an analytical method for the detection of exudates in retinal … · 2016-07-21 · fig. 3. cake...

An Analytical Method for the Detection ofExudates in Retinal Images Using Invertible

Orientation ScoresSurya Rajan, Taraprasad Das, R. Krishnakumar

Abstract—Diabetic retinopathy is an eye related complicationof diabetes mellitus caused due to damage to the retinal bloodvessels, resulting in micro-aneurysms, hemorrhages and exudates.DR is asymptomatic, necessitating the development of an auto-mated screening system. Exudates appear in the later stages ofdiabetic retinopathy and their presence would classify the diseaseas moderate or severe. In developing countries, where access totraining data is limited, there is a greater need for analyticalmethods than machine learning techniques. The proposed methoduses the orientation scores of the retinal image to detect exudates.The 2D orientation score framework, proposed by Duits et al.,inspired by the visual system of mammals, is a mapping whichassigns the position and orientation angle of each pixel to acomplex scalar and has been so far used to detect vasculature treeon the retina. This paper proposes the use of orientation scoresto form an orientation enhanced image, from which a binarymask of exudates can be obtained by intensity thresholding. Itachieves a sensitivity of 86.2% and a specificity of 85% on imagesof DIARETDB1 database.

Index Terms—Diabetic Retinopathy, Exudates, OrientationScores, Cake Wavelets

I. INTRODUCTION

Diabetic retinopathy is increasingly becoming the commoncause of blindness among the middle-aged population in Indiaand the rest of the world. In a study conducted in [1], by 2050,there will be 244 million people (14.9% of the population)with diabetic retinopathy, compared with 42 million (4.5% ofthe population) in 1995. This indicates that retinal diseaseswould soon become the major cause of blindness in India.But, with limited number of specialist doctors, an automatedscreening system is essential to ensure an early diagnosis of thedisease. The main symptoms of diabetic retinopathy are micro-aneurysms, exudates and hemorrhages. Micro-aneurysms oc-cur due to weakening and bulging of blood vessel walls ofthe retina. They are typically round lesions of the size of onlya few pixels. Hemorrhages form when the walls of a vesselruptures. When the leakage from a vessel also contains lipidsand proteins, it creates yellow spots on the retina known asexudates. They have well-defined edges, no consistent shapeand their sizes vary widely. There are also lesions called cottonwool spots or soft exudates, which do not have sharp edges.An example of a retinal image containing exudates is given inFig. 1.

Surya Rajan and R. Krishnakumar are with the Department of EngineeringDesign, Indian Institute of Technology Madras, India

Dr. Taraprasad Das is with L.V. Prasad Eye Institute, Hyderabad, IndiaCorresponding author: [email protected]

Fig. 1. An image from DIARETDB1, with portions of exudates marked asshown

Exudate detection in the retina is typically split into twosteps: pre-processing of the images, and segmentation of thelesions. A noise mask was generated in [2] for pre-processingto exclude noisy regions and background subtraction was donein [3], [4], [5]. In-painting to remove dark structures andanalysis using adaptive templates to remove bright structureswere done in [6]. Contrast of the image was enhanced in[7] using a combination of complex shock filter and globalhistogram stretching.

The segmentation of exudates may be achieved by machinelearning techniques, or analytical methods. Machine learningtechniques give good results, if one has a large training set.But, in a country like India, where access to images arelimited, and different camera setups are used to capture retinalimages, it is not practical to rely only on machine learningtechniques for lesion detection. Therefore, there is a needto develop better analytical techniques. Morphological oper-ations were carried out in [8] in order to generate candidateregions for exudates and optic disc, and then the OD wasremoved from the candidates to obtain the bright lesions.Morphological operations were used also in [9], [10], [11]to segment exudates. Several inverse segmentation methodswere proposed in [12] such as region growing, region growingwith background correction and adaptive region growing withbackground correction.

Many techniques have been used for a coarse extractionstage, including intensity thresholding [5], [13] and morpho-logical techniques [6], [14], [15], [16]. In the latter case, themorphological operators were used to obtain a more accuratelocalization of exudates. The classification of candidates wasdone analyzing their edge strength in [5] by applying kirschanalysis and setting a threshold. In [17], a new approach using

Proceedings of the World Congress on Engineering 2016 Vol I WCE 2016, June 29 - July 1, 2016, London, U.K.

ISBN: 978-988-19253-0-5 ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)

WCE 2016

ant colony optimization was tested, with the output informa-tion inside a pheromone matrix, which changes according tolocal image intensity variations. Machine learning techniqueshave also been used in the classification stage [2], [3], [4], [6]by analyzing various intensity, shape and size features.

In resource deficient areas, one has to depend on onlineretinal datasets to train a model.

Such a machine learning network trained on publicly avail-able datasets fails to work when tested on images obtainedfrom hospitals. A model based on Random Forests [18],has been trained on the images from the online databaseDIARETDB1, and was tested on images obtained from L.V. Prasad Eye Institute, Hyderabad. This model fails to detectexudates accurately on the images from the hospital. Thus,the need for more analytical techniques is clear and is themotivation for this paper.

This work is an attempt towards fulfilling the need for moreanalytical techniques for lesion detection in the retina. Thispaper proposes the use of orientation scores to detect exudates,using the fact that they are either isotropic or moderatelyelongated structures. Though orientation score domain haspreviously been used to analyze blood vessels of the retina[19], it has never been used to detect lesions. In this work ananalytical method has been developed for detection of exudatesusing orientation scores framework. Hard exudates are goodcandidates for such a technique due to the presence of well-defined edges, and therefore, prominent orientations.

II. ORIENTATION SCORE DOMAIN

The mathematical framework for the orientation scores wasdeveloped in [20], inspired by the cortical orientation columnsin the visual system of mammals, where a simple cell receptivefield can be parameterized by its position and orientation.Perceptual organization can be similarly done on an image,on the basis of orientation similarity, as long as a linearwell-posed invertible transformation exists. An orientationscore of an image f can be obtained by convolution with ananisotropic wavelet ψ [20]. Cake wavelets, proposed in [20],[21], [22],[23] have been used in this work to construct theorientation scores.

Uf (x, eiθ) =

∫R2

ψ(R−1θ (x′ − x))f(x

′)dx

′(1)

where Rθ is the rotation matrix, x = (x,y) is the positionand eiθ denote the local orientation. The domain of Uf isR2xS1 where S1 denotes the space of orientation and themapping f → Uf is a wavelet transformation. Thus, all theorientations in the spatial domain are effectively disentangledin the orientation score domain [19], [20]. While implementingthis method, the orientation scores and cake wavelets wereconstructed in the Fourier domain. The Fourier transform ofthe image was multiplied with each cake of the filter and theinverse Fourier transform of the product is taken to be thecorresponding orientation score.

III. ORIENTATION ENHANCED IMAGE

A. Pre-processingPre-processing of the image is done as laid out in [7]. First,

a complex shock filter is applied to the image to smooth out

random fluctuations, and then the background is estimated andcorrected. The complex shock filter smooths out the randomnoise fluctuations, without blurring the edges. To find thebackground image, an irregular grid is laid out, and the meanand standard deviation are found inside a window of size 125pixels [7]. These values are interpolated and the Mahalanobisdistance is calculated at each pixel. Those pixels with distanceless than 1 are assumed to belong to the background [7]. Theintensity is computed at each of the background pixels andinterpolated to get the background intensity image. Illumina-tion correction is performed, by subtracting this backgroundfrom the original image. Contrast of the image is enhancedin two steps edge enhancement and intensity separation.Edge enhancement is obtained by applying the complex shockfilter to the illumination corrected image. For increasing theintensity separation between regions, a global contrast stretchis used and the intensities are saturated at extreme 1intensities.The R, G, B planes of the original image are then remappedby a post-processing step. The results are shown in Fig. 2. Theoutput image has a Gaussian histogram in all three planes. Theblue plane of the pre-processed image is used in further stages,since only pixels with high intensities are visible in the bluechannel.

B. Orientation Enhanced Image

The orientation scores of an image, calculated using theequation in section II, are a column of complex values at eachpixel location. Fig. 3 shows the real and imaginary parts of theorientation score of a retinal image at θ = 3π/8.The real partof the orientation score gives information regarding orientedstructures, while the imaginary part contains edge information.An enhanced image can be formed using the real and theimaginary parts of the orientation scores, as explained below.The responses of the blue channel of the image to each of thecake filters, make up a conceptual stack of N0 = 16 layers;each layer containing the pixels oriented in that particular di-rection. Exudates are either isotropic or moderately elongatedstructures. When isotropic, pixel information is uniformlysmeared out over the different orientation layers. Therefore,at an exudate pixel, every orientation gives a large response,though lower than the original pixel value. Information aboutmoderately elongated exudate is available in one of the stackedlayers. To form the enhanced image for exudate detection,the maximum intensity through the layers for each pixel is

Fig. 2. (a) An image from DIARETDB1 database (b) Output image afterpre-processing



WCE 2016

Fig. 3. Cake filters applied on a retinal image with N0=16. (a) Pre-processed image. Results at θ = 3π/8 (b) Real part of the orientation score (c) Imaginarypart of the orientation score

Fig. 4. Enhanced images in case of exudate detection, formed from (a) real part and (b) imaginary part of the orientation scores, (c) Final enhanced image,by adding (a) and (b)

extracted. It should be noted that blood vessels which are lowintensity structures, give rise to negative orientation scores andare not confused with exudates, which are at the other endof the spectrum. The maximum intensity projection is donefor real as well as imaginary parts of the orientation scores.An enhanced image is obtained by adding these two images,so that both the oriented structure information and the edgeinformation are present in one image (Fig. 4).

IV. DETECTION AND ELIMINATION OF LANDMARKS

To ensure that other bright structures like optic disk andoptic nerve striations on the retina do not get falsely detectedas exudates, they are first detected and eliminated from theenhanced image.

A. Optic Disc

Optic disc is the brightest structure in a retinal image andtherefore, it needs to be eliminated from the candidate imagefor accurate exudate detection. A mask of the optic disc isgenerated by the method laid out in [24]. The algorithminvolves creating template orientation fields of the optic discfor the left and the right eye images and correlating themwith a normalized orientation field of the input image toobtain a response map. Depending on whichever templategenerates the maximum response, the retina image is classified

as belonging to the left or the right eye. Further processingis then performed on the response map, which has a localmaximum at the center of the optic disc, to identify the correctlocation of the optic disc in the image. The exact boundaryof the optic disc is found by using an area-conscious activecontour model [24]. In order for the blood vessels lying in thisregion, not to introduce discontinuities along the boundary ofthe OD, they are removed by thresholding the region near theidentified location at µc − σc, where µc and σc are the meanand standard deviation of all pixel values in it.

A sobel edge detection filter is applied to the vessel removedimage and a corresponding gradient magnitude image is calcu-lated. If the boundary of the optic disc is sufficiently strong andvessels have been successfully removed, an image is obtainedwith a relatively high value of gradient magnitude along theboundary of the optic disc and small gradients inside the ODregion. A potential surface P is obtained and the Derivativeof Gaussian (DoG) filters is applied in X and Y directions onP, to calculate a smooth force field. The gradient vector flow(GVF) method in [25] is used to set up a vector field basedon the image gradients, and a special outward force is addedto the contracting contour to ensure that it does not collapseto a point in images where the boundary of the OD is weakor absent. This resistance force is calculated as:

Fres =

(area enclosed by the evolving contour

area enclosed by the initial contour

)−k

n (2)



WCE 2016

Fig. 5. (a) Orientation enhanced image (b) OD Mask (c) Parabola Mask foroptic nerves (d) Orientation enhanced image after applying the masks

where k is set to 1.5 empirically and n is the outward normalvector to the evolving contour. The total external force is:

Fexternal = −ζFimage + ηFres (3)

where ζ is the image force weight factor and η is theresistance force weight factor. Fimage is the external forcefield derived from the image. The internal force and the energyfunctional are constructed as laid out in [24]. The contour isevolved so as to minimize the energy functional, using theEuler-Lagrange equation. Thus the optic disc is segmentedand is used as a mask in the detection of exudates (Fig. 5(b)).

B. Optic Nerves

Another important structure in the retinal image are theoptic nerves. They are bright striations, usually observed tooccur alongside the major blood vessels. They get picked upin the candidate image for exudates, because of their edges andbright nature. In order to eliminate these striations, the methodproposed in [6] is followed. After obtaining the vasculaturemask, those vessels which are less than 2/3rd the diameterof the largest vessel, are removed from the mask. This givesonly the major blood vessels, lying along a shape that mightbe approximated to a C- shaped parabola. Hence parabolasof varying curvatures, and dilated to 1/15th of the aperturediameter, are tested to see which among them contained thelargest number of blood vessel pixels [6]. This mask (Fig. 5(c))is applied to ensure that the optic nerve striations are removedfrom the orientation enhanced image.

V. BINARY MASK OF EXUDATES

From the orientation enhanced image, after applying themasks, a binary mask of exudates can be obtained by simpleintensity thresholding. The grayscale level 110 was foundto be suitable for images having pixels within the range 0-255. This threshold value gave good performance for imagesfrom DIARETDB1. Multi-thresholding using Otsu methodmay also be used to generate the binary mask. The resultsof the proposed method are discussed in VI.

VI. RESULTS AND DISCUSSIONS

The proposed algorithm was tested on the online databaseDIARETDB1, for the sake of evaluation. Receiver OperatingCharacteristics was used to determine the performance accu-racy of the method. The ROC curve obtained is shown in Fig. 6and the corresponding sensitivity, specificity and accuracy ofperformance are given in Table I. They show that the methodperforms better than other existing algorithms for exudatedetection that do not use any type of machine learning forclassification. The comparison was made against [8], [9], [10],[11] and not with the methods that use machine learning. Asurvey of literature gave only few analytical techniques, whichwere validated on DIARETDB1 and therefore a comparisonwas made against such methods only. From the images inFig. 7, it can be observed that the proposed technique workswell to detect lesions in images. Exudate detection shows goodperformance, irrespective of the size or shape of the lesion. Ifthe retinal image is very bright and has a low contrast, a fewfalse positives were detected, as shown in Fig. 7(e, f).

VII. CONCLUSION

A novel method to detect exudates on the retina was pro-posed. Recently, the field of machine learning has seen a rise inpopularity, and hence most papers propose techniques that usesa trained model to classify lesions.Machine learning generallygives good results, but the outcome is heavily dependenton the data that is used to train the model. In a resourcedeficient country like India, access to datasets of retinal imagesare limited, and one has to depend on online datasets andground truths for training models. But such a model fails toperform well on realworld images obtained from hospitals.Hence, it is not advisable to rely only on machine learning

Fig. 6. ROC curve: Performance of exudate detection for DIARETDB1

TABLE ICOMPARISON OF PERFORMANCE MEASURES OF EXUDATE DETECTION

METHODS

Method Sensitivity Specificity Accuracy

Proposed Method 86.2 % 85 % 85.39 %

Alharthi et al.[8] 86 % 80 % 84.6 %

Pourreza et al.[9] 78.28 % - -

Welfer et al. [10] 70.48 % 98.84 %

Omar et al. [11] 85.39 % - -



WCE 2016

Fig. 7. Results obtained for exudate detection on DIARETDB1. (a), (b) Size of exudates very small (c), (d) Moderately sized exudates (e), (f) Images withhigh illumination and low contrast, gives rise to a few false positives.

techniques to develop automatic screening mechanisms fordiabetic retinopathy.

The aim of this work is to develop an analytical methodfor the automatic segmentation of exudates on the retina. Themethod uses the orientation scores of the retinal image, fromwhich, an enhanced image is generated using the orientationcharacteristics of exudates. After further processing, a binaryimage or mask of the lesion is obtained. The proposed methodwas found to perform better than other analytical methods onthe DIARETDB1 dataset.

ACKNOWLEDGMENT

The authors would like to thank Hem Rampal and NikhilChandwadkar, Department of Engineering Design, Indian In-stitute of Technology Madras, for laying out the algorithmicframework for the detection of diabetic retinopathy. Theauthors also thank Erik Bekkers, Medical Image Analysisgroup, Eindhoven University of Technology, for his valuablesuggestions.

REFERENCES

[1] P. K. Nirmalan, J. Katz, A. L. Robin, J. M. Tielsch, P. Namperumalsamy,R. Kim, V. Narendran, R. Ramakrishnan, R. Krishnadas, and R. D.Thulasiraj, “Prevalence of vitreoretinal disorders in a rural populationof southernindia: The aravind comprehensive eye study,” Archives ofophthalmology, vol. 122, no. 4, pp. 581–586, 2004.

[2] U. M. Akram and S. A. Khan, “Automated detection of dark and brightlesions in retinal images for early detection of diabetic retinopathy,”Journal of medical systems, vol. 36, no. 5, pp. 3151–3162, 2012.

[3] M. U. Akram, A. Tariq, S. A. Khan, and M. Y. Javed, “Automateddetection of exudates and macula for grading of diabetic macularedema,” Computer methods and programs in biomedicine, vol. 114,no. 2, pp. 141–152, 2014.

[4] A. Tariq, M. U. Akram, A. Shaukat, and S. A. Khan, “Automateddetection and grading of diabetic maculopathy in digital retinal images,”Journal of digital imaging, vol. 26, no. 4, pp. 803–812, 2013.

[5] L. Giancardo, F. Meriaudeau, T. P. Karnowski, Y. Li, S. Garg, K. W.Tobin, and E. Chaum, “Exudate-based diabetic macular edema detectionin fundus images using publicly available datasets,” Medical ImageAnalysis, vol. 16, no. 1, pp. 216–226, 2012.

[6] X. Zhang, G. Thibault, E. Decenciere, B. Marcotegui, B. La, R. Danno,G. Cazuguel, G. Quellec, M. Lamard, and P. Massin, “Exudate detectionin color retinal images for mass screening of diabetic retinopathy,”Medical image analysis, vol. 18, no. 7, pp. 1026–1043, 2014.

[7] H. Rampal, R. K. Kumar, B. Ramanathan, and T. P. Das, “Complexshock filtering applied to retinal image enhancement,” in World Congresson Medical Physics and Biomedical Engineering May 26-31, 2012,Beijing, China. Springer, Conference Proceedings, pp. 900–903.

[8] A. S. A. Alharthi and V. Emamian, “An automated mechanism for earlyscreening and diagnosis of diabetic retinopathy in human retinal images.”

[9] H. Pourreza, “Localization of hard exudates in retinal fundus imageby mathematical morphology operations,” in Computer and KnowledgeEngineering (ICCKE), 2012 2nd International eConference on. IEEE,Conference Proceedings, pp. 185–189.

[10] D. Welfer, J. Scharcanski, and D. R. Marinho, “A coarse-to-fine strategyfor automatically detecting exudates in color eye fundus images,”Computerized Medical Imaging and Graphics, vol. 34, no. 3, pp. 228–235, 2010.

[11] M. Omar, A. Hossain, L. Zhang, and H. Shum, “An intelligent mobile-based automatic diagnostic system to identify retinal diseases usingmathematical morphological operations,” in Software, Knowledge, Infor-mation Management and Applications (SKIMA), 2014 8th InternationalConference on. IEEE, Conference Proceedings, pp. 1–5.

[12] C. Kose, U. evik, C. kiba, and H. Erdl, “Simple methods for segmenta-tion and measurement of diabetic retinopathy lesions in retinal fundusimages,” Computer methods and programs in biomedicine, vol. 107,no. 2, pp. 274–293, 2012.

[13] C. I. Sanchez, M. Garca, A. Mayo, M. I. Lpez, and R. Hornero, “Retinalimage analysis based on mixture models to detect hard exudates,”Medical Image Analysis, vol. 13, no. 4, pp. 650–658, 2009.

[14] T. Walter, J.-C. Klein, P. Massin, and A. Erginay, “A contribution ofimage processing to the diagnosis of diabetic retinopathy-detection ofexudates in color fundus images of the human retina,” Medical Imaging,IEEE Transactions on, vol. 21, no. 10, pp. 1236–1243, 2002.

[15] A. Sopharak, B. Uyyanonvara, S. Barman, and T. H. Williamson,“Automatic detection of diabetic retinopathy exudates from non-dilatedretinal images using mathematical morphology methods,” ComputerizedMedical Imaging and Graphics, vol. 32, no. 8, pp. 720–727, 2008.

[16] B. Harangi, B. Antal, and A. Hajdu, “Automatic exudate detection withimproved nave-bayes classifier,” in Computer-Based Medical Systems



WCE 2016

(CBMS), 2012 25th International Symposium on. IEEE, ConferenceProceedings, pp. 1–4.

[17] C. Pereira, L. Gonalves, and M. Ferreira, “Exudate segmentation infundus images using an ant colony optimization approach,” InformationSciences, vol. 296, pp. 14–24, 2015.

[18] A. C. Leo Breiman, “Random Forests,” https://www.stat.berkeley.edu/∼breiman/RandomForests/cc home.htm, [Accessed on 31 March 2016].

[19] E. Bekkers, R. Duits, T. Berendschot, and B. ter Haar Romeny, “Amulti-orientation analysis approach to retinal vessel tracking,” Journalof mathematical imaging and vision, vol. 49, no. 3, pp. 583–610, 2014.

[20] R. Duits, M. Duits, M. van Almsick, and B. ter Haar Romeny, “Invertibleorientation scores as an application of generalized wavelet theory,”Pattern Recognition and Image Analysis, vol. 17, no. 1, pp. 42–75, 2007.

[21] R. Duits, M. Felsberg, G. Granlund, and B. ter Haar Romeny, “Imageanalysis and reconstruction using a wavelet transform constructed froma reducible representation of the euclidean motion group,” InternationalJournal of Computer Vision, vol. 72, no. 1, pp. 79–102, 2007.

[22] H. Fuhr, “Abstract harmonic analysis of continuous wavelet transforms,”Thesis, 2005.

[23] R. Duits, “Perceptual organization in image analysis: A mathematicalapproach based on scale, orientation and curvature,” Thesis, 2005.

[24] N. Chandwadkar, “Critical feature extraction in colour fundus images,”Thesis, 2013.

[25] C. Xu and J. L. Prince, “Snakes, shapes, and gradient vector flow,” ImageProcessing, IEEE Transactions on, vol. 7, no. 3, pp. 359–369, 1998.



WCE 2016

an analytical method for the detection of exudates in retinal … · 2016-07-21 · fig. 3. cake...

Documents