Page 259 - Computational Retinal Image Analysis

3 Automatic analysis of drusen and amd-related pathologies 257

ROI identification k-means clustering was used to develop a system called
THALIA [54] for the automatic detection of drusen for AMD assessment. The system
first detects the macular region of interest using a seeded mode tracking approach,
and then map this into a new representation using a hierarchical word transform
(HWI) which generates generate structured pixels which embed local context. These
structured pixels are then clustered using hierarchical k-means clustering, and these
clusters are classified using a SVM-based classifier. THALIA was evaluated on a
dataset of 350 images, achieving an accuracy of 95.5%.
The idea of using SVM was extended to AMD detection by Mookiah et al.
[55], who proposed an automated AMD detection system using a discrete wavelet
transform (DWT) and feature ranking. Many statistical measures including the first
four-order statistical moments, energy, entropy, and Gini index-based features were
extracted. Five feature-ranking strategies were used to identify an optimal feature
set. SVM, decision tree, k-nearest neighbor (k-NN), Naive Bayes, and probabilistic
neural network were used to evaluate the highest performance measure using the
minimum number of features in classifying normal and dry AMD classes. The pro-
posed framework obtained an average accuracy of 93.7%, sensitivity of 91.1%, and
specificity of 96.3% using KLD ranking and SVM classifier.
Methods have also been developed toward detecting drusen for diabetic retinopa-
thy diagnosis. Niemeijer et al. [56] presented a machine learning-based system to
automatically detect exudates and cotton-wool spots in color fundus photographs and
differentiate them from drusen, aiming for the early diagnosis of diabetic retinopathy.
The authors used 300 retinal images from 300 diabetic patients (1 eye per patient)
from a diabetic retinopathy tele-diagnosis database. 100 with lesions and 200 with-
out. The images were annotated manually by two expert graders to reach a consensus.
They developed a machine learning approach involving a k-nearest neighbor classi-
fier to classify pixels, producing a probability map that was used to group lesion
clusters. Based on these characteristics, each probably lesion cluster was assigned a
probability that the cluster was in fact a lesion using a linear discriminant classifier.
Finally, each cluster was distinguished as (hard) exudate, cotton-wool spot or drusen.
They achieved sensitivity/specificity of 95%/88% for lesion detection, 95%/86%
for exudate, 70%/93% for cotton-wool spots, and 77%/88% exudates. These results
were generally lower than, but similar to, the performance of a third retinal special-
ists annotation results of 95%/74% for lesions, 90%/98% for exudates, 87%/98% for
cotton wool spots, and 92%/79% for drusen.

3.2 Drusen segmentation and measurement

Brandon and Hoover [57] aimed to automatically detect and segment drusen in reti-
nal fundus images in an unsupervised way. They used a multi-level approach, be-
ginning with pixel level classification and proceeding to region, area, and finally
image level. This is to allow the classification to be tuned to detect faint drusen and
those that are difficult to distinguish from the background, relying upon the higher
levels of classification for broader context. The method was tested on 119 images