Proceedings of the World Congress on Engineering and Computer Science 2013 Vol I

WCECS 2013, 23-25 October, 2013, San Francisco, USA

Automated Classification of Normal, Cataract

and Post Cataract Optical Eye Images using
SVM Classifier
Jagadish Nayak

Abstract— Eye is a very important organ of the human body, Fig. 1 shows the complete system for the automated
which has many complex sensory elements such as lens, retina classification of three classes of optical images such as
etc. Eye disorder is a prominent issue in the health care sector. normal, cataract and post cataract.
Cataract is an eye disorder, which occurs due to the clouding of
Optical Eye Normal
lens. Over a period of time, cataract will lead to reduced Feature SVM
Images Cataract
eyesight. If cataract is not treated in proper time, then it will Extractio Classifier
Post-cataract
lead to blindness. This is common in aged people. In this work
image processing techniques are used to detect the features in Figure 1. Proposed system for the automated
the three classes of optical eye images such as normal, cataract classification of optical eye images
and post-cataract images. The features of the optical eye image The normal eye [1,2,3,4] is made up of the sclera,
such as Big Ring Area (BRA), Small Ring Area (SRA), Edge cornea, pupil, aqueous humor, iris, conjunctiva, lens,
Pixel Count (EPC) and Object Perimeter are extracted. The vitreous humor, ciliary body, macula, retina, fovea and the
features are statistically analyzed and found to be significant
for the automatic classification. The same features are then
optic verve. Lens is the clear part of the eye behind the iris
used in the automatic classifier such as Support Vector that helps to focus light on the retina. The lens helps to focus
Machines (SVM) for the automatic classification. The results on both far and near objects so that they are perceived
are found to be clinically significant with 94% sensitivity and clearly and sharply. The ciliary muscle helps to change the
93.75% specificity. The classification rate is nearly 90%. shape of the lens. This changing of the lens shape is called
accommodation. It is said that the diameter of the lens is
Index Terms- Cataract, Eye disorder, Big Ring Area (BRA), 10mm. Fig. 2(a) shows the optical image of a typical normal
Small Ring Area (SRA), Edge Pixel Count (EPC) and Object eye.
Perimeter, SVM A cataract is a clouding of the natural lens, which lies
behind the iris and the pupil. The lens is contained in a
I. INTRODUCTION sealed bag or capsule. Fig. 2(b) shows the image of the eye
The research on medical images has been adopted by affected with cataract. The main symptom is gradual,
most of the scientists and physicians, which can assist in the painless vision blurring. As old cells die they become
detection of the abnormalities. The image processing trapped within the capsule. Over time, the cells accumulate
algorithms are developed to identify the disease specific causing the lens to cloud, making images look blurred.
features on the medical images, which can help in the Cataracts generally develop slowly over years. Early
detailed study of abnormalities. The cataract is an eye symptoms may be loss of contrast, glare, needing more light
disorder, which is basically detected by looking at the lens to see well, and problems distinguishing dark blue from
of the human eye. The anatomical changes occurred in the black. Later, progressive, painless blurring of vision occurs.[
lens due to the abnormality is one of the decision factors for 1,2,3,4].
the detection of cataract. Computer processing of a medical When the cataracts have progressed enough to seriously
image such as optical eye image can provide us with impair the vision and affects daily life, the cataract surgery
objective measure of the symptoms of the cataract. These is necessary. Cataract surgery is simple and relatively
objective features can be the decision factor for the painless procedure to regain vision. Cataract surgery is very
abnormality detection. Pattern recognition systems can be successful in restoring vision. During surgery, the surgeon
used for the automated detection of eye abnormality using will remove the clouded lens , and in most cases replace it
this feature vector. with a clear ,plastic intraocular lens (IOL). Fig 2(c) shows
the optical image of post cataract eye.[1,2,3,4]
Manuscript received February 23, 2013;
Jagadish Nayak is with the BITS PILANI Dubai, Dubai International
Acedemic City , Dubai, phone: +971 55 4907979 e-mail: jagadishnayak@
dubai.pilani-ac.in.

(a) (b) (c)

Fig. 2. Optical Images
(a) Normal (b) cataract (c) Post-cataract

ISBN: 978-988-19252-3-7 WCECS 2013

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the World Congress on Engineering and Computer Science 2013 Vol I
WCECS 2013, 23-25 October, 2013, San Francisco, USA

In 1989 , Nuclear Magnetic Resonance (NMR) microscopic inner surface of the cornea images is more whitish as
imaging method is used to detect the early stages of cataract, compared to that of normal and post-cataract images. The
but this method is limited by the resolution problem[5]. pictorial representation of SRA is given by the template as
Wong et al, have studied the galactose cataract using a shown in Figure 3. It is the region between Ci and Co with
conventional whole-body Magnetic Resonance Imaging ro=5 and ri= 15 in this work. Fig. 4(a), (b), and (c) shows
(MRI) operated at 0.6 Tesla [6]. The optical aberrations the results of computing SRA in normal, cataract and post
caused by cataract were detected and quantified objectively cataract images of different subjects respectively.
using a new system called focus detection system (FDS) [7].
Monestam et al, have compared the functional outcome of
cataract surgery in terms of visual ability between patients Co
younger than 84 years, 85 to 89 years, and 90+ years [8].
Best corrected visual acuity (BCVA), contrast sensitivity Ci ri
(CS) and functional visual complaints in early cataract and
ro
after the cataract surgery were examined [9]. Ferraro et al,
have assessed the validity of a digital non-mydriatic fundus
camera in detecting cataract as a cause of visual impairment
[10]. They showed that the non-mydriatic fundus camera Fig. 3 Template for small ring area detection
may be an alternative method for screening for visually
significant cataract in the community. An early transvaginal
anomaly scan at 14-16 gestational weeks has been proposed
to diagnose fetal eye anomalies (especially cataract) using
ultrasound [11]. Specifically, using the Ultrasonography
technique, cataract was detected in the foetus in more than
80% of the cases. A novel method for the automated (a) (b) (c)
classification of normal, cataract and post-cataract optical Fig. 4. Small ring area detection in (a) normal (b)
images is presented in this paper. Details of the algorithm cataract (c) pos-cataract images for different subjects
development and implementation is shown in Section II,
results are given in section III, discussion is given in section C. Detection of Big Ring Area
IV and conclusion is given in section V.
The color at the outer surface of the cornea is not the same
in all the three classes. In cataract images, the outer surface
II. MATERIALS AND METHOD of the cornea images is bright in color as compared to that of
The optical images were collected from Kasturba Medical the normal and Post-cataract images. The pictorial
College, Manipal, India. Total of 174 images of all the age representation of BRA is given by the template as shown in
group (both male and female) is collected. Out of those 56 Fig. 5. It consists of two circles: an inner circle Ci and an
images are normal, 89 images are of cataract images and 29 outer circle Co. In this work, the region between the Ci and
post-cataract images. All the images are of size 128x128, Co is the BRA with ro= 59 and ri= 55. Fig. 6(a), (b), and (c)
stored in the TIFF format. shows the schematic diagram of the BRA detection for
normal, cataract and Post-cataract images of different
A. Image Preprocessing subjects respectively.
Usually, there is a large variation in the color of eye images
among the patients. This variation is strongly correlated to
the person’s skin pigmentation and iris color. Hence, it is
mandatory to identify a reference frame and normalize the C
colors of all other images against it. One reference eye r
image is selected from the dataset and colors of all other r
images in the dataset are normalized in accordance with the C
reference image [12]. The contrast of the eye images are not Fig. 5. Template for big ring area detection
distributed evenly throughout the image. The objective of
this preprocessing was to reduce this effect and to normalize
the mean intensity. The intensities of the three color bands
were transformed to an intensity-hue-saturation
representation [13]. This allowed the intensity to be
processed without affecting the perceived relative color
values of the pixels.
B. Detection of Small Ring Area (a) (b) (c)
Fig. 6. Big Ring Area detection in (a) normal (b) cataract (c) pos-
The color at the inner surface of the cornea is not the
cataract images for different subjects
same in all the three kinds of images. In cataract images the

ISBN: 978-988-19252-3-7 WCECS 2013

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the World Congress on Engineering and Computer Science 2013 Vol I
WCECS 2013, 23-25 October, 2013, San Francisco, USA

D. Edge pixel count (EPC) applied in various fields like classification, regression etc
There are different methods available to extract the edges [16].
of the eye images namely: sobel, canny, prewitt, roberts, log,
etc. Canny method is the most powerful edge-detection III. RESULTS
method among them. It uses two different thresholds (to Table I shows the range of four features used for
detect strong and weak edges), and it includes the weak the classifications. It can be seen from the table that these
edges in the output only if they are connected to strong features show very low ‘p-value’ (p<0.0001) indicating that
edges. In the computation of EPC, we counted the number they are clinically significant. Table II shows the
of white pixels in the output of the edge detection. Fig. 7 classification results of the two classifiers. We used 125 eye
shows the three class of eye images with EPC detected. images for training and 49 features for testing. Our result
shows that we are able to detect the early stage of cataract
and classify normal, cataract and post-cataract eye images
with an average accuracy of 88.39%. The SVM classifier is
able to identify the (average) correct class to the tune of
88%. It is able to classify almost all the cataract and normal
(a) (b) (c) eye images correctly.
Fig. 7 Edge pixel count in (a) normal (b) cataract (c) Post cataract subjects
The sensitivity of a test is the proportion of people
E. Object Perimeter with the disease who have a positive test result. Higher the
Morphological image processing is a type of processing in sensitivity, greater will be the detection rate and the lower
which the spatial form or structure of objects within an the false negative (FN) rate. The specificity of the test is the
image is modified. Dilation, erosion and skeletonization are proportion of people without the disease who have a
the three fundamental morphological operations. With negative test. Higher the specificity, lower will be the false
dilation, an object grows uniformly in spatial extent, positive rate and the lower the proportion of people with the
whereas with erosion an object shrinks uniformly. disease who will be unnecessarily worried or exposed to
Skeletonization results in a stick figure representation of an unnecessary treatment. The MedCalc statistical software
object. The normal, cataract and post-cataract images have [17] was used for analysis. As we can see from the Table III
too many sudden changes in the gray levels. Hence, there that our system has the sensitivity of 94% and specificity of
will be many edges in these images which can be used as a 93.75% for all the classes indicating that the results are
feature to reflect normality and abnormalities in the eye clinically significant.
images. There are many different types of morphological
TABLE I
operations available, however; in this case, the best result
obtained was from using an operation called erosion. RANGES OF DIFFERENT PARAMETERS FOR NORMAL,
Erosion is performed using the structuring element of size CATARACT AND POST- CATARACT EYE CLASSES.
(3x3) with all ones. Fig. 8 shows the object perimeter of the
eye images of the three classes specified in this work. Features Post
Normal Cataract p value
Cataract
Big Ring 709.41 716.88 491 p<
Area ± 211 ±178 ±242 0.0001
Small Ring 277.75 336.30 153.97 p<
Area ±240 ±219 ±231 0.0053
EPC 9137.3 8435.2 6276.0±1 p<
±1858 ±2831 433 0.0001
(a) (b) (c) Perimeter 295.16 464.42 596.52 p<
Fig. 8 Object Perimeter (a) normal (b) cataract (c) Post cataract subjects. ±272 ±214 ±371 0.0001

TABLE II
F. Support Vector Machine (SVM) Classifier NUMBER OF TRAINING, TESTING DATA (FEATURES) AND
SVMs are new type of pattern classifier based on a PERCENTAGE OF CORRECT CLASSIFICATION.
novel statistical learning technique that has been recently
proposed by Vapnik et al [14]. Unlike traditional methods Classifier TN TP FP FN Sensitivity Specificity
(%) (%)
(eg. Neural Networks), which minimize the empirical
training error, SVMs aim at minimizing an upper bound of SVM 15 31 01 02 94 93.75
the generalization error through maximizing the margin
between the separating hyperplanes and the data [15].
SVMs are known to generalize well even in high
dimensional spaces under small training samples conditions.
They have been proven to be superior as compared to the
traditional empirical risk minimization principle employed
in most of neural networks. SVMs have been successfully

ISBN: 978-988-19252-3-7 WCECS 2013

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the World Congress on Engineering and Computer Science 2013 Vol I
WCECS 2013, 23-25 October, 2013, San Francisco, USA

TABLE III
RESULTS OF SENSITIVITY, SPECIFICITY FOR COMPLETE EYE ACKNOWLEDGMENT
CLASSES SVM CLASSIFIERS
The authors wish to thank Head of Ophthalmology
Type of No. of No. of Test SVM Classifier department, Kasturba Medical College, Manipal, INDIA, for
Image Training Samples providing images and clinical support required for this work.
Samples No. of % Result
samples REFERENCES
correctly
classified [1] Snell RS, Lemp MA, “Clinical Anatomy of the Eye”, Second Edition,
Normal 40 16 15 93.75 Blackwell Science, Inc.,
[2] Finlay RD, Payne PAG, “The Eye in General Practice”, Butterworth-
Cataract 63 26 26 100 Heinemann, 1988.
[3] Vander JF, Gault JA, “Ophthalmology Secrets”, Second Edition,
Post- 22 07 05 71.42 Hanley & Belfus, Inc., USA. 2002.
Cataract [4] Remington LA, “Clinical Anatomy of the Visual System”,
Butterworth-Heinemann, USA, 1998
[5] Cheng H-M, Yeh LI, Barnett P, Miglior S, Eagon JC, Gonzalez G,
Brady TJ, “Proton magnetic resonance imaging of the ocular lens”,
IV. DISCUSSION Exp Eye Res., 45(6),1987,875-882.
[6] Wong EK, Cho ZH, Gardner BP, Ahn CB, Kim I, Jo JM, Juh SC, and
In our work, cataract eye can be identified with 100% Anderson JA: In vivo magnetic resonance imaging studies of
accuracy. With the inclusion of more cataract images at the galactose cataracts. ARVO Abstracts. Invest Ophthalmol Vis Sci
early stages, the proposed system can be made more robust 28(Suppl), 80, 1987.
to identify cataract at the early stage too. The post-cataract [7] Nusz KJ, Congdon NG, Ho T, Gramatikov BI, Friedman DS, Guyton
DL, Hunter DG, “Rapid, objective detection of cataract-induced blur
eye images have been used to test the efficacy of cataract using a bull's eye photodetector”, J Cataract Refract Surg., 31(4),
operations. The accuracy of the system can further be 2005, 763-770.
improved by increasing the size and quality of the training [8] Monestam E, Wachmeister L, “Impact of cataract surgery on the
set. The classification results can be enhanced by extracting visual ability of the very old”, Am J Ophthalmol., 137(1), 2004,145-
155.
additional features from the optical images. Environmental
[9] Hejcmanova D, Langrova H, Bytton L, Hejcmanova M, “Changes of
conditions such as the reflection of light influences the visual function and visual ability in daily life following cataract
quality of the optical images and hence, the efficiency of surgery”, Acta Medica (Hradec Kralove), 46(4), 2003, 189-194.
classification. The software for feature extraction and the [10] Ferraro JG, Pollard T, Muller A, Lamoureux EL, Taylor HR,
program for classification of eye images are written in “Detecting cataract causing visual impairment using a nonmydriatic
fundus camera”, Am J Ophthalmol., 2005 139(4):725-726.
MATLAB 7.0.4. [11] Mashiach R , Vardimon D, Kaplan B, Shalev J, Meizner I, “Early
sonographic detection of recurrent fetal eye anomalies”, Ultrasound in
Obstetrics and Gynecology, 24( 6) , 2004, 640 – 643.
V. CONCLUSION [12] Gonzalez RC, Wintz P, “Digital Image Processing”, 2nd edn. Addison-
Wesley. Reading, MA, 1993.
In this paper, we have discussed the performances of [13] Gonzalez RC, Wintz P, “Digital Image Processing”, 2nd edn. Addison-
SVM classifier as diagnostic tools to aid physicians in the Wesley. Reading, MA, 1987.
[14] Vapnik V, “Statistical Learning Theory”, Springer, N.Y., 1997.
detection of cataract even at the early stages. These [15] Vapnik V, “The Nature of Statistical Learning Theory”, Springer,
classifiers are also suitable for diagnosing the effectiveness N.Y., 1995 ISBN 0-387-94559-8.
of cataract operations using the post-catact images. [16] Ye M, Wang X, “Modeling of nonlinear dynamic system using nu-
However, these tools generally do not yield results with support vector machines”, Intelligent Control and Automation, 2004.
WCICA 2004. Fifth World Congress, 1(1), 2004, 272 – 275.
100% accuracy. The accuracy of these tools depend on [17] MedCalc is a complete statistical program for Windows designed to
several factors such as the size and quality of the training closely match the requirements of biomedical researchers.
set, the rigor of the training imparted, and parameters http://www.medcalc.be
chosen to represent the input. From the results obtained as
listed in Table II and III, it is evident that the SVM classifier
is effective to the tune of about 88% accuracy.

ISBN: 978-988-19252-3-7 WCECS 2013

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)