This may be the author’s version of a work that was submitted/acceptedfor publication in the following source:

Mookiah, Muthu RamaKrishnan, Acharya, Rajendra, Koh, Joel, Chan-dran, Vinod, Chua, Chua Kuang, Tan, Jen Hong, Min, Lim, Ng, Yin Kwee(E.Y.K.), Noronha, Kevin, Tong, Louis, & Laude, Augustinus(2014)Automated diagnosis of Age-related Macular Degeneration usinggreyscale features from digital fundus images.Computers in Biology and Medicine, 53, pp. 55-64.

This file was downloaded from: https://eprints.qut.edu.au/75738/

c© Consult author(s) regarding copyright matters

This work is covered by copyright. Unless the document is being made available under aCreative Commons Licence, you must assume that re-use is limited to personal use andthat permission from the copyright owner must be obtained for all other uses. If the docu-ment is available under a Creative Commons License (or other specified license) then referto the Licence for details of permitted re-use. It is a condition of access that users recog-nise and abide by the legal requirements associated with these rights. If you believe thatthis work infringes copyright please provide details by email to qut.copyright@qut.edu.au

License: Creative Commons: Attribution-Noncommercial-No DerivativeWorks 2.5

Notice: Please note that this document may not be the Version of Record(i.e. published version) of the work. Author manuscript versions (as Sub-mitted for peer review or as Accepted for publication after peer review) canbe identified by an absence of publisher branding and/or typeset appear-ance. If there is any doubt, please refer to the published source.

https://doi.org/10.1016/j.compbiomed.2014.07.015

Automated Diagnosis of Age-Related Macular Degeneration us-ing Grayscale Features from Digital Fundus Images

Muthu Rama Krishnan Mookiaha,∗, U. Rajendra Acharyaa,b, Joel E.W. Koha, VinodChandranc, Chua Kuang Chuaa, Jen Hong Tana, Choo Min Lima, E.Y.K. Ngd, Kevin

Noronhae, Louis Tongf,g,h,i, Augustinus Laudej

aDepartment of Electronics and Computer Engineering, Ngee Ann Polytechnic, Singapore 599489bDepartment of Biomedical Engineering, Faculty of Engineering, University of Malaya, Malaysia 50603cSchool of Electrical Engineering and Computer Science, Queensland University of Technology, Brisbane

QLD 4000, AustraliadSchool of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore 639798

eDepartment of E&C, MIT Manipal, India 576104fSingapore National Eye Center, Singapore 168751

gOcular Surface Research Group, Siganpore Eye Research Institute, Singapore 168751hDuke-NUS Graduate Medical School, Singapore 169857

iYong Loo Lin School of Medicine, National University of Singapore, Singapore 117597jNational Healthcare Group Eye Institute, Tan Tock Seng Hospital, Singapore 308433

Abstract

Age-related Macular Degeneration (AMD) is one of the major causes of vision loss and blind-ness in ageing population. Currently, there is no cure for AMD, however early detectionand subsequent treatment may prevent the severe vision loss or slow the progression of thedisease. AMD can be classified into two types: dry and wet AMD. The people with maculardegeneration are mostly affected by dry AMD. Early symptoms of AMD are formation ofdrusen and yellow pigmentation. These lesions are identified by manual inspection of fun-dus images by the ophthalmologists. It is a time consuming, tiresome process, and hencean automated diagnosis of AMD screening tool can aid clinicians in their diagnosis signifi-cantly. This study proposes an automated dry AMD detection system using various entropies(Shannon, Kapur, Renyi and Yager), Higher Order Spectra (HOS) bispectra features, FractalDimension (FD), and Gabor wavelet features extracted from gray scale fundus images. Thefeatures are ranked using t-test, Kullback-Lieber Divergence (KLD), Chernoff Bound andBhattacharyya Distance (CBBD), Receiver Operating Characteristics Curve (ROC)-basedand Wilcoxon ranking methods in order to select optimum features and classified into nor-mal and AMD classes using Naive Bayes (NB), k-Nearest Neighbour (k-NN), ProbabilisticNeural Network (PNN), Decision Tree (DT) and Support Vector Machine (SVM) classifiers.The performance of the proposed system is evaluated using private (Kasturba Medical Hos-pital, Manipal, India), Automated Retinal Image Analysis (ARIA) and STructured Analysisof the Retina (STARE) datasets. The proposed system yielded the highest average classi-fication accuracies of 90.19%, 95.07% and 95% with 42, 54 and 38 optimal ranked featuresusing SVM classifier for private, ARIA & STARE datasets respectively. This automated

1

AMD detection system can be used for mass fundus image screening and aid clinicians bymaking better use of their expertise on selected images that require further examination.

Keywords:Age-related macular degeneration, Entropy, Texture, Higher order spectra, Gabor wavelet,Computer aided diagnosis

1. Introduction

AMD is a chronic irreversible medical condition characterized by drusen or hyper orhypopigmentations [1]. It is caused due to cell damage in the macula region resulting incentral vision loss. AMD is a multi-factorial disease and has various risk factors viz., age,ocular risk factors namely drusen, retinal pigmentation, choroidal neovascularisation andsystemic risk factors namely family history, hypertension and smoking [2, 3, 4, 5]. AMD isone of the leading sight-threatening disease among people above 50 years [1, 6]. Between20-25 million people are affected globally and it may increase three times in the next 30-40years with the increase in ageing population [2, 7, 8]. World Health Organization reported8 million people have already affected with severe blindness due to AMD [2, 7, 8]. AMD ismainly classified into two types wet and dry [1, 9]. They are briefly explained below.

(i) Wet macular degeneration: affects the retina due to filling of fluid under theretina [1]. It leads to bleeding (See Figure 1b) and scarring causing loss of vision.It progresses rapidly and may respond to laser treatment in the early stage [1, 10].Approximately 10% of all people with macular degeneration have the ‘wet’type [1, 10,11].

(ii) Dry macular degeneration: caused by lack of functioning of visual cells [1]. Initialsymptoms of dry AMD is the presence of fatty deposits, called drusen (See Figure 1c),on the retina [12]. There is no treatment for ‘dry’type [13]. Most of the people withmacular degeneration are affected by ‘dry’type [1, 10].

∗Corresponding Author: Tel.: 65 64607889Email address: mkm2@np.edu.sg (Muthu Rama Krishnan Mookiah)

Preprint submitted to Elsevier July 1, 2014

Figure 1: Typical non-stereo fundus images (a) Normal; (b) wet AMD; (c) dry AMD.

The Wisconsin AMD grading system categorizes drusen based on size and visibility ofthe boundary, and is adopted in this study. Drusen can be classified as hard or soft. Fur-ther, soft drusen are classified as distinguishable or indistinguishable. Retinal pigmentarylesions are grouped into hypo and hyper pigmentation [10, 2, 14]. The presence of clinicalfeatures such as drusen, retinal pigmentation and Geographic Atrophy (GA), dry AMD arecategorised in to three stages: early, intermediate, and advance [6, 10, 2] which are brieflydescribed in Table 1.

Table 1: Stages of dry AMD [6, 10, 2].

Dry AMD stages Presence of clinical featuresEarly (See Figure 2a) -Many small (63µm in diameter) and few

medium (63µm to 124µm in diameter)sized drusens (i.e., distinguishable andindistinguishable soft drusen).

-Retinal pigmentation.Intermediate (See Figure2b)

-At least one large (124µm) and manymedium sized drusens.

-Lesions present away from the centre ofthe macula.

Advanced (See Figure2c)

-Drusen in the centre or periphery of themacula.

-Geographic atrophy (area with ≥175 µmdiameter).

-Neovascular AMD lesions.

3

Figure 2: Stages of dry AMD (a) Early; (b) Intermediate; (c) Advanced.

The drusen can be identified by the manual evaluation of retinal fundus images by trainedclinicians [6]. Several authors have developed automated drusen segmentation methods,which is the initial step in AMD classification. Automated drusen segmentation methodsare briefly described in Table 2.

Table 2: Automated drusen segmentation methods.

Authors Methods Highlights Performancemeasure

Sbeh etal. (2001)[15]

h-Maxima trans-formation

Segmentation ac-curacy dependson the h parame-ter

Not mentioned

Rapantzikos etal. (2003) [16]

Histogram-basedAdaptive LocalThresholding(HALT) operator

Able to segmentsmall and vaguedrusen

Sensitivity-98%

Niemeijer et al.(2007) [17]

Gaussian deriva-tive filtersk-NN classifier

May miss fewbright lesions

Sensitivity-77%Specificity-88%

Soliz et al.(2008) [18]

IndependentComponent Anal-ysis (ICA)

Drusen pheno-types

Accuracy-100%

Continued on next page

4

Table 2 – Continued from previous pageAuthors Methods Highlights Performance

measureBarriga et al.(2009) [19]

Amplitude Mod-ulation (AM)-Frequency Mod-ulation (FM)method

Characterizeslowly varyingintensity

Area Under re-ceiver operatorcharacteristicsCurve (AUC)-1

Freund et al.(2009) [20]

Mexican hatwaveletSupport VectorData Description(SVDD)

Scale parameterselected manually

Accuracy-100%

Liang et al.(2010) [21]

Intensity-basedsegmentation

Maximal region-based pixel inten-sity

Accuracy-100%

Burlina et al.(2011) [22]

Multi-resolutionlocally adaptivesegmentationmethod

Constant falsealarm rate detec-tor

Sensitivity-95%Specificity-96%

Santos-Villaloboset al.(2011) [23]

Statistical mod-elling

Post process-ing is requiredto remove falsepositives

AUC-1

Mora et al.(2011) [24]

Gradient-basedsegmentation

Drusen is mod-elled using Gaus-sian function

Kappa agree-ment of 0.60

Quellec et al.(2011) [25]

Optimal filterframework

Able to differen-tiate drusen andflecks

Not mentioned

Cheng et al.(2012) [26]

BiologicallyInspired Fea-tures (BIF)

Gabor functions Sensitivity-86.3%Specificity-91.9%

The important challenge in the above mentioned methods is the identification of drusenand also differentiating drusen from background noise and other lesions. Moreover, thevariation in the shape and size of drusen varies from image to image causing inaccuratedetection of drusen. Hence, traditional image segmentation techniques are not very effectivein isolating drusen within the retinal photographs [6].

In this paper we are proposing to classify normal and dry AMD classes without seg-mentation of drusen. The fundus images are preprocessed using adaptive histogram equal-ization. Further, various non-linear features (FD, HOS entropies, and Gabor wavelet) are

5

extracted from the preprocessed images and ranked using t-test, KLD, CBBD, ROC-basedandWilcoxon ranking techniques. Finally, the ranked features are fed to the set of supervisedclassifiers to identify the best performing classifier. The black diagram of the automatedAMD diagnosis system is shown in Figure 3.

Fundus Image Preprocessing Feature Extraction Feature Selection Classification

Figure 3: Overview block diagram of automated AMD diagnosis system.

Paper is organized as follows: Image acquisition, preprocessing, feature extraction, rank-ing and selection are presented in Section 2. Classifiers used for AMD classification arebriefly described in Section 3. The results of feature ranking method and various classifiersare explained in Section 4. The obtained results are discussed in Section 5 and a conclusionis provided in Section 6.

2. Materials and methods

2.1. Retinal fundus imagingPrivate dataset

The Normal and dry AMD fundus images were acquired using TOPCON non-mydriaticretinal camera (TRC-NW200) from Department of Ophthalmology, Kasturba Medical Col-lege, Manipal, India. The images were photographed and labelled by the group of ophthal-mologist. The hospital ethics committee has approved the image data for research. Thefive hundred and forty 24-bit Red Green Blue (RGB) color images (Normal-270 and dryAMD-270) were stored in lossless JPEG format with an image size of 480 × 364 pixels.The stored AMD images were reviewed by an experienced ophthalmologist and graded in toearly, intermediate and advance (See Figure 2) dry AMD images according to Age-RelatedEye Disease Study (AREDS) [27] classification.

Public datasetTwo publicly available datasets viz., ARIA (http://www.eyecharity.com/aria_online)

and STARE (http://www.ces.clemson.edu/~ahoover/stare) were also used to test theperformance of the proposed AMD detection system. The ARIA dataset consists of 101normal and 60 AMD images acquired using Carl Zeiss Meditec fundus camera with 50field of view and a resolution of 768×576 pixels. The STARE dataset consists of 36 normaland 47 AMD images acquired using TOPCON fundus camera with 35 field of view and aresolution of 700×605 pixels.

2.2. PreprocessingThe green band of the RGB color fundus images was subjected to contrast enhancement

using Contrast Limited Adaptive Histogram Equalisation (CLAHE). This technique divides6

the images into blocks and performs adaptive histogram equalization on the divided blocksby considering the pixels in the adjacent area. Hence, each pixel in the image is mapped toan new intensity value which is proportional to its rank in its context region [28], causinganatomical structures Optic Disc (OD), macula drusen, and Choroidal neovascularization(CNV) appear more visible.

2.3. Feature extractionThis section describes the different techniques used to extract features from each green

band AMD fundus image. In this work, we have extracted FD, Gabor wavelet, entropiesand HOS features from the preprocessed fundus images. They are briefly explained below.

2.3.1. Fractal dimension (D)Fractal Dimension is used to model the natural objects reflecting roughness and self-

similarity [29]. Let us consider A is a bounded set in n dimensional Euclidean space and itsfractal dimension (D) can be computed using the following equation

1 = NrrD

(1)

D = logNrlog( 1

r)

(2)

The differential box counting method [30] considers an image havingM×M size which isreduced to a grid size of s×s where 2 6 s 6 1

2M . Further, the scale factor r can be estimated

as r = sM. Each grid is considered as a set of s×s×s′ sized boxes, which are stacked one by

one, where s′ = [s× GM

] and satisfies the relation [Ms

] = [Gs′

], G is total gray levels. The graylevel difference of the image in k and lth box is computed as nr(i, j) = k− l+ 1, which is thecontribution of Nr in (i, j)th grid [30]. The Nr is computed by adding all grids contributionswhich is denoted as follows

Nr =∑

i,j = nr(i, j)(3)

Nr is calculated by changing the values of r. Further, linear least square estimation oflog(Nr) against log(1

r) [30] is used to compute D. In this work, sequential modified differential

box counting algorithm [31, 32] is implemented to compute the D of normal and AMD fundusimages. This algorithm considers the grid sizes as s = 2i, where i is an integer (here i = 2).Moreover, the grid size is doubled until max (R,C)<2, where R= rows and C = columns of

7

the image. The Nr is computed from the grid boxes by considering minimum and maximumintensity values using equation (3). Similarly, D is estimated using equation (2).

2.3.2. Gabor waveletThe Gabor wavelet based features provide a proven feature set for texture quantifica-

tion [33]. Fourier transform (G(m,n)) of 2D Gabor function (g(x, y)) is computed as follows

g(x, y) =(

12πσxσy

)exp

[−1

2

(x2

σ2x

+ y2

σ2y

)+ 2πjωx

](4)

Gl,k(m,n) = 12πσmσn

exp(−1

2

(m2

σ2m

+ n2

σ2n

)+ 2πjωm

)(5)

where ω is the frequency of sinusoid, σm = 12πσx

and σn = 12πσy

is the standard deviationof the Gaussian envelops [34]. Two dimensional Gabor wavelets are obtained using dilationand rotation of the mother wavelet G(m,n) and it is given below

Gl,k(m,n) = a−lG[a−l (m cos θ + n sin θ) , a−l (−m sin θ + n cos θ)

], a > 1

(6)

where a−l is the scale factor; l and k are the integers; θ = nπK

is the orientation and Kis the total number of orientations [34]. The σm and σn are filter parameters, which arecalculated as follows

a =(UhUl

) 1S−1 , σm = (a−1)Uh

(a+1)√

2ln2, σn = tan

(π2k

) [Uh − 2ln

(2σ2m

Uh

)] [2ln2− (2ln2)2σ2

m

U2h

]− 12

(7)

where Uh and Ul are the lower and upper frequency of the sinusoid, S is number ofscales [34].

Let us consider I(m,n) is a given fundus image. Its Gabor wavelet transform can be

8

computed using the following equation

gl,k(m,n) = I (m,n) ∗Gl,k (m,n) for l= 1, 2,...,S and k= 1, 2,...,K

(8)

where ∗ denotes the convolution operator. The mean (µlk) and the standard deviation(σlk) of the magnitude of the transform coefficients are used as feature vector [34] which areextracted using the following equation

µl,k(m,n) = 1M2

M∑m=1

N∑n=1

|gl,k(m,n)| and σl,k(m,n) =

(1M2

M∑m=1

N∑n=1

(|gl,k(m,n)| − µl,k)2

) 12

(9)

we use K = 6 orientations and S = 4 scales, provides total 48 features (Gabor-1 toGabor-48), which are denoted as

f = (µ1, σ1, ..., µ48, σ48)(10)

2.3.3. EntropyEntropy is mainly used to capture the uncertainty in the image pixels [35, 36, 37]. The

entropy measures such as Shannon, Kapur, Renyi and Yager are used to quantify the ran-domness in the intensity distribution of the normal and dry AMD fundus images. Consid-ering I(x, y) is either normal or AMD image having Ni(i = 0, 1, 2, 3, 4, ..., L − 1) intensityvalues. Hence, the normalized intensity histogram (Hi) can be computed for the given image(I(x, y)) having size of M ×N is given below [35, 36, 37]

Hi = NiM×N

(11)

where M and N are rows and columns of the given image respectively.

Shannon entropy is a measure of average information content [35, 36, 37] present in theimage and is computed as follows

9

Shannon’s entropy S = −∑L−1

i=0 Hi log2(Hi)

(12)

where L represents number of gray levels present in the given image.

Likewise Renyi, Kapur and Yager entropies are calculated. Renyi entropy is a generalizedentropy [38, 39] which is defined as

Renyi’s entropy: R = 11−α log

∑L−1i=0 H

αi , For α 6= 1, α > 0

(13)

where α is diversity index. Here α = 3.

The Kapur entropy is a further generalized version of Renyi entropy [38, 39] and it canbe defined as

Kapur’s entropy: K(α,β) = 11−α ln

L−1∑i=0

Hαi

L−1∑i=0

Hβi

, For α 6= 1, β > 0, α + β − 1 > 0

(14)

where α and β are diversity indices. Here α = 0.5 and β = 0.7.

Yager’s entropy measure is generally used to capture the visible information present inthe image [40] and can be computed as follows

Yager’s measure: Y = 1−∑L−1i=0 |2Hi−1||M×N |

(15)

where M and N are rows and columns of the given image respectively, Hi is the proba-bility distribution of normalized intensity histogram.

2.3.4. Radon Transform (RT)The Radon Transform (RT) provides many useful features which may not give any phys-

ical meaning as per human perception, however it has mathematical properties which candiscriminate the objects [41]. This technique is generally used to reconstruct the imagefrom scattering data in computed tomography. RT transforms the images pixels into lineparameters in the Radon domain for a given set of angles [41]. RT [42] can be computed

10

using the following equation

R(ρ, θ) =´ ´∞−∞ g(x, y)δ(ρ− x cos θ − y sin θ)dxdy

(16)

where θ is the angles at which the line parameters are computed, g(x, y) is given imageand δ(r) is the Dirac function [42, 43].

where δ(x cos θ + y sin θ) is the line integration function of g(x, y) [42, 43] defined using

ρ− x cos θ − y sin θ = 0(17)

In this work, we have used the step size θ = 10 to compute RT of all images. The HOSbispectrum is performed on these Radon projections.

2.3.5. HOS bispectrumHigher order spectra, or polyspectra, are Fourier spectral representations of third and

higher order cumulants of ergodic random processes. They have been extended to apply todeterministic signals and images. In this work, we the third order HOS or the bispectrumto extract features [44, 45, 46]. The bispectrum [47, 48, 49] of a deterministic signal x(t)can be estimated as

B(f1, f2) = E[X(f1)X(f2)X∗(f1 + f2)](18)

where X(f) is the discrete-time Fourier transform and estimated using Fast FourierTransform (FFT), f1 and f2 are the frequency components lies between 0 and 1 which arenormalized with Nyquist frequency. The bispectrum of a signal can be denoted with thetriangle 0 ≤ f2 ≤ f1 ≤ f1+ ≤ f2 ≤ 1, considering no bispectral aliasing. Features areextracted by integrating along the slope (a) passing through the bifrequency space [46, 50].

Different features can be extracted from the bispectrum. Here, bispectral phase entropy(Ph), entropy 1 (P1), entropy 2 (P2) and entropy 3 (P3) are extracted for each radon-transformed (θ− 0, 45, 90, 135) normal and AMD fundus image. The phase entropy can

11

be computed as follows

Ph =∑〈n〉 p(Ψn) log(p(Ψn))

(19)

p(Ψn) = 1L

∑〈Ω〉 l (Φ (B(f1f2)) ∈ Ψn)

(20)

Ψn = Φ| − π + 2πn/N ≤ Φ < −π + 2π(n+ 1)/N with n = 0, 1, . . . , N − 1

(21)

where B is the bispectrum of the signal, L is the number of points in Ω region, Φ is thephase angle, and l(.) is an indicator function and Ψn is the bin, as depicted by in Equation20. We chosen each bin as 45 (with N = 4) and histogram is computed. Further, Shannonentropy [39] is used to compute three bispectrum entropies which are computed as follows:

Normalized bispectral entropy (P1) can be defined as

P1 = −∑〈k〉 pk log(pk)

(22)

where pk = |B(f1,f2)|∑〈Ω〉 |B(f1,f2)| , pk is the probability density function and Ω is the bifrequency

space [46, 50].

Normalized bispectral squared entropy (P2) can be defined as

P2 = −∑〈i〉 pi log(pi)

(23)

where pi = |B(f1,f2)|2∑〈Ω〉 |B(f1,f2)|2 , pi is the probability density function and Ω is the bifrequency

space [46, 50].

12

Normalized bispectral cubed entropy (P3) can be defined as

P3 = −∑〈n〉 pn log(pn)

(24)

where pn = |B(f1,f2)|3∑〈Ω〉 |B(f1,f2)|3 , pn is the probability density function and Ω is the bifrequency

space [46, 50].

2.4. Feature selection and rankingFeature selection removes the redundant features and retain features which contributes

to the classifier performance. It avoids “curse of dimensionality” to enhance the classifierperformance [51]. In this work we have evaluated five feature ranking methods namely t-test,KLD, CBBD, ROC-based and Wilcoxon ranking methods. t-test compares the populationmeans of the two groups to identify the correlation among the features [52]. KLD rank thefeatures based on divergence measure between normal and dry AMD classes [53]. Chernoffbound provides exponential decay on tail distributions using independent variables. Henceit is used with Bhattacharyya distance for feature ranking [54]. ROC-based feature rankingselects the features using area between ROC and random classifier slope [54]. The Wilcoxonmethod is a non-parametric test, which does not assume normal distribution.

3. AMD classification

Efficient classification systems can help reduce the screening time and can be used pro-vided they do not trade off accuracy more than desirable. In this work we have used SVM,DT, PNN, k-NN and NB classifiers. They are briefly described below.

SVM uses data from normal and dry AMD groups to construct a hyperplane, whichseparates the two classes. It can be used for non-linearly separable data using kernel func-tions [55, 54, 56]. In this work, different kernel functions namely linear, Polynomial of order2, 3 and Radial Basis Function (RBF) are used. Moreover, we have selected the RBF sigma(σ = 2) value to tune the classifier performance. The DT classifier performs the classificationtask based on set of rules derived from the input features of the training data [54] (normaland AMD). PNN is a feed forward neural network based on Parzens’ results on probabilitydensity estimators. It has three layers such as input, pattern and summation layer. Thecompete transfer function in the summation layer used the distance vector probabilities tomake the decision during testing [57]. The summation layer spread value (σ) of 0.1 is cho-sen to obtain maximum accuracy. k-NN is a non-parametric lazy learning algorithm usingnearest neighbour principle and makes decision by computing distance between the nearestneighbours [54, 58]. In this work we have used k-NN classifier and found that K = 1 gavethe best performance. NB classifier is derived using Bayes theorem. It assumes that thedata is obtained from normal distribution and also independent. The maximum likelihoodalgorithm is used to compute the class prior probability and feature probability. This model

13

is then used along with a maximum a posteriori decision rule to determine the class label ofunknown data [54].

4. Results

The performance of the proposed automated AMD detection system is evaluated onprivate, ARIA and STARE datasets. AMD progression can be detected by the presenceof drusen, pigmentary irregularities and new vessel proliferation. These abnormal lesionschanging the pixel pattern in the retinal fundus images, which can be quantified usingentropy and non-linear features.

The summary of statistics (Mean ± StandardDeviation(SD)) of extracted features isshown in Figure 4a–4d for private dataset. It shows that the texture values are low forAMD and high for normal (See Figure 4b & 4c), due to the presence of homogeneous pixelpatterns of drusens in the AMD images. Also, most of the HOS entropy values are high incase of AMD (See Figure 4a & 4d), due to the presence of drusens.

(a) (b)

14

(c) (d)

Figure 4: Statistical variation (Mean±SD) of features for private dataset.

The 3D scatter plot of private dataset (Figure 5) shows that the Gabor wavelet basedfeatures are distinct for normal and AMD classes with some overlapping. Gabor functionsoften provide the optimal resolution both in spatial and frequency domain compared to otherwavelet basis functions [33]. Hence it has more discrimination power (Figure 5). Figure 6shows the box plot of FD for normal and AMD classes.

Figure 5: 3D plot of Gabor wavelet features for normal and AMD classes using private dataset.

15

Figure 6: Box plot of FD for normal and AMD classes using private dataset.

The extracted features are ranked using t-test, KLD, CBBD, ROC-based and Wilcoxonranking methods. Further, these ranked features are incrementally nested and then fed toNB, DT, k-NN, PNN and SVM classifiers using the bootstrapping approach [59]. The clas-sification performance is evaluated using 10-fold cross validation. The data set is separatedinto 10 distinct sets, nine sets are used to build the training model and the remaining setis used for testing. It is repeated 10 times with different train-test combinations and finallythe performance measures namely sensitivity, specificity, Positive Predictive Value (PPV)and accuracy are computed over the 10-folds (Table 3–5). A linear SVM outperformed allother classifiers investigated for this problem on DT, PNN, k-NN and NB classifiers.

16

Figure 7: Plot of performance measures versus number of features for various datasets.

Figure 8: Statistical variations (Mean±SD) plot of accuracy for various datasets.

The performance measures plot (Figure 7) of SVM classifier with preprocessing showsthat the 42, 54 and 38 ranked features are significantly contributed to achieve the maximumaccuracy (Mean±SD) of 90.19±5.38%, 95.07±4.81% and 95±6.46% respectively for private,ARIA and STARE datasets using CBBD, t-test and KLD ranking methods. However, SVMclassifier without preprocessing shows that the 56, 19, and 67 ranked features obtained anaverage accuracy(Mean±SD) of 92.96±3.88%, 92.57±7.04% and 85.14±12.95% respectivelyfor private, ARIA and STARE datasets using Wilcoxon, CBBD and t-test ranking methods.The obtained results (with preprocessing) are comparable, since the SD is deviating 5%from the Mean (See Figure 8). From the results we can infer that using a larger set of testimages is good and gives a better estimate of how the method generalizes to unseen data.

17

Table 3: Performance measures of various classifiers for private dataset using CBBD ranking method.

Classifier Numberof featuresused

Accuracy(%)

PPV (%) Sensitivity(%)

Specificity(%)

NB 38 72.78 72.40 73.70 71.85k-NN 57 84.07 86.51 80.74 87.41PNN 56 86.11 89.09 82.59 89.63DT 29 82.59 83.40 81.48 83.70SVM-Linear 42 90.19 91.28 88.89 91.48SVM-Quadratic 31 88.33 90.67 85.93 90.74SVM-Polynomial 8 84.44 87.22 80.74 88.15SVM-RBF 15 87.41 87.72 87.04 87.78

Table 4: Performance measures of various classifiers for ARIA dataset using t-test ranking method.

Classifier Numberof featuresused

Accuracy(%)

PPV (%) Sensitivity(%)

Specificity(%)

NB 6 83.90 85.87 90.18 73.33k-NN 20 82.02 86.11 86.18 75PNN 49 86.99 88.76 92.18 78.33DT 39 80.81 84.11 86.18 71.67SVM-Linear 54 95.07 96.52 96.09 93.33SVM-Quadratic 17 85.77 88.19 90.09 78.33SVM-Polynomial 69 81.47 82.83 89.27 68.33SVM-RBF 44 89.45 92.02 92.09 93.33

Table 5: Performance measures of various classifiers for STARE dataset using KLD ranking method.

Classifier Numberof featuresused

Accuracy(%)

PPV (%) Sensitivity(%)

Specificity(%)

NB 3 78.19 79.05 90 59.17k-NN 6 74.44 85.77 76 66.67PNN 3 76.94 77.86 90 55.83DT 34 79.58 84.33 81 75.83SVM-Linear 38 95 96.67 96 93.33SVM-Quadratic 30 82.64 87.62 88 74.17SVM-Polynomial 3 78.06 77.43 85.50 64.17SVM-RBF 3 82.92 86.10 88 71.67

18

5. Discussion

Fundus image analysis helps in the detection of various posterior segment eye diseasesnamely AMD, Diabetic Retinopathy (DR) and glaucoma [60]. In this work, various non-linear features are extracted from normal and AMD fundus images to perform automateddetection. The yellow pigmentation and drusen results in contrast in the green band andthen it is extracted as information in the features [33]. The main effect of feature rankingwas that the Gabor wavelet and HOS entropies were ranked higher than that of the otherfeatures. Hence, the mean accuracy of the proposed system using private, ARIA and STAREdataset is more than 90% and the SD is 5% (See Figure 8).

A full list of ranked features is not shown in this manuscript because of space limitationsand a summary of observations is provided instead. The ranking of features differed from onedataset to another and also depended on whether the images were subjected to pre-processingwith adaptive histogram equalization. It was observed that Gabor features dominated thelist of top features for both ARIA and STARE datasets when the images were pre-processed.However, entropy and HOS features also figured higher up on the list when there was nopre-processing. For the ARIA dataset and for the private dataset, HOS features figuredin the top 10 features in terms of statistical variation, when there was no pre-processing.On the private dataset, HOS and entropy features appeared in the top 10 even with thepre-processing. The fact that a linear SVM classifier performed the best suggests that thefeatures distributed are clustered and are probably not on non-linear manifolds where akernel based method may have worked better.

The classification results shows that SVM linear kernel achieved highest classificationaccuracy (See Table 3–5) compared to other kernels. Other classifiers such as NB, k-NNand PNN yielded maximum accuracies (Mean±SD) of 83.90±5.82%, 84.07±4.55% and86.99±7.40% using 6, 57 and 49 features respectively for private and ARIA dataset, whereas DT achieved maximum accuracy of 82.59±5.25% using 29 features for private dataset.Hence, we can infer that the SVM classifier with linear kernel achieved maximum accuracyin classifying normal and dry AMD classes. The summary of our results obtained for theautomated diagnosis of AMD is shown in Table 6.

Köse et al. [61] have proposed inverse segmentation method to identify healthy and un-healthy area in the AMD fundus image. Initially, the OD and macula are located usingSobel filter and relative locations of these two structures. Further, region growing algorithmis initialized around macula to segment out healthy and unhealthy regions. Next, the inverseof segmented image is taken to generate the unhealthy segmentation result. Their methodachieved a segmentation accuracy of more than 90%.

Ahmad Hijazi et al. [62] have proposed two stage Case Based Reasoning (CBR) techniqueapplied to histogram of the retinal image to detect AMD. Their approach involved two stages(i) case base generation and (ii) classification of unknown image. Their algorithm achieveda mean sensitivity of 86% compared with graders.

19

Köse et al. [63] have presented statistical segmentation method to differentiate healthyand unhealthy region of macula for automated AMD detection. The statistical texture fea-tures such as intensity variation, distribution, low and high extreme distribution are com-puted on characteristics and sample images after preprocessing. Finally, difference betweenthe intensity distributions of characteristics and sample images segment the unhealthy area.Their method obtained the accuracies of 92.76% and 96-100% to segment out unhealthy andhealthy area respectively.

Hijazi et al. [63] have presented two data mining techniques such as spatial histogramsand hierarchical image decomposition for automated screening of AMD. Initially, the spatialhistogram is generated from different regions by utilizing texture and shape information ofthe AMD featured images. Further, an average signature histogram is generated for eachregion with respect to each class. In the hierarchical decomposition approach the image isdecomposed by angular and circular manner. Their approach achieved an accuracy of 74%and 100% using spatial histogram and hierarchical decomposition respectively.

20

Table 6: Summary of literatures on automated detection of AMD

Authors Method Imageused

Detectedfeature

Classifier Performance mea-sure

Köse et al.(2008) [61]

Sobel filter and re-gion growing

60 Statistical tex-ture features

NA Accuracy-90%

Ahmad Hi-jazi et al.(2010) [62]

Dynamic TimeWarping (DTW)

144 Histogramfeatures

CBR Accuracy-75%

Köse et al.(2010) [63]

Statistical segmenta-tion

80 Statistical tex-ture features

NA Accuracy-92.76%(Unhealthy area)Accuracy-96-100%(Healthy area)

Agurto et al.(2011) [64]

AM-FM decomposi-tion

507 Statisticalmoments andhistogrampercentiles

Partial leastsquare

AUC-0.84 (RetinaInstitute of SouthTexas (RIST))AUC-0.77 (Uni-versity of TexasHealth ScienceCentre in San Anto-nio (UTHSCSA))

Hijazi et al.(2012) [6]

Spatial histogramand hierarchical im-age decomposition

161 Image feature CBR andDTW

Spatial histogramaccuracy-74%Hierarchical decom-position accuracy-100%

Zheng et al.(2012) [65]

hierarchical treesand weighted fre-quent subgraphmining

258 Image feature SVM Accuracy-99.6%

The currentstudy

FD, Gaborwavelet, Shannon,Kapur, Renyi andYager entropymeasures andHOS

Private(540)ARIA(161)STARE(83)

Entropiesand Texture

SVM Accuracy-90.19%(Private)Accuracy-95.07%(ARIA)Accuracy-95%(STARE)

NA 1; NM 2

Zheng et al. [65] have presented automated AMD grading system using image-miningtechniques. The hierarchical trees based decomposition is applied on preprocessed images.

1Not Applicable2Not Mentioned

21

Then weighted frequent subgraph mining algorithm is applied to extract features. Theirmethod obtained an highest classification accuracy of 99.6% using SVM classifier.

Agurto et al. [64] have proposed AM-FM based AMD detection using retinal fundusimages. Their approach decomposed the green channel image into different representationsusing AM-FM method on 140 × 140 pixels Region Of Interest (ROI) which reflects the in-tensity, geometry and texture of the different lesions present in the image. Further, featuressuch as statistical moments and histogram percentiles are computed. Finally, partial leastsquare method is used to classify the features as DR and AMD. Their method was evaluatedin two datasets namely RIST and UTHSCSA and achieved AUC of 0.84 and 0.77 respectively.

The main advantages and uniqueness of the proposed system are summarized below:

• The proposed system does not require any segmentation.

• State of the art accuracy (90.19%, 95.07% and 95%) on larger datasets (private, ARIAand STARE) used for testing (See Table 6).

• Our results are consistent over 10-fold cross validation, which shows that the obtainedresults are robust and reproducible.

• The proposed system uses non-linear features such as HOS entropies and Gaborwavelet, which are more robust to noise. Hence, it can provide accurate results duringmass dry AMD screening.

• The execution time of the optimal system for feature extraction (image data size=540/161/83),training (10-Fold) and testing (10-Fold) are 266.318/39.7368/39.0108 sec, 0.0310/0.0041/0.0037sec and 0.0235/0.0033/0.0032 sec respectively for private/ARIA/STARE dataset us-ing Intel i7-4770 3.47GHz processor, 16GB 1600 MHz CL9 DDR3-RAM and MATLAB2012b computational environment. Hence our proposed system is fast.

• The proposed method can be extended for the detection of other eye diseases like, DRand glaucoma.

6. Conclusion

Regular eye screening will help to diagnose dry AMD and may avoid the loss of visionin old aged subjects. In this work, an automated dry AMD screening system using fundusimages is proposed. The non-linear features extracted using HOS and Gabor wavelet areable to identify the abrupt changes in the normal and dry AMD images. The proposedsystem achieved an average accuracy of more than 90% and the SD of 5% consistently (inall 10-folds) using SVM classifier for all three datasets namely private, ARIA and STARE.Moreover, the testing time of the proposed system is less, hence this system can be used asan adjunct tool for mass screening of eye to diagnose dry AMD cases.

22

Conflicts of interest

The authors do not have any related conflict of interest.

Acknowledgements

Authors thank Social Innovation Research Fund (SIRF/Project Code: T1202), Singaporefor providing grant for this research. Also authors would like to thank National medicalresearch council (NMRC/CSA/045/2012).

AppendixA. Bibliography

[1] P. T. de Jong, Age-related macular degeneration, New England Journal of Medicine 355 (14) (2006)1474–1485.

[2] A. Chopdar, U. Chakravarthy, D. Verma, Age related macular degeneration, BMJ 326 (7387) (2003)485–488.

[3] M. P. S. Group, Risk factors for choroidal neo vascularisation in the second eye of patients withjuxtafoveal or subfoveal choroidal neovascularisation secondary to ageÂrelated macular degen eration,Archives of ophthalmology 115 (0) (1997) 741 – 747.

[4] D. J. Pieramici, S. B. Bressler, Age-related macular degeneration and risk factors for the developmentof choroidal neovascularization in the fellow eye, Current Opinion in Ophthalmology 9 (3) (1998) 38 –46.

[5] J. R. Evans, Risk factors for age-related macular degeneration, Progress in Retinal and Eye Research20 (2) (2001) 227 – 253.

[6] M. H. A. Hijazi, F. Coenen, Y. Zheng, Data mining techniques for the screening of age-related maculardegeneration, Knowledge-Based Systems 29 (0) (2012) 83 – 92.

[7] National statistics office. national population projections,Accessed: 04/07/2013.URL www.statistics.gov.uk/CCI/nugget.asp

[8] World health organization. fact sheet no. 144. blindness and visual dis ability. part iii of vii: otherleading causes worldwide,Accessed: 04/07/2013.URL worldwide.www.who.int/infÂ\discretionary-fs/en/fact144html

[9] M. R. K. Mookiah, U. R. Acharya, C. K. Chua, C. M. Lim, E. Y. K. Ng, A. Laude, Computer-aideddiagnosis of diabetic retinopathy: A review, Computers in biology and medicine 43 (12) (2013) 2136–2155.

[10] T. Lim, A. Laude, Age-related macular degeneration: An asian perspective, ANNALS-ACADEMY OFMEDICINE SINGAPORE 36 (10) (2007) S15.

[11] P. Mitchell, W. Smith, K. Attebo, J. J. Wang, Prevalence of age-related maculopathy in australia. theblue mountains eye study., Ophthalmology 102 (10) (1995) 1450–1460.

[12] Y. Kanagasingam, A. Bhuiyan, M. D. Abràmoff, R. T. Smith, L. Goldschmidt, T. Y. Wong, Progresson retinal image analysis for age related macular degeneration, Progress in retinal and eye research 38(2014) 20–42.

[13] A.-R. E. D. S. R. Group, et al., A randomized, placebo-controlled, clinical trial of high-dose supplemen-tation with vitamins c and e, beta carotene, and zinc for age-related macular degeneration and visionloss: Areds report no. 8, Arch Ophthalmol 119 (10) (2001) 1417–1436.

[14] F. G. Holz, T. J. Wolfensberger, B. Piguet, M. Gross-Jendroska, J. A. Wells, D. C. Minassian, I. H.Chisholm, A. C. Bird, Bilateral macular drusen in age-related macular degeneration. prognosis and riskfactors., Ophthalmology 101 (9) (1994) 1522–1528.

23

[15] Z. Ben Sbeh, L. Cohen, G. Mimoun, G. Coscas, A new approach of geodesic reconstruction for drusensegmentation in eye fundus images, Medical Imaging, IEEE Transactions on 20 (12) (2001) 1321–1333.

[16] K. Rapantzikos, M. Zervakis, K. Balas, Detection and segmentation of drusen deposits on human retina:Potential in the diagnosis of age-related macular degeneration, Medical Image Analysis 7 (1) (2003) 95– 108.

[17] M. Niemeijer, B. van Ginneken, S. R. Russell, M. S. Suttorp-Schulten, M. D. Abramoff, Automateddetection and differentiation of drusen, exudates, and cotton-wool spots in digital color fundus pho-tographs for diabetic retinopathy diagnosis, Investigative ophthalmology & visual science 48 (5) (2007)2260–2267.

[18] P. Soliz, S. Russell, M. Abramoff, S. Murillo, M. Pattichis, H. Davis, Independent component analysisfor vision-inspired classification of retinal images with age-related macular degeneration, in: ImageAnalysis and Interpretation, 2008. SSIAI 2008. IEEE Southwest Symposium on, 2008, pp. 65–68.

[19] E. Barriga, V. Murray, C. Agurto, M. Pattichis, S. Russell, M. Abramoff, H. Davis, P. Soliz, Multi-scale am-fm for lesion phenotyping on age-related macular degeneration, in: Computer-Based MedicalSystems, 2009. CBMS 2009. 22nd IEEE International Symposium on, 2009, pp. 1–5.

[20] D. Freund, N. Bressler, P. Burlina, Automated detection of drusen in the macula, in: BiomedicalImaging: From Nano to Macro, 2009. ISBI ’09. IEEE International Symposium on, 2009, pp. 61–64.

[21] Z. Liang, D. W. K. Wong, J. Liu, K.-L. Chan, T. Y. Wong, Towards automatic detection of age-related macular degeneration in retinal fundus images, in: Engineering in Medicine and Biology Society(EMBC), 2010 Annual International Conference of the IEEE, 2010, pp. 4100–4103.

[22] P. Burlina, D. Freund, B. Dupas, N. Bressler, Automatic screening of age-related macular degenera-tion and retinal abnormalities, in: Engineering in Medicine and Biology Society,EMBC, 2011 AnnualInternational Conference of the IEEE, 2011, pp. 3962–3966.

[23] H. Santos-Villalobos, T. P. Karnowski, D. Aykac, L. Giancardo, Y. Li, T. Nichols, K. Tobin, E. Chaum,Statistical characterization and segmentation of drusen in fundus images, in: Engineering in Medicineand Biology Society,EMBC, 2011 Annual International Conference of the IEEE, 2011, pp. 6236–6241.

[24] A. D. Mora, P. M. Vieira, A. Manivannan, J. M. Fonseca, Automated drusen detection in retinal imagesusing analytical modelling algorithms, Biomedical engineering online 10 (1) (2011) 59.

[25] G. Quellec, S. R. Russell, M. D. Abràmoff, Optimal filter framework for automated, instantaneousdetection of lesions in retinal images, Medical Imaging, IEEE Transactions on 30 (2) (2011) 523–533.

[26] J. Cheng, D. Wong, X. Cheng, J. Liu, N. M. Tan, M. Bhargava, C. Cheung, T. Y. Wong, Earlyage-related macular degeneration detection by focal biologically inspired feature, in: Image Processing(ICIP), 2012 19th IEEE International Conference on, 2012, pp. 2805–2808.

[27] A.-R. E. D. S. R. Group, et al., The age-related eye disease study system for classifying age-relatedmacular degeneration from stereoscopic color fundus photographs: the age-related eye disease studyreport number 6., American journal of ophthalmology 132 (5) (2001) 668.

[28] S. M. Pizer, E. P. Amburn, J. D. Austin, R. Cromartie, A. Geselowitz, T. Greer, B. ter Haar Romeny,J. B. Zimmerman, K. Zuiderveld, Adaptive histogram equalization and its variations, Computer vision,graphics, and image processing 39 (3) (1987) 355–368.

[29] B. B. Mandelbrot, The Fractal Geometry of nature, San Francisco, CA: Freeman, CA, 1982.[30] N. Sarkar, B. Chaudhuri, An efficient differential box-counting approach to compute fractal dimension

of image, Systems, Man and Cybernetics, IEEE Transactions on 24 (1) (1994) 115–120.[31] M. K. Biswas, T. Ghose, S. Guha, P. K. Biswas, Fractal dimension estimation for texture images: a

parallel approach, Pattern Recogn. Lett. 19 (3-4) (1998) 309–313.[32] P. Chowriappa, S. Dua, U. R. Acharya, M. R. K. Mookiah, Ensemble selection for feature-based classi-

fication of diabetic maculopathy images, Computers in biology and medicine 43 (12) (2013) 2156–2162.[33] W. Ma, B. Manjunath, Texture features and learning similarity, in: Computer Vision and Pattern

Recognition, 1996. Proceedings CVPR ’96, 1996 IEEE Computer Society Conference on, 1996, pp.425–430.

[34] B. S. Manjunath, W.-Y. Ma, Texture features for browsing and retrieval of image data, Pattern Analysisand Machine Intelligence, IEEE Transactions on 18 (8) (1996) 837–842.

24

[35] A. Singh, B. Singh, Texture features extraction in mammograms using non-shannon entropies, in: S.-I.Ao, B. Rieger, M. A. Amouzegar (Eds.), Machine Learning and Systems Engineering, Vol. 68 of LectureNotes in Electrical Engineering, Springer Netherlands, 2010, pp. 341–351.

[36] P. A. P. Singh, B. Singh, Shannon and non-shannon measures of entropy for statistical texture fea-ture extraction in digitized mammograms, in: Proceedings of the World Congress on Engineering andComputer Science 2009 Vol II, WCECS 2009, Vol. 2, 2009, pp. 1–6.

[37] A. Laude, M. R. K. Mookiah, C. K. Chua, L. C. Min, E. Y. K. Ng, U. R. Acharya, A novel, integratedcomputer tool for identifying digital fundus images with diabetic retinopathy, in: American Academy ofOphthalmology, Asia-Pacific Academy of Ophthalmology Annual Meeting, 2012, Chicago, USA, 2012,p. 184.

[38] Karmeshu, Entropy measures, maximum entropy principle and emerging applications, Vol. 119,Springer, 2003.

[39] V. P. Singh, Entropy theory and its application in environmental and water engineering, John Wiley& Sons, 2013.

[40] Q. Hu, D. Yu, Entropies of fuzzy indiscernibility relation and its operations, International Journal ofuncertainty, fuzziness and knowledge-based systems 12 (05) (2004) 575–589.

[41] M. R. K. Mookiah, O. Faust, Automated glaucoma detection using hybrid feature extraction in retinalfundus images, Journal of Mechanics in Medicine and Biology 13 (01).

[42] J. Radon, On the determination of functions from their integral values along certain manifolds, MedicalImaging, IEEE Transactions on 5 (4) (1986) 170–176.

[43] Q. Zhang, I. Couloigner, Accurate centerline detection and line width estimation of thick lines usingthe radon transform, Image Processing, IEEE Transactions on 16 (2) (2007) 310–316.

[44] J. M. Mendel, Tutorial on higher-order statistics (spectra) in signal processing and system theory:Theoretical results and some applications, Proceedings of the IEEE 79 (3) (1991) 278–305.

[45] C. L. Nikias, J. M. Mendel, Signal processing with higher-order spectra, Signal Processing Magazine,IEEE 10 (3) (1993) 10–37.

[46] K. C. Chua, V. Chandran, U. R. Acharya, C. M. Lim, Application of higher order statistics/spectra inbiomedical signals –a review, Medical engineering & physics 32 (7) (2010) 679–689.

[47] U. R Acharya, S. Dua, X. Du, S. Vinitha Sree, C. K. Chua, Automated diagnosis of glaucoma usingtexture and higher order spectra features, Information Technology in Biomedicine, IEEE Transactionson 15 (3) (2011) 449–455.

[48] U. R. Acharya, C. K. Chua, E. Y. K. Ng, W. Yu, C. Chee, Application of higher order spectra for theidentification of diabetes retinopathy stages, Journal of Medical Systems 32 (6) (2008) 481–488.

[49] M. R. K. Mookiah, U. R. Acharya, R. J. Martis, C. K. Chua, C. M. Lim, E. Y. K. Ng, A. Laude,Evolutionary algorithm based classifier parameter tuning for automatic diabetic retinopathy grading:A hybrid feature extraction approach, Knowledge-Based Systems 39 (2013) 9–22.

[50] M. R. K. Mookiah, U. R Acharya, C. M. Lim, A. Petznick, J. S. Suri, Data mining technique forautomated diagnosis of glaucoma using higher order spectra and wavelet energy features, Knowledge-Based Systems 33 (2012) 73–82.

[51] B. H. Cho, H. Yu, K.-W. Kim, T. H. Kim, I. Y. Kim, S. I. Kim, Application of irregular and unbal-anced data to predict diabetic nephropathy using visualization and feature selection methods, ArtificialIntelligence in Medicine 42 (1) (2008) 37–53.

[52] A. M. Goon, M. K. Gupta, B. Dasgupta, An outline of statistical theory, World Press Private, 1973.[53] M. Bressan, J. Vitria, On the selection and classification of independent features, Pattern Analysis and

Machine Intelligence, IEEE Transactions on 25 (10) (2003) 1312–1317.[54] R. O. Duda, P. E. Hart, D. G. Stork, Pattern classification, John Wiley & Sons, 2012.[55] V. N. Vapnik, Statistical learning theory, Wiley, 1998.[56] S. Chaplot, L. Patnaik, N. Jagannathan, Classification of magnetic resonance brain images using

wavelets as input to support vector machine and neural network, Biomedical Signal Processing andControl 1 (1) (2006) 86 – 92.

[57] D. F. Specht, Probabilistic neural networks, Neural networks 3 (1) (1990) 109–118.

25

[58] W.-L. Zuo, Z.-Y. Wang, T. Liu, H.-L. Chen, Effective detection of parkinson’s disease using an adaptivefuzzy k-nearest neighbor approach, Biomedical Signal Processing and Control 8 (4) (2013) 364 – 373.

[59] K. Singh, M. Xie, Bootstrap: A statistical method (2008).[60] S. Nirmala, S. Dandapat, P. Bora, Wavelet weighted blood vessel distortion measure for retinal images,

Biomedical Signal Processing and Control 5 (4) (2010) 282 – 291.[61] C. Köse, U. Şevik, O. Gençalioğlu, Automatic segmentation of age-related macular degeneration in

retinal fundus images, Computers in Biology and Medicine 38 (5) (2008) 611–619.[62] M. Ahmad Hijazi, F. Coenen, Y. Zheng, Retinal image classification using a histogram based approach,

in: Neural Networks (IJCNN), The 2010 International Joint Conference on, 2010, pp. 1–7.[63] C. Köse, U. Sevik, O. Gencalioglu, C. Ikibas, T. Kayikicioglu, A statistical segmentation method for

measuring age-related macular degeneration in retinal fundus images, Journal of Medical Systems 34 (1)(2010) 1–13.

[64] C. Agurto, E. S. Barriga, V. Murray, S. Nemeth, R. Crammer, W. Bauman, G. Zamora, M. S. Pattichis,P. Soliz, Automatic detection of diabetic retinopathy and age-related macular degeneration in digitalfundus images, Investigative ophthalmology & visual science 52 (8) (2011) 5862–5871.

[65] Y. Zheng, M. H. A. Hijazi, F. Coenen, Automated disease/no disease grading of age-related maculardegeneration by an image mining approach, Investigative ophthalmology & visual science 53 (13) (2012)8310–8318.

26