[email protected]

Retinal Image Analysis for Diagnosis of MacularEdema using Digital Fundus Images

Zainab Yousaf Zaidi , M. Usman Akram, Anam TariqCollege of Electrical and Mechanical Engineering,

National University of Sciences and TechnologyBahria University

PakistanEmail: [email protected],[email protected], [email protected]

AbstractDigital fundus images are one of the modern andadvanced approaches of creating image of inner surface of humaneye emphasizing retina. These fundus images are really helpfulin diagnosis of possible abnormalities and severe diseases likediabetic macular edema and its various types. Research hasshown that early detection and treatment can prevent total visionloss and severe impacts on human visual system. Hence anautomated system for diagnosing macular edema will help theophthalmologists and patients. In this paper, we have proposed anovel method for diagnosing macular edema using fundus images.The technique has four steps which constitutes of preprocessing,macula detection, feature extraction of possible exudates regionfollowed by classification using Nave Bayes classifier. The pro-posed system is tested using MESSIDOR database and resultsshow that our method outperformed others in terms of accuracy.

I. INTRODUCTION

Study in medical imaging has shown that most of the eyediseases screen themselves in the human retina. This leads toa clear road for the researchers to devise and improve themethods for analyzing the human retina. So to obtain theproper image of retina for analysis, specialized techniques areused. The first retinal image creation effort dates back to 1823when Jean Mery, a French physician applied his method ofcreating retinal image on cat [1].

After his successful experiment the process of creatingretinal images evolved and today we have very sophisticatedand reliable equipment for generating digital fundus imagesof human retina. This derives to increased usage of retinalimages by medical professionals for observation of diseaseprogression in recent years. This technique not only helpsthe medical professionals in timely disease diagnosis but alsoimproves medical facilitation available to the patients. Retinalanalysis can really help in detecting the presence of eyediseases and modern medication and treatment methodologycan prevent from major impacts on human visual system [1].Figure 1 shows digital image of human retina and its maincomponents.

The two types of macular edema are Non Clinically Signif-icant Macular Edema (Non-CSME) and Clinically SignificantMacular Edema (CSME). Non-CSME is a mild form of edemain which there are no symptoms of the disease because thelocations of exudates are at a distance from fovea and the

Fig. 1. Digital image of human retina along with its main components andexudates

central vision is not affected. CSME is the severe form ofedema in which the exudates leak out and get deposited veryclose to or on fovea affecting central vision of the eye [1].Figure 2 shows the affect of macular edema on human vision.

Fig. 2. a) Vision with Normal Eye, b) Vision with Moderate Diabetic MacularEdema, c) Vision with Severe Diabetic Macular Edema

Many approaches for diagnosing macular edema nd otherretinal diseases are discussed in literature [2]-[5]. Deepak et.al. [6] have described their solution of automatic assessmentof DME from color fundus images. Global characteristics ofimage are captured in feature extraction and disease severity ismeasured using rotational symmetry. Experimental results areobtained by classifying the infected images and accuracy ofmore than 74% is observed in all classes. In [7], Hunter et. al.throws light on very intense condition of maculopathy calledreferable maculopathy, which needs straightaway attention.The specified algorithm uses optic nerve head position to

2013 IEEE Jordan Conference on Applied Electrical Engineering and Computing Technologies (AEECT)

978-1-4799-2303-8/13/$31.00 2013 IEEE

locate the centre of macula. Image is then analyzed forpossible bright lesions which are classified using multi-layerperceptron. In [8], the authors have described the presenceand cause of cystoid macular edema (CME) in patients. Theproposed method includes a computationally fast bilateral filterfor speckle de-noising while maintaining CME boundaries.The results showed the accuracy of more than 91%. Sid-dalingaswamy et. al. [9] proposed a method in which theyfind macula using darkest region and exudates are detectedusing clustering and mathematical morphology techniques.Disease severity is detected using location of exudates frommacula. The sensitivity was 95% and the specificity was 96%but the method is tested on local dataset only. Osareh et.al. [10] identified macula as minimum intensity region andexudates are segmented using contrast enhancement, FCMclustering followed by neural network based classification. Thesensitivity was 92% and the specificity was 82%. Lim et. al.[11] used marker controlled watershed segmentation for thesegmentation of exudates region and tested the algorithm on88 images of MESSIDOR and found the sensitivity/specificityto be 80/90%. In [12], Giancardo et. al. proposed a newmethodology for DME diagnosis. A feature set for completeimage based on color, wavelet decomposition and lesionprobability map is used. Classification was performed usingdifferent individual classifiers. The have also developed adatabase HEI-MED and the AUC was found to be 0.88.

In this paper, we present an automated system for detectionof macular edema in colored fundus images. The proposedmethodology consists of preprocessing stage in which back-ground estimation is done and the noise is eliminated fromthe fundus image. The next stage is to extract macula usingminimum intensity region in fundus images. The vesselsare also extracted and removed in this stage. This stage isfollowed by possible extraction of candidates exudates regionand classification is done using Nave Bayes classifier.

This paper comprises of four sections. Section 2 containsproposed solution followed by experimental results in section3. Section 4 contains conclusion.

II. PROPOSED SYSTEM

Diabetes has spread around the globe and we can easilyfind people suffering from diabetes around us no matter inwhich part of earth we are in. This ratio is even increased incountries with more industries. CAD based systems in certaincountries are necessary to prevent the patients from retinaldiseases and vision degradation problems. We have presenteda system for classification of macular edema to identify itsseverity so that proper precautionary measure can be takento avoid the possibility of vision loss. The proposed systemconsists of a series of steps going through which the severityof macular edema is classified as one of the predeterminedseverity levels by the system. The accuracy of the algorithmis increased due to its ability to remove the blood vesselsand optic disc from the retinal images so that the presenceof macula and exudates can be detected accurately. Figure 3shows the flowchart of the proposed solution.

Fig. 3. Flowchart Of Proposed Solution

A. Preprocessing

In preprocessing, background and noise removal is done.The objective of filtering un-necessary pixels from backgroundand foreground is to pass better quality image in followinglayers for advanced processing. For this purpose, backgroundestimation is done using mean and variance based method andHSV channels are utilized for noise removal. The detail ofpreprocessing is explained in our earlier published work [13].

B. Macula Detection

Automated systems for classifying diabetic macular edematake it very important to detect macula from fundus images.As it is one of the most important steps on which the wholetreatment will be based. Macula detection in the proposedscheme works by localizing OD and taking the distance fromthe center of optic disk for detecting the darkest pixels. Bloodvessels are segmented [15] and removed already so that theymay not interfere in the macula detection as both are of similarintensity. The details of macula detection are proposed in ourearlier published work [14]. Figure 4 shows the outputs ofpreprocessing and macula detection.

C. Candidate Exudates Detection

Exudates can be defined as bright lesions visible on theretina in cases when blood contains fats and proteins alongwith the water. This is one of the deadliest threats to the visualsystem. Optic disc sometimes makes it harder to identify thepresence of exudates as both are of same intensity. To make iteasier, the proposed system effectively detects and removes the

2013 IEEE Jordan Conference on Applied Electrical Engineering and Computing Technologies (AEECT)

Fig. 4. Preprocessing and macula detection: a) original retina image; b)background mask; c) detected macula

optic disc from image pixels so that the chances of identifyingthe presence of exudates can be increased. Exudate detectionprocess consists of following steps [4].

1) Morphological closing is applied on the preprocessedimages to remove the presence of blood vessels.

2) Adaptive contrast enhancement techniques are appliedto increase the contrast on the retinal surface.

3) Gabor kernel based filter bank is created for furtherenhancing the bright lesions.

4) Possible exudate areas are embedded in a binary mapafter applying threshold using OTSU algorithm.

5) OD is detected using Hough transform and is removedfrom binary map. This is done in our earlier publishedwork [16].

Figure 5 shows the outputs of candidate exudate detectionphase.

Fig. 5. Candidate exudate detection: a) Filter bank response to enhanceexudate regions; b) Optic disc localization; c) Binary map for candidateexudate regions

D. Feature Extraction

In this step, the possible exudate regions are taken asinput. The threshold value for filtering is kept low by designso that the possibility of detecting smaller exudates can beincreased. Pixels not lying in the exudate category are removedin classification phase. The appearance of exudates can benoticed as bright yellow spots which can have different sizesand shapes having strong and sharp edges. If there are npossible regions for image x, in which presence of exudatescan be noticed, the set of exudates can be represented asfollows X = v1, v2, v3, , vn and this whole set hasto be processed for detecting each exudate and non-exudateregion. Every potential candidate exudate is considered asa sample vector input for the classification algorithm. The

classification algorithm uses certain features for making thedecisions. The feature vector comprises of area, compactness,mean intensity, mean HSV values, mean gradient magnitude,energy, entropy and third moment.

E. Classification

In the process of identifying lesions, we have used the lowvalue of threshold by choice so that we can get most out of theimage in forms of possible exudates. The doubtful detectionsare removed in the phase of classification. In the proposedscheme, we have used Nave Bayes classifier [17].

We have used two classes R1 = Exudate region and R2 =Non exudate region. The final decision of classifying regionsis made by using a supervised classification method. The inputdata set is divided into subsets for training. The classifier istrained using the earlier created sample dataset. Bayes decisionrule is used to fetch a decision rule based on estimates fromthe training set. Bayes decision rule is stated as [17].

Choose R1 if, p(v|R1)P (R1) > p(v|R2)P (R2)otherwise choose R2 (1)

Here p(vj |Ri) can be defined as the class conditional probabil-ity Density Function and P (Ri) is the prior probability of classRi which is calculated as the ratio of class Ri samples in thetraining set. Nave Bayes classifier assumes that the presenceor absence of a particular feature is unrelated to the presenceor absence of any other feature, given the class variable. Itis a probabilistic model and can be trained very efficiently ina supervised learning setting. An advantage of Nave Bayesis that it only requires a small amount of training data toestimate the parameters (means and variances of the variables)necessary for classification [18].

After the classifier has received all the inputs, it classifiesthe image as either of healthy, Non CSME and CSME basedon the location and distance between macula and exudates ifpresent in retinal image. Table 1 show the three conditionswhich have been used to grade the input image into differentstages of macular edema.

TABLE ICONDITIONS FOR GRADING OF MACULAR EDEMA

Grade Condition Class

0 No exudate present Healthy

1 A few exudates present and distance between Non-

macula and exudates > one disc diameter CSME

2 Exudates present and distance between CSME

macula and exudates one disc diameter

III. EXPERIMENTAL RESULTS

We have used standard retinal image database MESSIDORfor the testing and evaluation of algorithms implemented forscreening of macular edema [19]. The database is collected

2013 IEEE Jordan Conference on Applied Electrical Engineering and Computing Technologies (AEECT)

using TopCon TRC NW6 Non-Mydriatic fundus camera with45o FOV and resolutions of 1440 960, 2240 1488 or2304 1536 with 8 bits per color plane. It contains total 1200images which are divided into three sets of 400 images andeach set is further divided into 4 parts to facilitate thoroughtesting. Annotations at pixel level are performed with the helpof an ophthalmologist who used a MATLAB based annotationtool to mark the regions as exudates.

The performance of the proposed system is computed usingmetrics such as sensitivity (true positive rate), specificity (truenegative rate) and accuracy. The parameters are calculated bycomparing the proposed system with ground truth data. Thevalues of these parameters are given in Table 2. Table 2 alsoshows the performance comparison of exudates segmentationat pixel level. It shows that our method surpasses the othermethods because of the use of detailed feature set and accurateclassifier. Table 3 shows the performance comparison of grad-ing of macular edema with other techniques in the literature.

TABLE IIPERFORMANCE COMPARISON OF PROPOSED SYSTEM WITH EXISTING

SYSTEMS FOR EXUDATE DETECTION

Method Sensitivity Specificity Accuracy(%) (%) (%)

Sinthanayothin et al. [21] 88.5 99.7 -Osareh et al.[10] 93 94.1 93.4Haniza et al. [22] 94.25 99.2 78.65Proposed Method 94.61 95.15 95.03

TABLE IIICOMPARISON OF OUR PROPOSED METHOD WITH EXISTING TECHNIQUES

FOR GRADING OF MACULA EDEMA USING MESSIDOR DATABASE

Author Sensitivity Specificity Accuracy

(%) (%) (%)

Lim et. al. [11] 80.9 90.2 -

Deepak [6] et al. 95 90 -

Aquino et. al. [20] - - 96.51

Proposed Method 93.9 95.8 94.1

Figure 6 and 7 shows the pictorial results for proposed systemusing MESSIDOR database. Figure 6 shows the results formacula and exudate detection. Figure 7 represents the gradingof macular edema.

IV. CONCLUSION

Diabetic patients often suffer from macular edema which ishard to detect at early stages and can lead to total vision loss.In this paper, we have proposed an automated system whichconsists of four stages comprising of preprocessing, maculadetection, candidate exudates detection and classification. Wehave tested our system on MESSIDOR database and it gives93.9% sensitivity, 95.8% specificity and 94.1% accuracy. Thissystem will help in early screening of diabetic macular edema.

Fig. 6. Row 1: Macula detection; row 2: exudate detection

Fig. 7. Grading of macula edema: a) healthy images; b) images graded asNon-CSME ; c) images graded as CSME

V. ACKNOWLEDGMENT

This research has been funded by National ICT R&D Fund,Pakistan.

REFERENCES

[1] M.D. Abramoff, M.K. Garvin, Sonka, M., Retinal Imaging and Im-age Analysis, IEEE reviews in Biomedical Engineering, vol.3, no.,pp.169,208, 2010

[2] 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 (JOMS), vol. 36, no. 5, 3151-3162, 2012.

[3] A. Tariq, M. U. Akram, A. Shaukat, S. A. Khan, Automated Detectionand Grading of Diabetic Maculopathy in Digital Retinal Images, Journalof Digital Imaging, vol. 26, no. 4, pp. 803-812, 2013.

[4] 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, 4858-4866,2012.

[5] M. U. Akram, S. Khalid, S. A. Khan, Identification and Classificationof Microaneurysms for Early Detection of Diabetic Retinopathy, PatternRecognition, vol. 46, no.1, 107-116, 2013

2013 IEEE Jordan Conference on Applied Electrical Engineering and Computing Technologies (AEECT)

[6] K. Sai Deepak and Jayanthi Sivaswamy,Automatic Assessment of Mac-ular Edema From Color Retinal Images,IEEE TRANSACTIONS ONMEDICAL IMAGING, VOL. 31, NO. 3, MARCH 2012

[7] A. Hunter et al, Automated Diagnosis of Referable Maculopathy inDiabetic Retinopathy Screening, 33rd Annual International Conferenceof the IEEE EMBS Boston, Massachusetts USA, August 30 - September3, 2011

[8] Wilkins et al, Automated Segmentation of Intraretinal Cystoid Fluidin Optical Coherence Tomography, IEEE TRANSACTIONS ONBIOMEDICAL ENGINEERING, VOL. 59, NO. 4, APRIL 2012

[9] P.C. Siddalingaswamy, K. G. Prabhu, Automatic Grading of DiabeticMaculopathy Severity Levels, International Conference on Systems inMedicine and Biology 2010, pp. 331-334.

[10] A. Osareh, M. Mirmehdi , B. Thomas and R. Markham, AutomaticRecognition of Exudative Maculopathy using Fuzzy C-means Clusteringand Neural Networks, International Conference on Medical ImageUnderstanding and Analysis 2010, pp. 49-53.

[11] S.T. Lim, W.M.D.W. Zaki, A. Hussain, S.L. Lim, S. Kusalavan, Clas-sification of Diabetic Macular Edema in Digital Fundus Images, IEEEColloquium on Humanities, Science and Engineering 2011, pp. 265 -269.

[12] Giancardo L, Meriaudeau F, Karnowski TP, Li Y, Garg S, Tobin KW Jr,Chaum E, Exudate-based Diabetic Macular Edema Detection in FundusImages Using Publicly Available Datasets, Medical Image Analysis, Vol.16, No. 1, pp.216-226, 2012.

[13] A. Tariq and M. U. Akram, An automated system for colored retinalimage background and noise segmentation, IEEE Symposium on Indus-trial Electronics and Applications (ISIEA), 2010, pp. 405-409.

[14] M. Mubashhar, A. Tariq, M.U. Akram, Automated System for MaculaDetection in Digital Retinal Images IEEE International Conference onInformation and Communication Technologies ICICT 2010, pp. 1-5.

[15] M. U. Akram and S. A. Khan, Multilayered Thresholding Based BloodVessel Segmentation for Screening of Diabetic Retinopathy, Engineeringwith Computers (EWCO), vol. 29, no. 2, pp. 165-173, 2013.

[16] M. U. Akram, A. Khan, K. Iqbal, and W. H. Butt, Retinal image: opticdisk localization and detection, Image Analysis and Recognition, LectureNotes in Computer Science 6112, 2010, pp. 40-49.

[17] G. H. John and P. Langley , Estimating Continuous Distributionsin Bayesian Classifiers. Proceedings of the Eleventh Conference onUncertainty in Artificial Intelligence, 1995. pp. 338-345

[18] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern Classifcation. NewYork: Wiley, 2001

[19] MESSIDOR Database, http://messidor.crihan.fr/index-en.php[20] A. Aquino, M.E. Gegundez, D. Marin, Automated Optic disc detection

in retinal images of patients with diabetic retinopathy and risk of macularedema, International Journal of Biological and Life Sciences, Vol. 8, No.2, pp. 87-92, 2012

[21] C. Sinthanayothin, J. F. Boyce, H. L. Cook, 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

[22] H. Yazid, H. Arof, H. Mohd Isa, Automated Identification of Exudatesand Optic Disc Based on Inverse Surface Thresholding, Journal ofMedical Systems, DOI 10.1007/s10916-011-9659-4, 2011.

2013 IEEE Jordan Conference on Applied Electrical Engineering and Computing Technologies (AEECT)