Escolar Documentos
Profissional Documentos
Cultura Documentos
3, MAY 2011
449
I. INTRODUCTION
LAUCOMA is caused by increased intraocular pressure
(IOP) due to the malfunction of the drainage structure of
the eyes [1]. It is estimated that more than four million Americans have glaucoma, and half of them are unaware that they
have it. Approximately 120 000 are blind from glaucoma, thus
accounting for 9%12% of all cases of blindness in the U.S.
About 2% of the population between 4050 years old and 8%
over 70 years old have elevated IOP [2], which increases their
risk of significant vision loss and even blindness.
Many studies have been conducted to develop computerbased decision support systems for the early detection of glaucoma. An artificial neural network (ANN) model using multifo-
cal visual evoked potential (M-VEP) data from the ObjectiVision perimetry system was studied in [3]. Their results showed
that a neural network model with M-VEP inputs was able to
detect glaucoma with a high sensitivity of 95%.
Qualitative assessment of the ability of optic nerve head
(ONH) stereo photographs (ONHSPs), confocal scanning laser
ophthalmoscopy (CSLO), scanning laser polarimetry (SLP), and
optical coherence tomography (OCT) to distinguish normal eyes
from those with early-to-moderate glaucomatous visual field
defects was performed in [4]. Receiver operating characteristic (ROC) curves were generated from discriminant analysis of
CSLO, SLP, and OCT measurements and from ONHP scores to
test the performance. It was found that the quantitative methods
CSLO, SLP, and OCT were not better than the qualitative assessment of disc ONHSPs conducted by observers experienced
at distinguishing normal eyes from those with early-to-moderate
glaucoma. They proposed that the combination of the imaging
methods might significantly improve this capability.
The optic disc topography parameters of Heidelberg retina
tomograph and neural network combination were used to differentiate between glaucomatous and nonglaucomatous eyes [5].
ROC curves were generated for the classification of eyes by three
neural network techniques: linear and Gaussian support vector
machines (Linear SVM and Gaussian SVM, respectively), a
multilayer perceptron (MLP), and linear discriminant function.
Bowd et al. showed that these neural network analyses helped
increase the diagnostic accuracy of glaucoma tests.
The performance of an ANN trained to recognize glaucomatous visual field defects was studied and its diagnostic accuracy
was compared with that of other algorithms proposed for the
detection of visual field loss [6]. ANN recorded a sensitivity of
93% at a specificity level of 94% and an area under the ROC
(AROC) curve of 0.984. The results of other compared methods
are as follows. The glaucoma Hemifield test had a sensitivity
of 92% at 91% specificity. The pattern standard deviation had a
sensitivity of 89% at 93% specificity. The cluster algorithm had
a sensitivity of 95% at 82% specificity.
An adaptive neuro-fuzzy inference system (ANFIS) was designed to differentiate between normal and glaucomatous eyes
from the quantitative assessment of summary data reports obtained using stratus OCT in Taiwan Chinese population [7].
Retinal nerve fiber layer thickness and ONH topography obtained using stratus OCT were used as input to the classifiers.
Two types of classifiers were studied: 1) ANFIS and 2) a backpropagation gradient descent method combined with the least
squares method. The results showed that ANFIS was better at
discriminating glaucomatous and normal eyes when using if
then rules and membership functions, which enhance the readability of the output.
450
Fig. 1.
IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 15, NO. 3, MAY 2011
An artificial intelligence system involving ANN and the analysis of the nerve fibers of the retina from the study with SLP
(NFAII; GDx), perimetry, and clinical data was developed [8].
The groups were defined as follows. Normal eyes were considered stage 0 and ocular hypertension as stage 1. Early glaucoma
was considered stage 2, and established glaucoma as stage 3.
Advanced glaucoma was considered stage 4, and terminal glaucoma as stage 5. The MLP using the LevenbergMarquardt
technique was used. The 100% specificity and sensitivity obtained indicates that 100% correct classification of each eye in
the corresponding stage of glaucoma was achieved.
An algorithm to detect the glaucoma using morphological
image processing was developed using fundus images. Nayak
et al. [1] used the cup-to-disc (c/d) ratio, the ratio of the distance
between the optic disc center and ONH to diameter of the optic disc, and the ratio of blood vessels area in inferiorsuperior
side to area of blood vessel in the nasal-temporal side as features in a neural network. The developed neural network system
identified glaucoma automatically with a sensitivity and specificity of 100% and 80%, respectively. This result implies that
the system can detect all subjects with glaucoma accurately, but
can detect only 80% of the normal subjects as normal. A new
framework for the detection of glaucoma based on the changes
in the ONH using the method of proper orthogonal decomposition was proposed in [9]. Any glaucomatous changes in the
ONH present during a follow-up examination were estimated
by comparing the follow-up ONH topography with its baseline
topograph subspace representation that was constructed earlier.
The changes in the ONH were quantified using image correspondence measuresL1-norm and L2-norm, correlation, and
image Euclidean distance (IMED). By using the L2-norm and
IMED in the new framework, good areas under the ROC curve
of 0.94 at 10 field of imaging, and 0.91 at 15 field of imaging
were obtained.
Linear discriminant analysis (LDA) and an ANN were used to
improve the differentiation between glaucomatous and normal
eyes in a Taiwan Chinese population based on the retinal nerve
fiber layer thickness measurement data from the SLP variable
corneal compensation [10].
The results showed that the individual parameter with the
best AROC curve for differentiating between normal and glaucomatous eye was the nerve fiber indicator (NFI, 0.932). The
AROCs for the LDA and ANN methods were 0.950 and 0.970,
respectively. Hence, the NFI, ANN, and LDF methods demonstrated equal diagnostic power in glaucoma detection. In contrast
to the method in [9], our proposed technique does not use any
pixel-based comparison between baseline image and a follow-up
image estimated to detect classes. Hence, minor topographical
changes within one image will not affect diagnoses.
The scheme proposed in this study is shown in Fig. 1. Higher
order spectra (HOS) based and texture-based features are com-
Fig. 2.
monly used in many medical image-processing areas. However, such studies have not yet been done on glaucoma images.
Therefore, these features were extracted in our study. After preprocessing the acquired fundus images, HOS-based and texturebased features are extracted from the preprocessed images. Subsequently, these features are fed to SVM, sequential minimal
optimization (SMO), random forest, and naive Bayesian (NB)
classifiers for classification. Feature ranking is also performed
to highlight and employ the discriminatory ability of the features
in the classification process. The layout of the paper is as follows. Section II contains an explanation of the data acquisition
process. In Section III, we address the preprocessing of the raw
normal and glaucoma images. Section IV contains an explanation of the extraction of the features using the HOS and texture
methods. A brief description of the various classifiers used is
discussed in Section V, and Section VI of the paper presents the
results of the proposed method. Finally, the paper concludes in
Section VII.
II. MATERIAL USED
The digital retinal images were collected from the Kasturba
Medical College, Manipal, India. We have used 60 fundus images: 30 normal and 30 open-angle glaucoma images from 20to-70-year-old subjects. The doctors in the Ophthalmology Department of the hospital certified the image quality and their
usability. The ethics committee, consisting of senior doctors,
approved the images for this research purpose. All the images
were taken with a resolution of 560 720 pixels and stored in
JPEG format. The fundus camera along with a microscope and
light source were used to acquire the retinal images to diagnose
diabetes retinopathy, glaucoma, etc. Fig. 2(a) and (b) shows the
typical normal and glaucoma fundus images, respectively.
III. IMAGE PROCESSING
The preprocessing step consists of image contrast improvement using histogram equalization and radon transform was
performed for HOS feature extraction. Histogram equalization
increases the dynamic range of the histogram of an image [11]
and assigns the intensity values of pixels in the input image such
ACHARYA et al.: AUTOMATED DIAGNOSIS OF GLAUCOMA USING TEXTURE AND HIGHER ORDER SPECTRA FEATURES
451
ity) and, hence, the entropies (Ent1, Ent2, and Ent3) computed
are also between 0 and 1.
B. Texture Features
that the output image contains a uniform distribution of intensities. As a result, the contrast of the image is increased. Radon
transform is used to detect features within an image. Radon
transform transforms lines through an image to points in the
radon domain, where each point in this domain is transformed
to a straight line in the image [11], [12]. This radon transformation is used before extracting the HOS parameters from the
image.
IV. FEATURE EXTRACTION
In this study, we have extracted two types of features:
1) HOS parameters and 2) texture descriptors. Brief explanations of these features are given in the following.
A. Higher Order Spectra
HOS elicits both amplitude and phase information of a given
signal. It offers good noise immunity and yields good results,
even for weak and noisy signals. HOS consist of moment and
cumulant spectra and can be used for both deterministic signals
and random processes [13]. We derived the features from the
third-order statistics of the signal, namely, the bispectrum. The
bispectrum is given by
B(f1 , f2 ) = E[X(f1 )X(f2 )X (f1 + f2 )]
(1)
where X(f) is the Fourier transform of the signal x(nT), and E[.]
stands for the expectation operation.
Features are calculated by integrating the bispectrum along
the dashed line with slope = a. Frequencies are normalized
by the Nyquist frequency (see Fig. 3). These bispectral invariants [14] contain information about the shape of the waveform
within the window and are invariant to shift and amplification
and robust to time-scale changes. In this study, we used these
bispectral invariant features for every 20 . Bispectral entropies
have been derived from bispectrum plots to find the rhythmic
nature of the heart rate variability and electroencephalogram
signals [15], [16]. The equations used to determine the various
HOS features are given in the Appendix [see (7)(11)]. The normalization in the equations ensures that entropy is calculated for
a parameter that lies between 0 and 1 (as required of a probabil-
(3)
Using (2) and (3), the following features were calculated: energy, contrast, homogeneity, entropy, and moments [see formulae given in Appendix: (12)(16)]. The difference vector, which
represents the gray-level difference statistics that can be obtained from the co-occurrence matrix can be derived using the
following equation [18]:
Cd (i, j)
(4)
Pd (k) =
i
452
IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 15, NO. 3, MAY 2011
Fig. 4.
Fig. 5.
and texture). HOS features are more widely distributed and have
very limited correlation among them. The texture features are
relatively correlated, but do exhibit discriminatory character for
each of the images. The classifiers were chosen based upon their
effectiveness in capturing the discriminative properties of these
features, the impact of the ranking of features and efficiency,
and the efficacy of the classification results. Four classifiers
were employed for supervised learning and testing: SVM [21],
SMO, random forest, and NB. Hardware consistency was maintained during the evaluation of these classifiers. The algorithmic control parameters for different classifiers are provided in
Table I, and a succinct discussion on these classifiers is presented
in the following.
SVMs perform classification by constructing an Ndimensional hyperplane that optimally separates the data into
two categories. SVMs view the classification problem as a
quadratic optimization problem. We chose SVM because of
its superior generalization in high-dimensional data and fast
convergence in training [22]. In general, SVMs plot the feature vector for each sample in the training set and result in a
high-dimensional feature space. Each feature vector is labeled
with its class identifier referred to as training IDs. A hyperplane
is drawn between the training IDs that maximizes the distance
between the different classes. The shape of the hyperplane is
generated by the kernel function, although many experiments
select the radial-basis kernel as optimal, which was employed
for our study. SMO speeds up the training of SVM [23], and
we implement the SMO algorithm by Platt [24] in our study.
The random-forest classifier discriminates classes by using a
collection of independent decision trees, instead of one tree,
where each tree is grown using a subset of the possible attributes. In random-forest classification, each tree votes for
one of the classes and, correspondingly, the most popular class
is assigned [25], [26]. NB is a statistical classifier based on
Bayes rules. The NB classifier can predict class membership
probabilities, such as the probability that a given tuple belongs
to a particular class, which is then employed for classification
purposes.
Since each of the features in the feature vector does not contribute equally to the classification process, we employ feature
ranking using the methods of chi-square, gain ratio, and information gain (IG). Consequently, each of the features is weighted
by their corresponding rank (normalized between 01). These
weighted features are then used for training and testing of instances. In an additional study, we investigated the impact of
normalization on the nonranked features. Lack of adequate normalization can skew the classification results causing significant
ACHARYA et al.: AUTOMATED DIAGNOSIS OF GLAUCOMA USING TEXTURE AND HIGHER ORDER SPECTRA FEATURES
TABLE II
RESULTS OF HOS FEATURES FOR NORMAL AND GLAUCOMA CLASSES
453
TABLE III
RESULTS OF TEXTURE FEATURES FOR NORMAL AND GLAUCOMA CLASSES
TABLE IV
CLASSIFICATION ACCURACIES (%) FOR DIFFERENT CLASSIFIERS
WITH/WITHOUT FEATURE RANKING AND WITH/WITHOUT FEATURE
SELECTION
v A
sA
(5)
(6)
be seen from the table that all the features show low p-values,
which indicate that there is a clinically significant difference
between the means of the two classes. In addition, Table III
shows the features extracted from the texture of the fundus
image. These features also show significantly low p-values. All
the texture features show higher variation for glaucoma images
as compared to the normal fundus image. However, most of the
HOS parameters show lower values for the glaucoma than the
normal images.
The ratio of the diameter of the optic cup to that of the optic
disc in a healthy eye is generally less than 0.5 [27]. When
the optic nerve is damaged by glaucoma, many of the individual
fibers that make up the nerve are lost and the optic nerve becomes
excavated. As glaucoma progresses, more optic nerve tissue
is lost and the optic cup grows larger. Thus, the cup-to-disc
ratio is higher for glaucoma subjects than for normal subjects,
thus leading to differences in the respective fundus images.
Moreover, in the case of optic nerve hemorrhage, another sign of
glaucoma-related damage, the blood typically collects along the
individual nerve fibers that radiate outward from the nerve. Such
physiological changes are manifested in the fundus images, and
our experiments show that the HOS and texture features are able
to quantify such differences in eye physiology.
Table IV summarizes our classification results with and without feature ranking, and with and without feature selection.
We applied fivefold cross validation for training and testing
purposes. LibSVM outperformed the other classifiers without
the ranking of features, while random forest performed better
than other methods with IG feature ranking. However, featureranking methods do not improve the performance of other
454
IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 15, NO. 3, MAY 2011
TABLE V
CLASSIFICATION ACCURACIES (%) OF CLASSIFIERS AFTER NORMALIZATION
AND FEATURE SELECTION OR FEATURE RANKING
Mean of magnitude:
Phase entropy:
Pe =
1
|B(f1 , f2 )|.
L
(7)
(8)
Normalized bispectral entropy (BE 1): Ent1 =
pn log pn
n
(9)
where pn = (|S(f1 , f2 )|)/( |B(f1 , f2 )|), and W is the region as shown in Fig. 3.
Normalized bispectral squared entropy (BE 2):
qn log qn
Ent2 =
(10)
where qn = (|B(f1 , f2 )|2 )/( |B(f1 , f2 )|2 ).
Normalized bispectral cubic entropy (BE 3):
rn log rn
Ent3 =
(11)
where rn = (|B(f1 , f2 )|3 )/( |B(f1 , f2 )|3 ).
Co-occurrence matrix based features:
[Pd (i, j)]2 .
Energy: E =
i
Contrast: Co =
i
Homogeneity:
H=
En =
(13)
i
Entropy:
(12)
i
Pd (i, j)
.
1 + (i j)2
(14)
(15)
Moments m1 , m2 , m3 , and m4 :
(i j)g Pd (i, j). (16)
mg =
i
ACHARYA et al.: AUTOMATED DIAGNOSIS OF GLAUCOMA USING TEXTURE AND HIGHER ORDER SPECTRA FEATURES
Difference-vector-based features:
Angular second moment:
n 1
Pd (k)2 .
(17)
k =0
Contrast:
n 1
k 2 Pd (k).
(18)
k =0
Entropy:
n 1
(19)
k =0
Mean:
n 1
kPd (k).
(20)
k =0
Run-length-matrix-based features:
(P (i, j)/j 2 )
i
j
Short-run emphasis:
.
i
j P (i, j)
2
j P (i, j)
i j
Long-run emphasis:
i
j P (i, j)
2
i
j P (i, j)
Gray-level nonuniformity:
.
i
j P (i, j)
P (i, j)]2
j [
i
Run-length nonuniformity:
i
j P (i, j)
i
j P (i, j)
Run percentage:
A
where A is the area of the image of interest.
(21)
(22)
(23)
(24)
(25)
REFERENCES
[1] J. Nayak, U. R. Acharya, P. S. Bhat, A. Shetty, and T. C. Lim, Automated
diagnosis of glaucoma using digital fundus images, J. Med. Syst., vol. 33,
no. 5, pp. 337346, Aug. 2009.
[2] Glaucoma research foundation. (2009). [Online]. Available: http://www.
glaucoma.org/learn/glaucoma_facts.php.
[3] R. Nagarajan, C. Balachandran, D. Gunaratnam, A. Klistorner, and
S. Graham, Neural network model for early detection of glaucoma using
multi-focal visual evoked potential (M-VEP), Invest. Ophthalmol. Vis.
Sci., vol. 42, 2002.
[4] M. J. Greany, D. C. Hoffman, D. F. Garway-Heath, M. Nakla, A. L.
Coleman, and J. Caprioli, Comparisons of optic nerve imaging methods to
distinguish normal eyes from those with glaucoma, Invest. Ophthalmol.
Vis. Sci., vol. 43, no. 1, pp. 140145, Jan. 2002.
[5] C. Bowd, K. Chan, L. M. Zangwill, M. H. Goldbaum, T. W. Lee,
T. J. Sejnowski, and R. N. Weinreb, Comparing neural networks and
linear discriminant functions for glaucoma detection using confocal scanning laser ophthalmoscopy of the optic disc, Invest. Ophthalmol. Vis.
Sci., vol. 43, pp. 34443454, 2002.
[6] D. Bizios, A. Heijl, and B. Bengtsson, Trained artificial neural network
for glaucoma diagnosis using visual field data: A comparison with conventional algorithms, J. Glaucoma, vol. 16, no. 1, pp. 2028, Jan. 2007.
[7] M. L. Huang, H. Y. Chen, and J. J. Huan, Glaucoma detection using
adaptive neuro-fuzzy inference system, Expert Syst. Appl., vol. 32, no. 2,
pp. 458468, Jan. 2007.
455