80

IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 1, JANUARY 2012

Wavelet-Based Energy Features for Glaucomatous Image Classication

Sumeet Dua, Senior Member, IEEE, U. Rajendra Acharya, Pradeep Chowriappa, Member, IEEE, and S. Vinitha Sree
AbstractTexture features within images are actively pursued for accurate and efcient glaucoma classication. Energy distribution over wavelet subbands is applied to nd these important texture features. In this paper, we investigate the discriminatory potential of wavelet features obtained from the daubechies (db3), symlets (sym3), and biorthogonal (bio3.3, bio3.5, and bio3.7) wavelet lters. We propose a novel technique to extract energy signatures obtained using 2-D discrete wavelet transform, and subject these signatures to different feature ranking and feature selection strategies. We have gauged the effectiveness of the resultant ranked and selected subsets of features using a support vector machine, sequential minimal optimization, random forest, and nave Bayes classication strategies. We observed an accuracy of around 93% using tenfold cross validations to demonstrate the effectiveness of these methods. Index TermsBiomedical optical imaging, data mining, feature extraction, glaucoma, image texture, wavelet transforms.

I. INTRODUCTION LAUCOMA is the second leading cause of peripheral blindness worldwide and results in the neurodegeneration of the optic nerve. As the revitalization of the degenerated optic nerve bers is not viable medically, glaucoma often goes undetected in its patients until later stages. The prevalent model estimates that approximately 11.1 million patients worldwide will suffer from glaucoma induced bilateral blindness in 2020 [1]. Furthermore, in countries, like India, it is estimated that approximately 11.2 million people over the age of 40 suffer from glaucoma [2]. It is believed that these numbers can be curtailed with effective detection and treatment options. In light of the diagnostic challenge at hand, recent advances in biomedical imaging offer effective quantitative imaging alternatives for the detection and management of glaucoma. Several imaging modalities and their enhancements, including optical coherence tomography [3] and multifocal electroretinograph (mfERG) [4], are prominent techniques employed to

Manuscript received April 12, 2011; revised July 19, 2011; accepted November 6, 2011. Date of publication November 18, 2011; date of current version February 3, 2012. S. Dua and P. Chowriappa are with the Computer Science Program, Louisiana Tech University, Ruston, LA 71272 USA (e-mail: sdua@coes.latech.edu; pradeep@latech.edu). U. Rajendra Acharya is with the Department of Electronics and Communications Engineering, Ngee Ann Polytechnic, Singapore 599489 (e-mail: aru@np.edu.sg). V. Sree S is with the School of Mechanical & Aerospace Engineering, Nanyang Technological University, Singapore 639798 (e-mail: vinithasree@ ntu.edu.sg). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TITB.2011.2176540

quantitatively analyze structural and functional abnormalities in the eye both to observe variability and to quantify the progression of the disease objectively [5]. Automated clinical decision support systems (CDSSs) in ophthalmology, such as CASNET/glaucoma [6], [7], are designed to create effective decision support systems for the identication of disease pathology in human eyes. These CDSSs have used glaucoma as a predominant case study for decades. Such CDSSs are based on retinal image analysis techniques that are used to extract structural, contextual, or textural features from retinal images to effectively distinguish between normal and diseased samples. Retinal image analysis techniques rely on computational techniques to make qualitative assessments of the eye more reproducible and objective. The goal of using such methods is to reduce the variability that may arise between different clinicians tracking the progression of structural characteristics in the eye. In CDSS, features extracted from the images are categorized as either structural features or texture features. Commonly categorized structural features include disk area, disk diameter, rim area, cup area, cup diameter, cup-to-disk ratio, and topological features extracted from the image [8]. Proper orthogonal decomposition (POD) is an example of a technique that uses structural features to identify glaucomatous progression [9]. In POD, pixel-level information is used to gauge signicant changes across samples that are location or region specic. The measurement of texture features, on the other hand, is roughly dened as the spatial variation of pixel intensity (gray-scale values) across the image. Textural features are, thus, not bound to specic locations on the image. Several feature extraction techniques, including spectral techniques, are available to mine texture features. The use of texture features and higher order spectra (HOS) features were proposed by [10] for glaucomatous image classication. The readers may refer to [10] and [11] for an introductory summary of the various supervised classication techniques used to detect glaucoma. Although the texture-based techniques have been proven successful, it is still a challenge to generate features that retrieve generalized structural and textural features from retinal images. Texture features using wavelet transforms (WTs) in image processing are often employed [12] to overcome the generalization of features. In WT, the image is represented in terms of the frequency of content of local regions over a range of scales. This representation provides a framework for the analysis of image features, which are independent in size and can often be characterized by their frequency domain properties. The use of wavelet-Fourier analysis (WFA) for the characterization of neuroanatomic disruption in glaucoma was proposed

1089-7771/$26.00 2011 IEEE

DUA et al.: WAVELET-BASED ENERGY FEATURES FOR GLAUCOMATOUS IMAGE CLASSIFICATION

81

by [12] and has achieved considerable success. WFA is used as a mathematical model to analyze and parameterize the temporal, superior nasal, inferior, and temporal (TSNIT) shapes. In this approach, discrete wavelet transform (DWT) using a fourth-order symlets wavelet is used to extract features and analyze discontinuities and abrupt changes contained in signals. DWT is a multiscale analysis method, in which analysis can be performed on various scales. Each level of the transformation provides an analysis of the source image at a different resolution, resulting in its independent approximation and detailed coefcients. In the WFA, the fast Fourier transform (FFT) is applied to the detailed coefcients. The resultant Fourier amplitudes are combined with the normalized approximation coefcients of the DWT to create a set of features. Though providing valuable insight into the analysis of images using mathematical generalization techniques with DWT and FFT, it is yet uncertain which DWT features need to be extracted to represent an image that is as discriminative as possible in the transform domain. In image processing, it is also common practice to use the wavelet energy of each wavelet subband by implementing wavelet packet-based texture classication [13] to gauge the discriminatory potential of the texture features obtained from the image [14]. The goal of this study is, thus, to automatically classify normal eye images and diseased glaucoma eye images based on the distribution of average texture features obtained from three prominent wavelet families. Hence, our objective is to evaluate and select prominent features for enhanced specicity and sensitivity of glaucomatous image classication. In this paper, we quantitatively examine the effectiveness of different wavelet lters on a set of curated glaucomatous images by employing the standard 2-D-DWT. We propose to use three well-known wavelet lters, the daubechies (db3), the symlets (sym3), and the biorthogonal (bio3.3, bio3.5, and bio3.7) lters [15]. We calculate the averages of the detailed horizontal and vertical coefcients and wavelet energy signature from the detailed vertical coefcients. We subject the extracted features to a myriad of feature ranking and feature selection schemes to determine the best combination of features to maximize interclass similarity and aid in the convergence of classiers, such as the support vector machine (SVM), sequential minimal optimization (SMO), random forest, and nave Bayes techniques. The paper is organized in the following way. Section II contains an explanation of the dataset used in this paper. Section III contains a detailed description of the methodology adopted for this research, including the feature preprocessing, feature ranking, and feature selection schemes that we chose. Section IV contains a description of the classiers and classier parameters employed for our experiments. Section V includes an interpretation of results obtained using the proposed method. Finally, the paper concludes with Section VI. II. DATASET The retinal images used for this study were collected from the Kasturba Medical College, Manipal, India (http://www. manipal.edu). The doctors in the ophthalmology department of the hospital manually curated the images based on the quality

Fig. 1. Typical fundus images (a) normal (b) glaucoma. In glaucoma, the pressure within the eyes vitreous chamber rises and compromises the blood vessels of the optic nerve head, leading to eventual permanent loss of axons of the vital ganglion cells.

and usability of samples. The ethics committee, consisting of senior doctors, approved the use of the images for this research. All the images were taken with a resolution of 560 720 pixels and stored in lossless JPEG format [10]. The dataset contains 60 fundus images: 30 normal and 30 open angle glaucomatous images from 20 to 70 year-old subjects. The fundus camera, a microscope, and a light source were used to acquire the retinal images to diagnose diseases. Fig. 1(a) and (b) presents typical normal and glaucoma fundus images, respectively. III. METHODOLOGY The images in the dataset were subjected to standard histogram equalization [16]. The objective of applying histogram equalization was twofold: to assign the intensity values of pixels in the input image, such that the output image contained a uniform distribution of intensities, and to increase the dynamic range of the histogram of an image. The following detailed procedure was then employed as the feature extraction procedure on all the images before proceeding to the feature ranking and feature selection schemes. A. Discrete Wavelet Transform-Based Features The DWT captures both the spatial and frequency informations of a signal. DWT analyzes the image by decomposing it into a coarse approximation via low-pass ltering and into detail information via high-pass ltering. Such decomposition is performed recursively on low-pass approximation coefcients obtained at each level, until the necessary iterations are reached. Let each image be represented as a p q gray-scale matrix I[i,j], where each element of the matrix represents the grayscale intensity of one pixel of the image. Each nonborder pixel has eight adjacent neighboring pixel intensities. These eight neighbors can be used to traverse the matrix. The resultant 2-DDWT coefcients are the same irrespective of whether the matrix is traversed right-to-left or left-to-right. Hence, it is sufcient that we consider four decomposition directions corresponding to 0 (horizontal, Dh), 45 (diagonal, Dd), 90 (vertical, Dv), and 135 (diagonal, Dd) orientations. The decomposition structure for one level is illustrated in Fig. 2. In this gure, I is the image, g[n] and h[n] are the low-pass and high-pass lters, respectively, and A is the approximation coefcient. In this study, the results from level 1 are found to yield signicant features. As is evident from Fig. 2, the rst level of decomposition results in four coefcient matrices, namely, A1, Dh1, Dv1, and

82

IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 1, JANUARY 2012

TABLE I WAVELET FEATURES AND CORRESPONDING P -VALUES

Fig. 2. 2-D-DWT decomposition: 2ds1 indicates that rows are down sampled by two and columns by one. 1ds2 indicates that rows are down sampled by one and columns by two. The operator indicates convolution operation.

Dd1. Since the number of elements in these matrices is high, and since we only need a single number as a representative feature, we employed averaging methods to determine such singlevalued features. The denitions of the three features that were determined using the DWT coefcients are in order. Equations (1) and (2) determine the averages of the corresponding intensity values, whereas (3) is an averaging of the energy of the intensity values. Average Dh1 = 1 pq 1 pq |Dh1(x, y)|
x={p} y ={q }

(1) D. Feature Ranking Feature ranking is a preprocessing step that precedes classication. In this paper, we focus on using lter-based approaches to rank the features based on their discriminatory potential across samples. Since our objective is to estimate the effectiveness of the wavelet features, our analysis consists of four widely used feature ranking schemes. These include chi squared ([2 ]) [18], gain ratio [19], information gain [20] feature evaluation techniques, and relief feature ranking schemes [21] that are based on an extensive ranking algorithm. Each of these algorithms is detailed as follows. 1) Chi-Squared (2 ) Feature Evaluation: In this method, the value of a feature is estimated by computing the value of its 2 statistic. The compuation of the 2 statistic is tied to the distribution of values that the feature possesses with respect to the class label [18]. This feature evaluation technique is divided into two phases. In the rst phase, each feature is sorted according to a signicance level (sigLevel). Set at a sigLevel of 0.5, the features are discretized into random intervals. The 2 value is then computed for every pair of adjacent intervals of the feature. Next, interval pairs of lowest 2 value are merged in a process that terminates when the 2 value exceeds the previous set sigLevel. Phase two of the feature evaluation is considered to be a ne tuning of the process performed in phase one. Once the merging of feature intervals is carried out independently, a consistency check is performed for the second phase. Any inconsistency in the merging of feature i does not pass the previously determined sigLevel(i) for that feature, may not be considered a potentially signicant feature, and is discouraged for future merging.

Average Dv1 =

|Dv1(x, y)|
x={p} y ={q }

(2)

Energy =

1 p2 q 2

(Dv1(x, y))2
x={p} y ={q }

(3)

B. Preprocessing of Features As shown in Table I, 14 features can be found among the Normal and Glaucomatous image samples. Their corresponding distribution across these samples is also shown in the table. It should be noted that the features that exhibited p values <0.0001 were chosen for analysis. C. Normalization of Features Each of the 14 features is subject to z-score normalization [17]. In the process of z-score normalization, a sample (vector) consisting of 14 features is converted to zero mean and unit variance. The mean and standard deviation of the input vector are computed as follows: ynew = yold mean std (4)

where yold is the original value, ynew is the new value, and the mean and std are the mean and standard deviation of the original data range, respectively. Fig. 3(a) and (b) shows the z-scored normalized distribution for each of the features across the 30 Glaucoma and 30 normal samples used in this study.

DUA et al.: WAVELET-BASED ENERGY FEATURES FOR GLAUCOMATOUS IMAGE CLASSIFICATION

83

Fig. 3.

Distribution of the 14 normalized wavelet features across (a) 30 glaucoma samples. (b) 30 normal samples in the dataset.

In this way, the features are ranked according to the level of signicance. 2) Gain Ratio and Information Gain Feature Evaluation: In this paper, we have utilized both information gain and gain ratiobased methods to rank features [19], [20]. For these techniques, the expected information needed to classify a sample in a dataset S with m classes is given by
m

I(S) =
i=1

pi log2 (pi )

(5)

where pi is the probability that an arbitrary sample belongs to class Ci and is estimated by Si /S. The entropy or expected information of a feature A having v distinct values is given as
m

Relief uses two measures near hit and near miss to describe the proximity of an instance to a subset of instances that belong to a class. An instance is a near hit of X if it belongs to a close neighborhood of X and to the same class as X. Similarly, an instance is considered a near miss if it belongs to the proximity of X but also belongs to a different class than X. This algorithm chooses a triplet of samples that include the <instance X, its Near Hit, and its Near Miss>, where the near hit and near miss are chosen using the Euclidean distance. Once the near hit and near miss are determined, a feature weight vector W is updated using the following: Wi = Wi diff(xi , nearhit i )2 + diff(xi , nearm iss i )2 . (10)

E(A) =
i=1

I(S)

(s1i + s2i + sm i ) . s

(6)

A relevance vector R is determined using every sample triplet. This relevance vector R is derived from the weight vector W , and is used to depict the relevance of each feature using the following: R= 1 m W. (11)

The information that would be gained by spliting the dataset on attribute A is gathered using the following relation: Info Gain(A) = I(S) E(A). (7)

The normalization of information gain using a constant called SplitInfo for each feature is obtained as follows:
v

The features in this paper are ranked based on increasing order of relevance R. E. Feature Selection To select a subset of relevant features, we subject the given set of features to the consistency subset evaluation (CSE) strategy. CSE nds the combinations of features that have values that divide the data into subsets containing a strong single class majority, i.e., high-class consistency. This consistency as a measure was rst presented by [22] as follows: |Di | |Mi | (12) N where s is a feature subset, J is the number of distinct combinations of feature values for s, |Di | is the number of occurrences of the ith feature value combination, |Mi | is the cardinality of the majority class for the ith feature value combination, and N is the number of instances in the dataset. To use the CSE, the dataset is discretized with numeric attributes using forward selection search, which produces a list of ranked attributes. The rank of the attribute is determined Consistencys = 1
J i=0

SplitInfoA (S) =
i=1

|Si | |S|

log2

|Si | |S|

(8)

The aforementioned SplitInfo for a feature represents the information generated by splitting training set S into v partitions corresponding to v outcomes of a test on feature A Thus, Gain Ratio is dened as Gain Ratio(A) = InfoGain(A) . SplitInfoA (S) (9)

The feature with the highest gain ratio is selected as the feature based on which the training set is split. 3) Relief Feature Ranking: The relief algorithm was rst proposed by [21] as a feature selection approach and is based on instance-based learning. The relief algorithm uses a relevancy parameter , which acts as a threshold that ranges between (0 < < 1) and is used to gauge the statistical relevancy of a feature to the target concept, or the class to which a sample belongs.

84

IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 1, JANUARY 2012

TABLE II CLASSIFIERS USED TO VALIDATE THE FEATURES

V. EXPERIMENTAL RESULTS The following section provides a detailed description of the results obtained from our feature ranking and feature selection experiments. A. Feature Ranking and Feature Selection Table III provides a snapshot of the results obtained from both the feature ranking and feature selection schemes described in the methodology section. The ranking algorithms include 2 , gain ratio, info gain, and the relief algorithm. Table III also shows the ranks of the features selected using four consistency subset evaluation strategies, best rst, random search, genetic search, and greedy search. The cells of the table highlighted with the symbol depict the selected features, and the number of each highlighted cell depicts the ranking of each of the features obtained. B. Classication Once the features are subject to both categories of feature ranking and/or feature selection, we perform both the tenfold cross validation and the 60:40 slip tests. Both tests are carried out on the entire 60-sample dataset. In the tenfold cross validation method, the dataset is split into ten parts. During the rst iteration, nine parts are used for training and the remaining part is used for testing. This procedure is carried out ten times using a different part for testing in each iteration. The results obtained from the ten iterations are then averaged to obtain an overall accuracy. The 60:40 split test provides the accuracy of the classication obtained when 60% of the total number of samples from the dataset are chosen randomly. The classier is tested against the remaining 40% of the samples that constitute the test set. Table IV provides the results obtained using the tenfold cross validation. Table V contains the results obtained using the 60:40 split test. From Table IV, it is evident that the highest accuracy obtained is 93.33% using the tenfold cross-validation method. In the case of 60:40 split up, all classiers except the random forest classier presented the highest accuracy of 95.83%. It can be seen from Tables IV and V that the use of feature ranking and CSE techniques had very little impact on the accuracy obtained using LibSVM, SMO, nave Bayes with the exception of the random forest classier. However, it is possible to obtain the highest accuracy using fewer features in CSE. For example, Table IV shows that the SMO_(2) classier gives 93.33% accuracy using just two features selected by the best rst and greedy stepwise strategies in the CSE method. These features are db3-Dh1_Average_l1_Norm and rbio3.3-cD_Energy (see Table I). Therefore, these features can be considered highly discriminatory. The consistency of results indicates that these features perform consistently independent to the classier. We can conclude that the CSE feature selection method does help in obtaining the highest accuracy using fewer features, thereby simplifying the implementation of the technique. For further validation, we conducted a sensitivity and specicity analysis of 42 training samples belonging to classes

according to its overall contribution to the consistency of the attribute set. The various feature subset selection schemes followed in this paper are part of the WEKA [23] suite and include the best rst, random search, genetic search, and greedy stepwise approaches. Best First: The best rst search strategy implementation is based on the beam search algorithm [24]. The rudiments of this algorithm are based on the standard hill climbing approach with backtracking to determine the best t of a subset of features. Random Search: In the random search approach, an exhaustive list of random feature combinations are generated and tested. The subset of features that generates the best accuracy is chosen as the subset of features that best represents the input set. Genetic Search: The genetic search technique was proposed in [25], and utilizes neural network feature ranking. This algorithm entails several iterations of the evaluation of a feature subset and includes training a neural network and computing its cost and accuracy. Greedy Stepwise: Greedy stepwise subset evaluation is performed using the greedy forward or backward search through the feature space. IV. CLASSIFIER SETTINGS We performed the validation of the ranked features and feature subsets using the standard C-SVC implementation of SVM, SMO, random forest, and nave Bayes (as shown in Table II) [26][28]. The SVM employs the radial basis function with a predetermined value of gamma set at 0.28. John C. Platts SMO algorithm is used to train the SVM [27]. The implementation of SMO is performed using both the polynomial kernel with the exponential set to 2.5, and the Pearson VII function-based universal kernel, each having parameters omega and sigma, which are both set at 0.1. The number of trees in the random forest algorithm is set at 10. The nave Bayes classier is set to use a kernal function to estimate the distribution of the data, rather than assuming that the data follow a normal distribution. The classier settings are determined based on repeated trials on the training set, until a classication accuracy of 100% is obtained on the training set.

DUA et al.: WAVELET-BASED ENERGY FEATURES FOR GLAUCOMATOUS IMAGE CLASSIFICATION

85

TABLE III FEATURE RANKING

TABLE IV CLASSIFICATION ACCURACIES OBTAINED USING TENFOLD CROSS VALIDATION

TABLE V CLASSIFICATION ACCURACIES OBTAINED USING THE 60:40 SPLIT TEST

TABLE VI SENSITIVITY AND SPECIFICITY ANALYSIS USING INDEPENDENT TEST SET: ACC: ACCURACY; SN: SENSITIVITY; SP: SPECIFICITY

TABLE VII OVERALL ACCURACIES OF INDIVIDUAL WAVELET FEATURES USING TENFOLD CROSS VALIDATION

glaucoma and normal, each consisting of 21 samples. In this study, the features were not subject to any feature ranking or feature selection technique. The test set consisted of 18 samples. Nine samples belonged to the class glaucoma, and the other nine

samples belonged to the class normal. Table VI shows the results obtained for each of the ve classiers. Nayak et al. [30] used structural features such as cup-to-disc (c/d) ratio, the ratio of the distance between the optic disc center

86

IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 1, JANUARY 2012

TABLE VIII COMPARISON OF ACCURACIES USING FIVEFOLD CROSS VALIDATION

and optic nerve head 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. Their technique detected glaucoma with a sensitivity and specicity of 100% and 80%, respectively. This result implies that although the system can detect all subjects with glaucoma accurately, it can detect only 80% of the normal subjects as normal. In our proposed technique using texture features based on wavelets, we were obtained a higher accuracy of over 93.33%. Finally, we compare our results with those obtained using HOS and texture features [10]. The training samples and the test sample identiers are mapped with those used in [10] to enable a direct comparision of accuracy. To establish a baseline analysis of the features used in this study, we gauge the reliability of individual wavelet features by performing standard sensitivity and specicity analysis on the independent training and test sets used in [10]. The classiers used for this analysis are described in Table II. We carried out a tenfold cross validation of independent wavelet features on the entire set, consisting of 60 samples. The results of this experimentation are shown in Table VII. It can be seen that the SMO_(2) classier performs consistently well using any wavelet-based feature. Table VIII contains the results obtained using vefold cross validation. From Table VIII, it can be observed that our features outperform the LibSVM, SMO, and Nave Bayes classiers. However, the results obtained using the random forest classier is almost at par with the results obtained in [10]. When the optic nerve is damaged by glaucoma, most of the individual bers in the nerve are lost, and the optic nerve becomes excavated. As glaucoma progresses, more optic nerve tissue is lost and the optic cup grows [31]. These changes make the fundus images obtained from glaucoma patients different from those obtained from normal subjects. We believe that the features selected and evaluated in this study adequately capture these subtle differences in the fundus images and, thus, help the classiers to detect normal and glaucoma cases with high accuracy. VI. CONCLUSION This study illustrates the dependence between features extracted using three wavelet lters that have been subjected to various feature ranking and feature selection methods. The ranked subsets of selected features have been fed to a set of classication algorithms to gauge the effectiveness of these features. From the accuracies obtained and contrasted, we can conclude that the energy obtained from the detailed coefcients can be used to distin-

guish between normal and glaucomatous images with very high accuracy. As observed the db3-Dh1_Average_l1_Norm and the rbio3.3-cD_Energy features are highly discriminatory. Furthermore, from Table IV that both LibSVM_(1) and SMO_(2) present the highest accuracy of 93.33%. REFERENCES
[1] R. Varma et al., Disease progression and the need for neuroprotection in glaucoma management, Am. J. Manage Care, vol. 14, pp. S15S19, 2008. [2] R. George, R. S. Ve, and L. Vijaya, Glaucoma in India: Estimated burden of disease, J. Glaucoma, vol. 19, pp. 391397, Aug. 2010. [3] K. R. Sung et al., Imaging of the retinal nerve ber layer with spectral domain optical coherence tomography for glaucoma diagnosis, Br. J. Ophthalmol., 2010. [4] J. M. Miquel-Jimenez et al., Glaucoma detection by wavelet-based analysis of the global ash multifocal electroretinogram, Med. Eng. Phys., vol. 32, pp. 617622, 2010. [5] B. Brown, Structural and functional imaging of the retina: New ways to diagnose and assess retinal disease, Clin. Exp. Optometry, vol. 91, pp. 504514, 2008. [6] S. Weiss, C. A. Kulikowski, and A. Sar, Glaucoma consultation by computer, Comp. Biol. Med., vol. 8, pp. 2440, 1978. [7] S. Weiss et al., A model-based method for computer-aided medical decision-making, Artif. Intell., vol. 11, pp. 145172, 1978. [8] R. O. Duncan et al., Retinotopic organization of primary visual cortex in glaucoma: A method for comparing cortical function with damage to the optic disk, Invest. Ophthalmol. Vis. Sci., vol. 48, pp. 733744, Feb. 2007. [9] M. Balasubramanian et al., Clinical evaluation of the proper orthagonal decomposition framework for detecting glaucomatous changes in human subjects, Invest. Ophthalmol. Vis. Sci., vol. 51, pp. 264271, 2010. [10] U. R. Acharya, S. Dua, X. Du, V. S. Sree, and C. K. Chua, Automated diagnosis of glaucoma using texture and higher order spectra features, IEEE Trans. Inf. Technol. Biomed., vol. 15, no. 3, pp. 449455, May 2011. [11] S. Dua, U. R. Acharya, and E. Y. K. Ng, Computational Analysis of the Human Eye With Applications. World Scientic Press, 2011. [12] E. A. Essock, Y. Zheng, and P. Gunvant, Analysis of GDx-VCC polarimetry data by wavelet-Fourier analysis across glaucoma stages, Invest. Ophthalmol. Vis. Sci., vol. 46, pp. 28382847, Aug. 2005. [13] K. Huang and S. Aviyente, Wavelet feature selection for image classication, IEEE Trans. Image Process., vol. 17, no. 9, pp. 17091720, Sep. 2008. [14] A. Arivazhagan and L. Ganesan, Texture classication using wavelet transform, Pattern Recog. Lett., vol. 24, pp. 15131521, 2003. [15] I. Daubechies, Ten Lectures on Wavelets. Philadelphia, PA: Society for Industrial and Applied Mathematics, 1992. [16] R. C. Gonzalez and R. E. Woods, Digital Image Processing. NJ: Prentice Hall, 2001. [17] M. H. Dunham, Data Mining Introductory and Advance Topics. NJ: Prentice Hall, 2002. [18] H. Liu and R. Setiono, Chi2: Feature selection and discretization of numeric attributes, in Proc. IEEE 7th Int. Conf. Tools With Artif. Intell., 1995, pp. 338391. [19] J. R. Quinlan, Induction of desion trees, Mach. Learning, vol. 1, pp. 81 106, 1986. [20] J. R. Quinlan, C4.5 Programs for Machine Learning. San Mateo: Morgan Kaufmann, 1993.

DUA et al.: WAVELET-BASED ENERGY FEATURES FOR GLAUCOMATOUS IMAGE CLASSIFICATION

87

[21] K. Kira and L. A. Rendell, A practical approach to feature selection, in Proc. 9th Int. Workshop Mach. Learning, San Francisco, CA, 1992, pp. 249256. [22] H. Liu and R. Setinono, A probabilistic approach to feature selectionA lter solution, in Proc. 13th Int. Conf. Mach. Learning, Bari, Italy, 1996, pp. 319327. [23] M. Hall et al., The WEKA data mining software: An update, SIGKDD Explorations, vol. 11, pp. 1018, 2009. [24] D. Furcy and S. Koenig, Limited discrepancy beam search, in Proc. Int. Joint Conf. Artif. Intell., 2005, pp. 125131. [25] D. E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning. Boston, MA: Addison-Wesley, 1989. [26] C.-C. Chang and C.-J. Lin. (2001). LIBSVM: A library for support vector machines. [Online]. Available: [Accessed: March 10, 2011.] http://www.csie.ntu.edu.tw/cjlin/libsvm [27] S. S. Keerthi et al., Improvements to Platts SMO algorithm for SVM classier design, Neural Comput., vol. 13, pp. 637649, 2001. [28] L. Breiman, Random forests, Mach. Learning, vol. 45, pp. 532, 2001. [29] G. H. John and P. Langley, Estimating continuous distributions in Bayesian classiers, in Proc. 11th Conf. Uncertainty Artif. Intell., San Mateo, 1995, pp. 338345. [30] 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. [31] Glaucoma guide. (2010). [Online]. Available: http://www.medrounds. org/glaucoma-guide/2006/02/section-1-b-meaning-of-cupping.html

U. Rajendra Acharya received the Ph.D. degree from the National Institute of Technology Karnataka, Surathkal, India, and the D.Engg. degree from Chiba University, Japan. He serves as visiting faculty at the Ngee Ann Polytechnic, Singapore, adjunct faculty at the Singapore Institute of Technology- University of Glasgow, Singapore and at the Manipal Institute of Technology, Manipal, India, and associate faculty at SIM University, Singapore. He has published more than 240 papers, in refereed international SCI-IF journals (140), international conference proceedings (41), textbook chapters (62), books (11 including in those in press) with h-index of 14 (Scopus). He has received funding of more than 1.5 million SGD, is on the editorial board of many journals, and has served as a guest editor for several journals. His major interests include biomedical signal processing, bioimaging, data mining, visualization, and biophysics for better healthcare design, delivery, and therapy.

Sumeet Dua (SM08) received the M.S. and Ph.D. degrees in computer science from Louisiana State University, Baton Rouge. He is currently the Upchurch endowed Associate Professor of computer science and the coordinator of IT research for the Computer Science Program at Louisiana Tech University, Ruston. He is also an adjunct faculty member at the School of Medicine, Louisiana State University Health Sciences Center at New Orleans. The U.S. National Institutes of Health NIH), the U.S. National Science Foundation (NSF), the Air Force Research Laboratory, and the Louisiana Board of Regents have funded his research in the past ve years. He has authored more than 50 research papers and is a co-author/editor of 3 books. He is a regular reviewer of more than 20 international journals, has served on more than 17 NIH study section and NSF/CISE review panels. His research interests include database mining and indexing, pattern recognition, dimensionality reduction, bioinformatics, and heterogeneous distributed data integration. Dr. Dua is a senior member of the ACM and the American Association for the Advancement of Science.

Pradeep Chowriappa (M09) received the B.Sc. degree in engineering from Osmania University, India, in 2000, the masters degree in computer applications from the University of Madras, India, in 2003, and the Ph.D. degree in computational analysis and modeling from Louisiana Tech University, in 2008. He is currently a Postdoctoral Research Associate in the Data Mining Research Laboratory at Louisiana Tech University, Ruston. His research interests include bioinformatics, data mining, feature discovery, and machine learning in high-dimensional domains.

S. Vinitha Sree received the M.S. degree in biomedical engineering and the Ph.D. degree from Nanyang Technological University, Singapore. She is a visiting scientist at Global Biomedical Technologies, Inc., CA. Her areas of interest include medical data mining, breast imaging systems, and healthcare systems and management. She has published over 30 peer-reviewed publications and is an associate editor for Journal of Medical Imaging and Health Informatics and a reviewer for the Journal of Mechanics in medicine and biology.