Dow

nloadedfrom

http://journals.lww.com

/jcejournalbyf0tkv+j1BdaEtB1IR

7T4ZHFBuyEw

JK5Vfxto7W1H

QO5U

ORd5aPkBJIg2lln1t4R

Zcm4N

XVXJFDkBw

I0bFzfbkKzM0Z8SC

NTyrM

L3CDMV+88H

/4uicN/QZH

dQRS0+hhX5

on10/18/2019

Downloadedfromhttp://journals.lww.com/jcejournalbyf0tkv+j1BdaEtB1IR7T4ZHFBuyEwJK5Vfxto7W1HQO5UORd5aPkBJIg2lln1t4RZcm4NXVXJFDkBwI0bFzfbkKzM0Z8SCNTyrML3CDMV+88H/4uicN/QZHdQRS0+hhX5on10/18/2019

Glaucoma Detection Using Retinal Nerve FiberLayer Texture Features

Arwa A. Gasm Elseid, PhD and Alnazier O. Hamza, Prof.

The retinal nerve fiber layer (RNFL) is one of the mostaffected parts of the eye retina by glaucoma disease.Progression of this disease results in RNFL texturechanges and can be observed from the fundus imagesand RNFL thickness; thus, the glaucoma is one of theocular diseases contributing to most of the blindnessworldwide. There are increasing demands for medicalimage–based computer-aided diagnosis system based onfundus image for glaucoma detection. Thus, the RNFLthickness changes can be assessed by optical coherencetomography. However, an examination using the opticalcoherence tomography device is rather expensive and stillnot widely available in resource-poor countries. On theother hand, a fundus camera is a fundamental diagnosticdevice and can be used for diagnosing other diseases.In this article, we introduce glaucoma detection usingfundus images. This algorithm uses texture analysisbased on co-occurrence matrix and Tamara features.The texture measures are extracted from the RNFLsegmented part. The method was tested on a set of158 images composed of 118 healthy retina images and40 glaucomatous images and achieved an area under thecurve of 93% and accuracy of 89.5%.

Computer-aided diagnosis, the way to automate the detec-tion process for glaucoma disease, attracts extensive atten-tion from clinicians and researchers.

Therefore, if glaucoma is diagnosed early enough, itcan be properly managed to prevent major loss of vision.Glaucoma is most common cause of permanent blindnessworldwide; there is no cure for glaucoma, but medicationcan be used to prevent vision loss.1 In the case of glaucomain which the visual field test result and intraocular pres-sure assessment are not reliable, the visual field measure-ment is usually carried out using a static perimeter, which

is difficult for different diseases, for example, advancedage-related macular degeneration, and it needs patientcorporation to follow the instruction and the intraocularpressure not sensitive because, in some types of glaucoma,the pressure is normal, which is called normal tensionglaucoma. The best way to detect glaucoma is optic nervehead (ONH), and the imaging modalities used (opticalcoherence tomography [OCT], Heidelberg retina tomogra-phy) are expensive. Then, it is necessary to search for othermethods that will enable to determine the glaucoma dis-ease. One suchmethod is the analysis of digital fundus im-ages of the eye fundus via taken fundus camera device.Features extraction techniques in fundus images are classi-fied based on the type of features. The types of features aredivided into 2 groups, namely, morphological and non-morphological.2 The morphological features based on op-tic disc and optic cup segmentation then calculate featuressuch as retinal nerve fiber layer (RNFL), Cup to DiscRatio, per-papillary atrophy (PPA), and Inferior, Superior,Nasal, Temporal rule.

Nonmorphological features are whole image featureswithout segmentation such as color, shape, and texture, thetypes of features that are captured from the whole image.

This article is a combination of the 2 methods by seg-menting the RNFL part and extracting the texture featuresfrom the specific part.

The use of texture features extraction was performedby Kavya and Padmaja,3 which detects glaucoma usingONH texture, which is the region that consists the opticcup and optic disc. The features are extracted from theONH fundus image by using Hough transformation andk-means for segmentation. From the segmentedONHpart,the different texture features used are gray level co-occurrencematrix (GLCM) andMarkov random field. Thus, the struc-tural changes take place in ONH; the texture and the inten-sity values also change. These features are used to classify thefundus images as normal and glaucoma. The obtained re-sults have approximately 94% accuracy in segmentationusing Hough transformation, 84% for segmentation usingk-means clustering, and approximately 86% for classifica-tion using support vector machine classifier. In another re-search done by Oh et al,4 they proposed a fully automaticmethod for detecting various forms and widths of RNFLdefects in color fundus images. Fundus photography isthe most common screening tool to detect RNFL defectsin various optic neuropathies. However, the detection of

Corresponding author: Arwa A. Gasm Elseid, PhD, is in the BiomedicalEngineering Department, College of Engineering, Sudan University ofScience and Technology, in Khartoum, Sudan, and can be reached atroro19842016@gmail.com.

Alnazier O. Hamza, is a professor and dean of the Radiology Department,Medical Sciences and Technology University, Khartoum, Sudan.

The authors have no conflicts of interest to disclose.

Copyright © 2019 Wolters Kluwer Health, Inc. All rights reserved.

DOI: 10.1097/JCE.0000000000000361

Feature Article

180 www.jcejournal.com Volume 44 • Number 4 • October/December 2019

Copyright © 2019 Wolters Kluwer Health, Inc. All rights reserved.

RNFL defects by using fundus photographs depends onthe experience of the examiner, and early defects may bemissed because of the low contrast of the RNFL. There-fore, they developed a simple and efficient algorithm to as-sist the ophthalmologist in the detection of RNFL defects.The strength of the proposed algorithm is that it can accu-rately differentiate very narrow defects in early-stage glau-coma to nonglaucomatous optic neuropathy involving thepapilla macular bundle. Thus, no previous studies have de-scribed specific methods for detecting RNFL defects with vari-ous forms and widths in fundus images. Their results showedthat the proposed algorithm was successful, with a sensitivityof 90% for glaucoma and 100% for papilla macular bundledefects in nonglaucomatous optic neuropathies.

Odstrcilika et al5 introduced a novel approach to cap-ture these variations using computer-aided analysis of theRNFL textural appearance in standard and easily availablecolor fundus images. The proposed algorithm was builtbased on the Gaussian Markov random fields and localbinary patterns features, together with various regressionmodels for prediction of the RNFL thickness. The algo-rithm describes the changes in RNFL texture, by reflectingvariations in the RNFL thickness. The method was testedon 16 healthy and 8 glaucomatous eyes. The results achievedsignificant correlation (normal = 0.72 ± 0.14; P = 0.05,glaucomatous = 0.58 ± 0.10; P = 0.05) between the resultsof the predicted output and the RNFL thickness measuredby OCT, which is the standard glaucoma assessmentdevice. The evaluation achieved good results to measurepossible RNFL thinning.

Therefore, those studies provide an expert system for real-time fundus image analysis, which gives an opportunity toimprove the image interpretation. The proposed system willimprove the diagnosis of glaucoma with another method,and therefore, it will minimize the missed detection rateand help in early diagnosis and treatment, which can signif-icantly improve the chance of managing glaucoma disease.

Materials and MethodsData were acquired from an open Retinal Image Databasefor Optic Nerve Evaluation (RIM-ONE), which contains

169ONH images, where the images are divided as follows:normal, 118 images; early glaucoma, 12 images; moderateglaucoma, 14 images; deep glaucoma, 14 images; and ocu-lar hypertension, 11 images. In this study, the RIM-ONE

FIGURE 1. An example of RIM-ONE database images: (A) normalimage, (B) glaucoma image (abnormal).

FIGURE 2. Proposed methodology.

Feature Article

Journal of Clinical Engineering www.jcejournal.com 181

Copyright © 2019 Wolters Kluwer Health, Inc. All rights reserved.

database is used to test the algorithm validation by Fumeroet al.6 Figure 1 shows an example of normal and abnormalglaucomatous images taken from the RIM-ONE database.

The proposed algorithm provides an automated glau-coma detection computer-aided system used for early di-agnosis of glaucoma with high accuracy. The algorithmtakes a preprocessed fundus image and segments the opticdisc followed by RNFL part extraction and extractingtexture features and classifying them. The results are usedto classify the images as glaucomaandhealthy cases. Figure 2illustrates the complete methodology.

Preprocessing StepImage preprocessing steps are as follows: resizing the im-age to 256� 256 pixels so that it has the specified numberof rows and column to reduce computational time anddenoising stage by median filters (5 � 5) by Elseid et al.7

Before any procedures are made, red channel has been ex-tracted because it appears as a good boundary andhas fewer blood vessels, which affects the segmentationaccuracy as to facilitate intensity analysis, as shown inFigure 3.

Optic Disc SegmentationThe optic disc segmentation algorithm was developed andtested in the RIM-ONE public database. The proposedmethod by Ahmed et al8 can be divided into 4 steps. Thefirst step is vessel removal to obtain accurate segmenta-tion, and it was done by morphological operation. Thesecond step is applying thresholding level of 180 to seg-ment disc region from the red channel, which appears tobest contrast the disc region; after that, in the third step,the boundary was smoothed and cleared from uncon-nected object, and binary image obtained at the final stepcircle image was constructed based on the radius and cen-ter of the detected region to minimize segmentation errors

resulting from the main blood vessels and PPA surround-ing the optic disc, using the formula and the result shownin Figure 4:

x−hð Þ2þ y−kð Þ2 ¼ r2 ð1Þ

where r = the radius from segmented object, and h, k = thecenter from segmented.

RNFL Region-of-Interest Extraction StepTheRNFL is the area surrounding the disc for the region of in-terest determined based on mathematical method to subtractthe whole image from the disc area, by the formula:

RNFL ¼ ROI image−OD ð2Þ

where the ROI image represents the ONH part, andOD isthe optic disc after circle reconstruction step. The resultsare shown in Figure 5.

The OD and RNFL are used for the third step, which isfeature extraction and selection, and the final step is theclassification.

Texture feature Extraction StepMany methods can be used to describe the main featuresof the textures such as directionality, smoothness, coarse-ness, and regularity. Gray-level co-occurrence matricesmeasure is one of the most important measures that can

FIGURE 3. The steps of an image preprocessing.

FIGURE 4. Fundus image optic disc segmentation steps.

Feature Article

182 www.jcejournal.com Volume 44 • Number 4 • October/December 2019

Copyright © 2019 Wolters Kluwer Health, Inc. All rights reserved.

be used to describe the texture. In this research, 2 tech-niques are used to describe the RNFL.

GLCMmethod: A method used to calculate spatial re-lationship of pixels is the gray level, also known as thegray-level spatial dependence matrix, by characterizing thetexture of an image by calculating how often pairs of pixelwith specific values and in a specified spatial relationshipoccur in an image, creating a GLCM, and then extractingstatistical measures, which are autocorrelation, contrast,correlation, cluster prominence, homogeneity, cluster shade,difference variance, dissimilarity, energy, entropy,maximumprobability, sum of squares, sum average, sum variance,sum entropy, difference entropy, information measure ofcorrelation, inverse difference, inverse difference normal-ized, and inverse difference moment normalized. Beloware examples of GLCM features equations.

Tamara method: A method used to calculate coarse-ness, contrast, and directionality features for digital fundusimage:

• Coarseness: The most fundamental texture fea-ture, which is a direct relationship to scale and rep-etition rates. It aims to identify the largest size atwhich a texture exists; even a smaller microtextureexists.

Ak x; yð Þ ¼ ∑xþ2k−1−1i¼x−2k−1−1

∑yþ2k−1−1

j¼y 2k−1f i; jð Þ=22k ð4Þ

• Contrast is a statistical distribution of the pixel in-tensity obtained.

Fcon ¼ σ

α1=44

ð5Þ

• Direction:Degrees we need to calculate the directionof the gradient vector are calculated at each pixel.

Fdir ¼ ∑nPp ∑ ;2ωρ θ−θρð Þ2 HD ;ð Þ ð6Þ

Feature Selection StepFeature selection is the process of selecting a subset of rel-evant features for use in model construction.

The objective of this step is 3-fold: improving the predic-tion performance of the predictors, providing faster andmorecost-effective predictors, and providing a better understand-ing of the underlying process that generated the data.

In this article, sequential feature selection method wasapplied to the 25 texture features and results in 1 feature,which was used to classify the fundus images to glaucomaand healthy cases.

Classification StepMany classification techniques are used such as support vec-tor machine (SVM), k-nearest neighbor (KNN), and ensem-bles, which use multiple learning algorithms to obtain betterpredictive performance than could be obtained from anyof the constituent learning algorithms alone. To get moremodel accuracy, ensemble learning is used to improve ma-chine learning results by combining several models thatensembles RUS Boost, an excellent technique for learningfrom imbalanced data (1 class outnumbers other classes bya large proportion), where synthetic minority oversamplingtechnique (SMOTE) algorithm is used to create artificialdata based on feature space (rather than data space) simi-larities from minority samples. Therefore, it generates arandom set of minority class observations to shift the clas-sifier learning bias toward minority class, to generate arti-ficial data, using bootstrapping and the same k-nearesttechnique used before.9-11

Results and DiscussionThe performance metrics used for the proposed glaucomadetection method evaluation are sensitivity, specificity,

FIGURE 5. Retinal nerve fiber layer for fundus image, which is a partsurrounding the disc.

Energy Feature Entropy Feature

Energy ¼ PN−1

i;j¼0pij� �2

Entropy ¼ PN−1

i;j¼0−In pij

� �pij� �

Contrast Feature Homogeneity Feature

Contrast ¼ PN−1

i;j¼0pij i−jð Þ2 Homogeneity ¼ PN−1

i;j¼0

pij1 i−j2ð Þ

Correlation Feature Shade feature

Correlation ¼ PN−1

i;j¼0pij

i−μð Þ j−μð Þσ2

Shade ¼ sgn Að ÞA13=

Prominence Feature

Prominence ¼ sgn Bð ÞB14=

(3)

Feature Article

Journal of Clinical Engineering www.jcejournal.com 183

Copyright © 2019 Wolters Kluwer Health, Inc. All rights reserved.

and accuracy, and the receiver operating characteristiccurve. Those values are defined as follows:

Sensitivity ¼ TP= TPþ FNð Þ ð7Þ

Specificity ¼ TN= FPþ TNð Þ ð8Þ

Accuracy ¼ TPþ TN= TPþ TNþ FPþ FNð Þ ð9Þ

Sensitivity (SE) is the ratio of glaucoma images thatweremarked and classified as glaucoma, to all marked images:

SE = TP/(TP + FN), from this equation it calculated tobe 0.83.

Specificity (SP) is the ratio of glaucoma images thatweremarked and classified as healthy, to all marked images:

SP = TN/(FP + TN) = TNR; from this equation, it wascalculated to be 0.92.

From Figure 6, we can notice that:

• True positive (TP) is the number images detected asglaucoma by an expert and the proposed method.

• True negative (TN) is the number of images detectedas normal by an expert and the proposed method.

• False positive (FP) is the number of images detectedas normal by an expert but detected as glaucomaby the proposed method.

• False negative (FN) is the number of images detectedas glaucoma by an expert but detected as normal bythe proposed method.

The values of sensitivity, specificity, and accuracy liebetween 0 and 1. Therefore, if the result of the proposedmethod is accurate, it should be close to 1.

The receiver operating characteristic curve shows TPrate versus FR rate for the currently selected trained classi-fier. The marker on the plot shows the performance ofthe currently selected classifier.

The marker shows the values of the FP rate and the TPrate for the currently selected classifier.

The selected feature is the RNFL coarseness. This fea-ture is evaluated by many classifiers such as SVM, KNN,and ensembles bagging classifier and ensembles boostingclassifier; the classification result in Figure 6 shows the

best result using ensembles RUSBoosted, which has accu-racy of 89.5% and area under the curve of 0.93.

The texture features classification errors equal 10.5 ex-plained as follows: in advanced glaucoma or optic atro-phy, RNFL defects cannot be detected, because the meanintensity of the RNFL is low in all directions, and a local-ized lesion is not distinguishable in this effect in the RNFLsegmentation. However, in these cases, the pathologicfeatures of disc cupping or atrophy are clearer than RNFLdefects and can easily be detected, compared with MorrisandMohammed,12 who used the Binary Robust Indepen-dent Elementary Features as a texture feature to detect theglaucoma and achieved an area under the curve of 0.84,and Ali et al,13 who detected glaucoma by local texturefeatures and achieved 95.1% success rate with a specific-ity of 92.3% and a sensitivity of 96.4%, which are betterthan texture features presented in this research becauseof 2 reasons. The first one is that the database used issmall and the texture features extracted from the wholeimage, therefore the suggested method is better inglaucoma detection.

The MATLAB codes used to extract the Tamara tex-ture features are available for free in MATLAB file ex-change side coded by Sornapudi14 and the GLCM featurecoded by Uppuluri15 (Table).

Conclusion and Future WorkThis article proposed a glaucoma detection algorithm basedon the analysis of digital fundus images using RNFL texturefeature (coarseness) classified by ensemble RUSBoostedtree classifier. The proposed method achieved an accuracyof 89.5%. The key contribution in this work is that theproposed features are suitable for glaucoma detectionwith high accuracy.

It is recommended that future work design a complete,integrated, automated system to classify all different typesof glaucoma that can be used for follow-up and test differ-ent types of features to improve the accuracy and test dif-ferent classifiers. Different segmentation techniques can beapplied to enhance the RNFL segmentation.

References1. Lim TC, Chattopadhyay S, Acharya UR. A survey and comparative

study on the instruments for glaucoma detection. J Biomed Eng Technol.2012;34(2):129-139.

FIGURE6. A,Confusionmatrix. B, Receiver operating characteristiccurve results from the texture selected feature.

TABLE. Final Proposed System Evaluation ParametersValuesSensitivity Specificity Accuracy Area Under the Curve

83% 92% 89.5% 93%

Feature Article

184 www.jcejournal.com Volume 44 • Number 4 • October/December 2019

Copyright © 2019 Wolters Kluwer Health, Inc. All rights reserved.

2. Anindita S, Agus H. Automatic glaucoma detection based on the type offeatures used: a review. J Theor Appl Inf Technol. 2015;72(3): ISSN:1992-8645.

3. Kavya N, Padmaja KV. Glaucoma detection using texture featuresextraction. In: 2017 51st Asilomar Conference on Signals, Systems, andComputers. Vol. 2017. Pacific Grove, CA: IEEE;2017:1471-1475.

4. Oh JE, Yang HK, Kim KG, Hwang J-M. Automatic computer-aideddiagnosis of retinal nerve fiber layer defects using fundus photo-graphs in optic neuropathy. Invest Ophthalmol Vis Sci. 2015;56:2872-2879.

5. Odstrcilika J, Kolar R, TornowRP, et al. Thickness related textural prop-erties of retinal nerve fiber layer in color fundus images.ComputMed Im-aging Graph. 2014;38:508-516.

6. Fumero F, Alayon1 S, Sanchez JL. (2011). RIM-ONE: an open retinalimage database for optic nerve evaluation. In: Proceedings of the IEEESymposium on Computer-Based Medical System.

7. Elseid AAG, Elmanna ME, Hamza AO. Evaluation of spatial filteringtechniques in retinal fundus images. Am J Artif Intell. 2018;2(2): 16-21.

8. Ahmed A, OsmanA.Optic disc segmentation using manual thresholdingtechnique. J Clin Eng. 2019;44(1):28-34.

9. Kumar Arun MN, Sheshadri HS. Building accurate classifier for theclassification of micro calcification. Int J Comput Sci Inform Technol.2012;3(6):5346-5350.

10. Alghamdi M, Al-Mallah M, Keteyian S, Brawner C, Ehrman J, Sakr S.Predicting diabetes mellitus using SMOTE and ensemble machine learn-ing approach: the Henry Ford Exercise Testing (FIT) project. PLoSOne.2017;12(7):e0179805.

11. Sui Y,Wei Y, ZhaoD. Computer-aided lung nodule recognition by SVMclassifier based on combination of random undersampling and SMOTE.Comput Math Methods Med. 2015;2015:368674.

12. Morris T,Mohammed S. (2015). Characterizing glaucoma using texture,The University of Manchester, School of Computer Science.

13. Ali MA, Hurtut T, Faucon T, Cheriet F. Glaucoma detection based onlocal binary patterns in fundus photographs. SPIE Med Imaging.2014;903531-903531-7. doi:10.1117/12.2043098.

14. Sornapudi S. Tamura Features MATLAB. 2014; https://github.com/Sdhir/TamuraFeatures. Accessed January 2018.

15. Uppuluri A. GLCM MATLAB Code. 2008; https://www.mathworks.com/matlabcentral/fileexchange/22187-glcm-texture-features. Copyright(c) 2008, All rights reserved.

Feature Article

Journal of Clinical Engineering www.jcejournal.com 185

Copyright © 2019 Wolters Kluwer Health, Inc. All rights reserved.