Automated Identification of Exudates for Detectionof Macular Edema

Umer Aftab† and M. Usman Akram∗

Department of Computer & Software EngineeringBahria University, Islamabad, Pakistan.

Email: umer6822@hotmail.com†, usmakram@gmail.com∗

Abstract—Macular edema is an advance stage of diabeticretinopathy which affects central vision of diabetes patients.The main cause of edema is the appearance of exudates nearor on macular region in human retina. An automated systemfor early detection of macular edema should identify all possibleexudates present on the surface of retina. In this paper, we presenta method for the identification of exudates in colored retinalimages which will help in building a computer aided diagnosticsystem for macular edema. The proposed system consists ofthree stages i.e. candidate exudate detection, feature extractionand classification. We use filter bank for candidate exudatedetection, basic properties of exudates for feature extraction andGaussian mixture model for classification. This paper presentsthe performance of our system on three retinal image databasesand comparative results with existing methods.

I. INTRODUCTION

Macular Edema (ME) is a common eye disease and advancestage of diabetic retinopathy (DR) which is caused due toincrease of insulin in blood. It is one of the leading cause ofblindness in industrialized countries [1]. ME is a progressivedisease but early detection and diagnosis of ME can savevision loss. Recent studies have shown that one out of fivepatients with newly discovered type II diabetes has DR at thetime of diagnosis where as in first five years after diagnosisof type I diabetes, DR almost never occurs [2].

The common symptoms of ME are blurred vision andsudden loss of central vision [2]. Human retina consists ofblood vessels, optic disc, macula and fovea and any change inthese components can affect the vision [1]. DR is a progressivedisease in which diabetes weakens the blood vessel boundariesand causes the leakage of blood into the retina. If the bloodleakage contains fats and proteins along with water theycause yellow spots known as exudates [3]. The appearanceof exudate on fundus surface decreases the vision but if theaccumulation of fats and proteins is near or on macula, thiscan lead to severe vision loss. This is known as ME and itis the most threatening sign as it can significantly reduce thevision [4]. Early detection of the exudates can save patient’svision which need mass screening. Automated and computeraided diagnostic systems can help the ophthalmologists inmass screening of ME [4]. Figure 1 shows a digital retinalimage with exudates present on surface of retina.

A number of automated systems for retinal image enhance-ment, component segmentation and lesion detection have beenproposed. Ahmed et al. [5] presented a fixed and variable

Fig. 1. Digital retinal image with exudates

threshold based method for OD and exudate detection. Theyused marker controlled watershed transformation for this pur-pose. A highest local variance based method was used bySinthanayothin et al. [6] to locate the position of the OD.Alireza et al. [7] used detailed feature set consisting of color,shape, size and texture for exudate classification. They useda multilayer neural network classifier for this purpose. In [8]candidate regions for exudate were extracted using morpholog-ical closing of the luminance channel, local standard variationin a sliding window and watershed transform. Wang et al. [9]proposed a system that combines brightness adjustment proce-dure with statistical classification method and local-window-based verification strategy for exudate detection. Another FCMbased method for OD and exudate detection was proposedby Haniza et. al [10]. Their system used the applicationsof FCM clustering, edge detection, OTSU thresholding andinverse surface thresholding for this purpose. Akram et al.[11] presented a method for dark and bright lesions detectionusing a hybrid fuzzy based classifier.

we present an automated system for detection of exudatesin colored retinal images. It enhances the candidate exudateregions using Gabor filter bank and creates a binary map forpossible exudate regions. A gaussian mixture model is used toclassify the candidate region as exudate and non exudate. Theproposed system is tested on three publicly available databasesi.e. STARE, DiaretDB0 and DiaretDB1 and compared withalready published techniques.

This paper consists of four sections. Section 2 describes the

27

YassinNew-conf

proposed system in detail including candidate exudate regiondetection, feature set formation and classification of exudate.The results are presented in section 3, followed by conclusionsin section 4.

II. PROPOSED SYSTEM

Digital image processing and machine learning based sys-tems are playing a vital role in biomedical now a days.Computer aided diagnostic systems have brought new horizonsin detection and treatment of many common diseases [3].Similarly, analysis of digital retinal images is now being usedfor the detection and diagnosis of DR. The proposed systemuses image processing and machine learning techniques forthe detection of exudates, a sign of ME, in retinal images.Figure 2 shows the flow diagram for our proposed system.

Fig. 2. Flow diagram for proposed system

A. Candidate Exudate Detection

Exudates also known as bright lesions, appear as brightspots and patches in fundus image with highest contrast inthe green plane of the color image [8]. An automated systemfor detection of exudates should enhance the contrast of brightregions with smoothing of dark regions. For exudate detection,morphological closing is used to smooth dark regions such ashaemorrhages and blood vessels using equation 1 [8]

φ(sB)f = min[maxf(x+ b)] (1)

here f is original colored image and b ∈ sB where sBrepresents the structuring element(SE) B with size s. Thisgives us a smooth fundus region φf containing bright regions

only but they need contrast enhancement. The objective ofcontrast enhancement is to improve the contrast of lesions foreasy detection using a w×w sliding window with assumptionthat w is large enough to contain a statistically representativedistribution of the local variation of lesions [6].

g = 255[Φw(φf )− Φw(φfmin)]

[Φw(φfmax)− Φw(φfmin)](2)

where Φw is the sigmoid function for a window defined as

Φw(φf ) = [1 + exp(mw − fσw

)]−1 (3)

φfmax, φfmin are maximum and minimum intensity value ofsmooth green channel image respectively. mw and σw are themean and variance of intensity values within the window.

We use Gabor filter bank for detection of all possible brightregions. Gabor filters are famous due to their fine frequencytuning and orientation selectiveness. They are appropriate fortexture representation and discrimination [12]. Gabor filter isrepresented by a Gaussian kernel function which can model awide range of shapes depending upon values of its parameters[12]. This property makes them suitable for detection ofexudates. Equation describes the filter bank used in proposedsystem.

GFB =1√πrσ

e−12 [(

d1σ )

2+(d2σ )

2](d1(cosΩ + ιsinΩ)) (4)

where σ, Ω and r are the standard deviations of Gaussian, spa-tial frequency and aspect ratio respectively θ is the orientationof filter and d1 = xcosθ + ysinθ and d2 = −xsinθ + ycosθ.The contrast enhanced image g is convolved with Gabor filterG centered at location(s,t) to generate Gabor filter response γfor selected values of σ, Ω and θ is given in equation 5 [12].

γ(σ,Ω, θ) =∑x

∑y

g(x, y)GFB(s− x, t− y, σ,Ω, θ, r) (5)

For considered frequency and scale values, the maximum Ga-bor filter bank response Mγ(σ,Ω) is computed using equation6 for θ spanning from 45o up to 180o at steps of 45o.

Mγ(σ,Ω) = max|γ(σ,Ω, θ)| (6)

The binary candidate regions for exudates are extracted fromMγ by applying a low adaptive threshold value T [13].The regions segmented by thresholding of filter bank basedenhanced image also contain OD region and pixels due totheir similarity with exudates. For accurate detection exudates,these false and spurious pixels should be removed before theclassification stage. The proposed system locates and segmentsoptic disc using image averaging and hough transformationrespectively [14]. Figure 3 shows the step by step outputs forcandidate exudate region detection.

B. Feature Extraction

Exudates appear as bright yellow spots with variable sizeand shape but they have strong and sharp edges. Candidate ex-udate region extraction stage gives all possible regions that can

28

Fig. 3. Candidate exudate region detection: a) Original colored retinal image;b)Smoothing of dark components with mathematical morphological closing;c)Contrast enhanced bright regions; d)Enhanced bright regions using filterbank; e)Segmented bright regions using adaptive thresholding; f)Candidateexudate regions after removal of OD pixels

be considered as potential exudates. Each object or candidateexudate region is considered as sample for classification andrepresented by a feature vector. The description of features,we used for classification of exudate and no exudate regionsare as following:Area (x1), is the count of number of pixels in candidateexudate region and defined as A =

∑υi

1 sum of all pixels incandidate region υi.Compactness (x2) is measure of shape defined as C =p2/4piA where p and A are the perimeter length and areaof candidate region respectively.Mean Intensity (x3) is the mean intensity value of contrastenhanced green channel for all pixels within the candidateregion.Mean Hue (x4), mean Saturation (x5) and mean Value (x6) foreach candidate region are calculated in order to differentiateexudate and non exudate regions on basis of their colorproperties.Mean gradient magnitude (x7) for edge pixels is computed todifferentiate between strong and blur edges.

C. Classification

In order to classify candidate region as exudate and nonexudate region, we use a Bayesian classifier using Gaussian

functions known as Gaussian Mixture Model (GMM) [15].We define two classes such as R1 = {Exudate region}and R2 = {Nonexudate region}. A supervised classificationmethod is used for final classification by dividing the datasetinto training and testing subsets which will be defined in nextsection. The classifier is trained using the training dataset andwe used Bayes decision rule to obtain a decision rule based onestimates from the training set. Bayes decision rule is statedas [16]

Choose R1 if, p(v|R1)P (R1) > p(v|R2)P (R2)otherwise choose R2 (7)

where p(v|Ri) is the class conditional probability densityfunction also known as likelihood and P (Ri) is the priorprobability of class Ri which is calculated as the ratio of classRi samples in the training set. We describe the likelihood aslinear combination of Gaussian function in equation 8

p(v|Ri) =κi∑j=1

p(v|j,Ri)ωi (8)

where κi is the number of Gaussian mixtures used forBayesian classification and p(v|j,Ri) is a m-dimensionalGaussian distribution of weight ωi and Ri = {R1, R2} are thetwo classes used in proposed system. We apply expectationmaximization to search for an optimal value of κ whichoptimizes the accuracy of GMM using different validation setsrandomly extracted from classified training data.

III. EXPERIMENTAL RESULTS

The evaluation and testing of computer aided diagnosticsystems are very important. We use three standard retinalimage databases which are publicly available for testingand evaluation of DR screening and diagnostic algorithms.First dataset is named as STurctured Analysis of the REtina(STARE) which is developed by Hoover et al. for analysisof vascular structure [19]. It consists of retinal images withresolution of 700 × 605 acquired using a TopCon TRV-50fundus camera. For more evaluation and comparison purposes,two more databases names as DiaretDB0 [17] and DiaretDB1[18] are used. Two trained human graders created groundtruths of manually labeled exudate regions for these databasesand we used them for evaluation purposes. Figure 4 showsthe segmented exudate regions using proposed method forDiaretDB0, DiaretDB1 and STARE databases.

The performance of proposed system is measured usingsensitivity, specificity, positive predictive value (PPV) andaccuracy as figures of merit. Sensitivity is true positive rateand specificity is true negative rate.

For performance comparison, we present the values of theseparameters for Sinthanayothin et al. [6], Wang et al. [9], walteret al. [8], ahmed et al.[5], Osareh et al.[7], Haniza et al.[10] and akram et al. [11]. Table-1 shows the performancecomparison of proposed method with these methods.

For further and more comparative purposes, we comparedthe results with recently published methods proposed by

29

Fig. 4. Randomly selected slices from retinal image. Exudate are highlightedwith green boundary

TABLE IPERFORMANCE COMPARISON OF EXUDATE DETECTION

Method Sensitivity Specificity PPV Accuracy

Sinthanayothin et al. [6] 88.5 99.7 - -Wang et al. [9] - 70 - 100walter et al. [8] 92.74 100 92.39 -ahmed et al.[5] 96.7 100 94.9 -Osareh et al.[7] 93 94.1 - 93.4Haniza et al. [10] 94.25 99.2 78.65akram et al. [11] - - - 94.73Proposed Method 96.36 98.25 97.45 97.59

Ahmed et al. [5] and Haniza et al. [10] for STARE database.Table-2 shows the performance comparison of exudate seg-mentation for STARE database in terms of true positive rate,true negative rate and PPV.

TABLE IIPERFORMANCE COMPARISON OF EXUDATE SEGMENTATION FOR STARE

DATABASE

Method Sensitivity Specificity PPV

Ahmed et al. [5] 63.6 98.6 79Haniza et al. [10] 97.8 99 83.3Proposed Method 97.72 96.15 95.56

IV. CONCLUSION

In this paper, we presented a method for exudate detectionin colored retinal image. The proposed system consisted ofthree phases that are candidate exudate detection, featureextraction and classification. The bright regions are enhancedand segmented using filter bank and adaptive thresholdingbut they contained spurious regions which we eliminated byremoving the OD pixels. Feature set for each candidate regionis formed using different properties of exudates such as color,shape and statistical. We implemented a GMM based classifierto divide the regions into two classes that are exudate and non

exudate regions. The results demonstrated that the proposedsystem can be used in computer aided diagnostic system forDR as it identified and detected exudates with high accuracies.

REFERENCES[1] Kohner EM, Aldington SJ, Stratton IM, Manley SE, Holman RR,

Matthews, ”DR: United Kingdom Prospective Diabetes Study, 30: diabeticretinopathy at diagnosis of noninsulin-dependent diabetes mellitus andassociated risk factors”, Arch Ophthalmol, vol. 116, No. 3, pp. 297-303,1998.

[2] Amos AF, McCarty DJ and Zimmet P, “The rising global burden ofdiabetes and its complications: estimates and projections to the year2010”, Diabet Med, 1997.

[3] Molven A, Ringdal M, Nordbo AM, Raeder H, Stoy J, Lipkind GM,”Mutations in the insulin gene can cause MODY and autoantibody-negative type 1 diabetes”, Diabetes, vol. 57, No. 4, 1131-1135, 2008.

[4] Frank R. N., “Diabetic retinopathy”, Prog. Retin. Eye Res., vol. 14, No.2, pp. 361-392, 1995.

[5] Ahmed Wasif Reza, C. Eswaran, Subhas Hati, “Automatic Tracing ofOptic Disc and Exudates from Color Fundus Images Using Fixed andVariable Thresholds”, Journal of Medical Systems, vol. 33, pp. 7380,2009.

[6] C. Sinthanayothin, J.A. Boyce, H.L. Cook, and T.H. Williamson, “Auto-mated localisation of the optic disc, fovea, and retinal blood vessels fromdigital colour fundus images”, British Journal of Ophthalmology, vol. 83,pp. 902-910, 1999.

[7] Alireza O., B. Shadgar and R. Markham, “A Computational-Intelligence-Based Approach for Detection of Exudates in Diabetic RetinopathyImages”, IEEE Trans on Information Tech in Biomedicine, vol. 13, No.4, 535-545, 2009.

[8] M. U. Akram, A. Tariq, M. A. Anjum, M. Y. Javed, “Automated De-tection of Exudates in Colored Retinal Images for Diagnosis of DiabeticRetinopathy”, OSA Journal of Applied Optics, vol. 51, No. 20, pp. 4858-4866, 2012.

[9] Wang, H., Hsu, W., Goh, K., and Lee, M., “An effective approach to detectlesions in colour retinal images”, Proceedings of the IEEE Conferenceon Computer Vision and Pattern Recognition 2, 181-187(2000).

[10] Haniza Yazid, Hamzah Arof, Hazlita Mohd Isa, “Automated Identifica-tion of Exudates and Optic Disc Based on Inverse Surface Thresholding”,Journal of Medical Systems, DOI 10.1007/s10916-011-9659-4 (2011).

[11] M. U. Akram and S. A. khan, “Automated Detection of Dark and BrightLesions in Retinal Images for Early Detection of Diabetic Retinopathy”,Journal of Medical Systems, DOI 10.1007/s10916-011-9802-2, 2011.

[12] J. Sung, S. Y. Bang, S. Choi, “A Bayesian Network Classifier andHierarchical Gabor Features for Handwritten Numeral Recognition”,Pattern Recognition Letters, 2005.

[13] R. C. Gonzalez and R. E. Woods, “Digital image processing”, Prenticehall. second edition, 2002.

[14] M. U. Akram, A. Khan, K. Iqbal and W. H. Butt, “Retinal Image:Optic Disk Localization and Detection. Image analysis and Recognition”,Lecture Notes in Computer Science, (Berlin, Heidelberg: Springer),LNCS 6112, Portugal, pp. 40-49, 2010.

[15] S. Theodoridis and K. Koutroumbas, “Pattern Recognition”, 1st ed.Burlington MA, Academic, 1999.

[16] R. O. Duda, P. E. Hart, and D. G. Stork, “Pattern Classification”, NewYork, Wiley, 2001.

[17] Kauppi, T., Kalesnykiene, V., Kamarainen, J.K., Lensu, L., Sorri, I.,Uusitalo, H., Klviinen, H., Pietil, J. “DIARETDB0: Evaluation Databaseand Methodology for Diabetic Retinopathy Algorithms”, Technical report,2005

[18] Kauppi, T., Kalesnykiene, V., Kamarainen, J.-K., Lensu, L., Sorri,I., Raninen A., Voutilainen R., Uusitalo, H., Klviinen, H., Pietil, J.,“DIARETDB1 diabetic retinopathy database and evaluation protocol”,Technical report, 2006.

[19] A. Hoover, M. Goldbaum, “Locating the optic nerve in a retinal imageusing the fuzzy convergence of the blood vessels”, IEEE Trans. onMedical Imaging, vol. 22, No. 8, pp. 951-958, 2003.

30