and Bag of feature approach
[92], where the latter was applied to a total of 3265 OCT scans from 1016 patients
with either no signs of AMD or with signs of early, intermediate, or advanced AMD.
The participants were randomly selected from a large European multicenter data-
base. Different from all the other approaches, Albarrak et al. proposed an OCT-tree
based method for the classification of three-dimensional OCT images directly and
provided promising results [93].
Although these methods have avoided the challenge posed by segmentation, they
still rely on certain features either directly (intensity, texture) or indirectly (graph
patterns), which still requires the knowledge and insight of the researchers. The re-
cent success of deep learning in computer vision has had a great impact in this field.
Recently, there have been four studies that used deep learning techniques for
the diagnosis of AMD. Lee et al. selected 52,690 normal macular OCT images and
48,312 AMD macular OCT images from 2.6 million OCT images linked to clinical
data points from the electronic medical records (EMRs) [94]. A modified version of
the VGG16 convolutional neural network (CNN) [95] was trained to categorize im-
ages as either normal or showing AMD. At the image level, the authors achieved an
AUC of 92.78% with an accuracy of 87.63%. At the patient level, an AUC of 97.45%
with an accuracy of 93.45% was achieved. Peak sensitivity and specificity with opti-
mal cut offs were 92.64% and 93.69%, respectively.
Grassman et al. reported an automated AMD classification algorithm using en-
semble learning [96], evaluating on a cross-sectional population-based study. 120,656
manually graded color fundus images were used from 3654 Age-Related Eye Disease
(AREDS) patients and 14 classes were defined, including 3 late AMD stages and
an ungradable class. Several architectures were tried but an ensemble was found to
obtain improved results. The authors also evaluated their method on a subset of the
Kooperative Gesundheitsforschung in der Region Ausburg (KORA) dataset, which
was not used for training. Good results were achieved, with the algorithm detecting
84% of images with definite signs of early and late AMD, and 94% of healthy images.