WINTER 2017-18B.Tech (I.
T)Digital Image ProcessingITE1010Slot – B2+TB2Diagnosis of Glaucoma on RetinalFundus Images using Artificial NeuralNetworksReview – 1Group Members:Umang Chaudhary 15BIT0074Apoorv Kshirsagar 15BIT0342Abstract:Glaucoma is the second leading cause of blindness in the world. It is estimated that 66.8 millionpeople in the world have glaucoma, with 6.
7 million bilaterally blind from this disease. One ofthe glaucoma symptoms is that the cup enlarges until it occupies most of the disc area. Thisresults in blindness. The cup-to-disc ratio is a measurement used in ophthalmology to detectglaucoma.
An automatic calculation of cup to disc ratio using accurate detection of disc and cup,and quantitative determination of their areas in retinal fundus images to diagnose glaucoma willbe carried out using Image Processing Techniques. This project is targeted to automate thediagnosis of Glaucoma and to provide an efficient way to detect it using ANN.Objective:Objective of our research based project is to -1) detect the glaucoma in one person’s eye by using it’s eyes fundus images with the help ofimage processing and convolutional neural network.Motivation:As we know, Glaucoma is the second leading cause of blindness in the world. It is estimated that66.8 million people in the world have glaucoma, with 6.7 million bilaterally blind from thisdisease.
One of the glaucoma symptoms is that the cup enlarges until it occupies most of the discarea. This results in blindness. The cup-to-disc ratio is a measurement used in ophthalmology todetect glaucoma which is cumbersome and it may not be correct. This led to the proposedmethodology for detection of glaucoma.Literature Survey:Several studies are reported in literature for Glaucoma detection and classification of glaucoma.The work is as follows:In Year 2007, Sangyeol Lee performed a work, “Validation of Retinal Image RegistrationAlgorithms by a Projective Imaging Distortion Model”2. A variety of methods for retinalimage registration have been proposed. Authors also present the validation tool for any retinalimage registration method by tracing back the distortion path and accessing the geometricmisalignment from the coordinate system of reference standard.
In Year 2008, S. Sekhar performed a work,” Automated localization of retinal optic disk usinghough transform”3. The retinal fundus image is widely used in the diagnosis and treatment ofvarious eye diseases such as diabetic retinopathy and glaucoma. The proposed methodologyconsists of two steps: in the first step, region of interest (ROI) is found by image by means ofmorphological processing, and in the second step, optic disk is detected using the Houghtransform.
In Year 2010, Zhuo Zhang performed a work,” ORIGA-light : An Online Retinal Fundus ImageDatabase for Glaucoma Analysis and Research”4. Author present an online dataset,ORIGA-light, which aims to share clinical retinal images with the public. Author had updatedthe system continuously with more clinical ground-truth images.
The proposed method focuseson optic disk and cup segmentation.In Year 2010, Vahabi Z proposed,” The new approach to Automatic detection of Optic Disc fromnon-dilated retinal images”5. Author describes a new filtering approach like Sobel edgedetection, Texture Analysis, Intensity and Template matching to detect Optic Disc. The proposedalgorithm is applied in wavelet domain on 150 images of Messidor dataset.In Year 2011, Zafer Yavuz performed a work,” Retinal Blood Vessel Segmentation Using GaborFilter And Tophat Transform”6. In this, Author gave a method for retinal blood vesselssegmentation by applying firstly Gabor filter to enhance blood vessels and then applying top-hattransform.
Later on, the output is converted to binary image with p-tile thresholding.In Year 2012, Nilan jan Dey performed a work,” Optical Cup to Disc Ratio Measurement forGlaucoma Diagnosis Using Harris Corner”7. In this paper, CDR is determined using HarrisCorner.
Harris comer detector 8,9 measures the local changes of the signal with patches shiftedin different directions by a small amount. It is based on the local auto-correlation function of asignal.In Year 2012, R. Geetha Ramani performed a work,” Automatic Prediction of DiabeticRetinopathy and Glaucoma through Retinal Image Analysis and Data Mining Techniques”10.This paper proposed a novel approach for automatic disease detection. Retinal image analysisand data mining techniques are used to accurately categorize the retinal images as either Normal,Diabetic Retinopathy and Glaucoma affected.In Year 2012, ManjulaSri Rayudu proposed,” Review of Image Processing Techniques forAutomatic Detection of Eye Diseases”11.
The review paper provides information about theapplication of image processing techniques for automatic detection of eye diseases. The keyimage processing techniques to detect eye diseases include image registration, fusion,segmentation, feature extraction, enhancement, pattern matching, image classification, analysisand statistical measurements.In Year 2013, Preeti and Jyothika Pruthi performed a work, ‘Review of Image ProcessingTechnique for Glaucoma Detection’12 and concluded that for detection and diagnosis ofglaucoma, firstly, optic disk need to be segmented. After image acquisition, preprocessing isdone by applying thresholding, illumination and histogram equalization. The optic disk and cupis segmented using various techniques like Hough transform, k-means clustering, fuzzy c-meansclustering, active contour method, matched filter approach, vessel bends, morphologicaloperations etc. Then CDR is calculated and classification is done for deciding whether conditionof eye is normal or glaucomatous.Detection of Hard Exudates And Red Lesions In The Macula Using A Multiscale Approach 13In this paper 13, they present an automatic system to detect pathologies on the sunspot likehard exudates microaneurysm, and hemorrhages.
They use the bottom up approach in which theytries to capture each unnatural structure in the macula which is used to observe DR lesions. Thistechnique starts by extinguish the non-uniform illumination thereby raising the contrast of redlesions in the images. Possible DR lesion candidates on the macula are draw out by usingamplitude-modulation frequency modulation (AM-FM) features. AM-FM features are used forextract texture information from different frequency scales, providing for an effectual method forthe detection of hard exudates and red lesions.
In this paper 14, it concentrates only in the automatic detection of one of the lesions associatedwith DR: hard exudates. They normally appear in the fundus photographs as small yellow whitesports with crisp margins and A Literature Survey on Glaucoma Detection Techniques usingFundus Images (IJSRD/Vol. 2/Issue 09/2014/173) All rights reserved by www.ijsrd.com 745different shapes. Among DR lesions, exudates are frequently occurring early lesions. Anautomatic method to detect hard exudates, a lesion related with diabetic retinopathy.
Thealgorithm are found by using a statistical classification. The blood vessels are metamericapplying the matched filter method which is described in to raise blood vessels and threshold theimage obtained.In this paper 15, the precise and automated characterization of anatomic structures from imagedata sequences is one of the lasting issues and there is a faster increase the interest in the researchcommunity to research on Automatic Detection of Hard and Soft Exudates in Fundus ImagesUsing Color Histogram Threshold. Diabetic retinopathy is considered as the root cause of visionloss for diabetic patients.In this paper 16, they used a fully automated fast method to observe the fovea and the opticdisc in digital color pictures of the retina is presented.
In this method they make few premisesabout the location of each and every structure in that image. They specify that a retinal imageproblem of localizing structures is a regression problem. A KNN regressor is employed to callthe distance in pixels in the image to the object of involvement at any given location in the imagebased on a set of features measured at that location. The method unites cues measured instantlyin the image with remind derived from a segmentation of the retinal vasculature.In year 2002, Conor et al. has used the skeleton operations to determine the change in retinalanatomy for DR detection in abnormal images 17. The features used in this work are vesselwidth and tortuosity.
The experiment is analyzed in terms of accuracy.In year 2003, Agostino et al. has used the neural network methodology to detect thekeratoconous abnormal retinal images 18. Experimental analysis is performed in this workbased on sensitivity and specificity. An extensive quantitative analysis is also yielded in thisreport.In year 2005, Harihar et al. has used the Bayesian algorithm for abnormal retinal imageclassification 19.
Multi-level classification is performed in this work with five classes. Theconcept of Markov random field is also used in this automated system. A better qualityclassification accuracy results are reported in this method.In year 2006, Multifractal analysis of human retinal images is performed by Tatijana et al 20.
This approach is mainly used to detect the blood vessels which 29 further aid in differentiatingthe different abnormal images. Texture based techniques and model based techniques areanalyzed in detail in this report. Lack of quantitative analysis is the major drawback of thisautomated system.Summary of Literature Survey:A survey of techniques for the automatic detection of glaucoma and diabetic retinopathy hasbeen presented in this chapter.
The proposed method involves two phases (i) detection ofglaucoma and (ii) detection of diabetic retinopathy. For the detection of glaucoma, a novelapproach is proposed initially for optic disc boundary detection to solve the problem of bloodvessel occlusion and detect the optic disc even when the boundary of the disc is not continuousor blurred. As the color intensity of the optic cup cannot be fixed, a method for optic cupdetection is proposed using the difference in pallor to estimate the cup-disc boundary. Currently,the CDR evaluation is manually performed and it is subjected to individual evaluation byophthalmologists.
Further dependence on manual grading limits its potential use for use in themass screening of populations for early glaucoma detection. Also, CDR does not take intoaccount the disc size and stages of the disease cannot be identified. Further, the blood collectsalong the individual nerve fiber that radiate outwards from the nerve. Such physiological changesare manifested in the fundus images and the texture features are used to quantify such differencein eye physiology. Therefore, structural features and textural features are to be combined forbetter analysis of the images as normal or abnormal. So a method is proposed to detect the earlystage of glaucoma using structural and textural features. The approach used can help clinicians inseveral eye care applications such as diagnosis, screening and monitoring.
An integrated systemis to be developed in the second phase for the detection of diabetic retinopathy, to enhance theperformance of the system by i) improving the results of other tasks such as the detection ofblood vessels, optic disc, localization of faint and small exudates using color and texturalfeatures and ii) proper selection of features to improve the performance of 59 classificationtechniques. An effective tool should therefore be developed to analyze fundus images to detectfeatures such as exudates comparable to that of an ophthalmologist in order to provide decisionsupport and reduce ophthalmologist’s workload.Proposed Architecture:Figure 1: Image Processing Steps Architecture.Figure 2: Work flow of Preprocessed images into CNN to detect Healthy or Glaucoma affected imagesModule Description:In this project we have proposed using Convolutional Neural Network for predicting glaucoma.The basic steps are1. Take retinal fundus images of the patient2. Using image processing crop the optic disk3.
Feed the cropped image in CNN for prediction4. Compare and train the CNN for improving accuracyProcedure for Image processing:i. Split in three channels r,g,bii. Take greeniii. Smoothen using medianiv. Detect edgesv. Subtract g – edgesvi.
Smoothen result again for removing noises median filtervii. Dilate the imageviii. Threshold the imageix.
Use threshold to find contourx. Limit the contour length from 1 – 2xi. Use contour to crop ROIxii.
Resize image to prepare for cnn 400*400xiii. Feed imagesxiv. Fetch the resultReferences :1 Kevin Noronha, Jagadish Nayak, S.N. Bhat, “Enhancement of retinal fundus Image tohighlight the features for detection of abnormal eyes”2 Sangyeol Lee, Michael D. Abr`amoff, and Joseph M.
Reinhardt.” Validation of RetinalImage Registration Algorithms by a Projective Imaging Distortion Model” 29th AnnualInternational Conference of the IEEE EMBS Cité Internationale, Lyon, France August 23-26,20073 S. Sekhar,” Automated localisation of retinal optic disk using hough transform”, Departmentof Electrical Engineering and Electronics, University of Liverpool, UK4 , Zhuo Zhang,” ORIGA-light : An Online Retinal Fundus Image Database for GlaucomaAnalysis and Research”, 32nd Annual International Conference of the IEEE EMBSBuenosAires, Argentina, August 31 – September 4, 20105 Vahabi Z,” The new approach to Automatic detection of Optic Disc from non-dilated retinalimages” Proceedings of the 17th Iranian Conference of Biomedical Engineering (ICBME2010),3-4 November 20106 Zafer Yavuz,” RETINAL BLOOD VESSEL SEGMENTATION USING GABOR FILTERAND TOPHAT TRANSFORM”, 2011 IEEE 19th Signal Processing and CommunicationsApplications Conference (SIU 2011) 978-1-4577-0463-511/11 ©2011 IEEE7 Nilan jan Dey,” Optical Cup to Disc Ratio Measurement for Glaucoma Diagnosis UsingHarris Corner”, ICCCNT128Harris C, Stephens, M. , 1 988, A Combined Corner and Edge Detector, Prooeedings of 4thAlvey Vision Conference9 Kon stantino s G Derpanis, 200 4, The Harris Corner Detector10 R. Geetha Ramani,” Automatic Prediction of Diabetic Retinopathy and Glaucoma throughRetinal Image Analysis and Data Mining Techniques”11 ManjulaSri Rayudu,” Review of Image Processing Techniques for Automatic Detection ofEye Diseases”, 2012 Sixth International Conference on Sensing Technology (ICST)978-1-4673-2248-5/12 ©2012 IEEE12 Preeti. and Pruthi J. (2013). Review of Image Processing Technique for GlaucomaDetection.
International Journal of Computer Science and Mobile Computing onlineIjcsmc.com. Accessed 11 Jan.
2018.13 Carla Agurto, Honggang Yu, Victor Murray, Marios S. Pattichis, Simon Barriga, PeterSoliz” Detection Of Hard Exudates And Red Lesions In The Macula Using A MultiscaleApproach” Doi: 10.1109/Ssiai.2012.6202441 Conference: Image Analysis and Interpretation(SSIAI), 201214 Sánchez CI, Hornero R, López Mi, Poza J.” Retinal Image Analysis To Detect AndQuantify Lesions Associated With Diabetic Retinopathy” Ieee Eng Med Biol Soc. 2004;3:1624-7.
15 S. Kavitha, K. Duraiswamy” Automatic Detection Of Hard And Soft Exudates In FundusImages Using Color Histogram Thresholding”10.2316/P.2012.
778-04416 Meindert Niemeijer, Michael D. Abràmoff, and Bram van Ginnekena” FAST DETECTIONOF The Optic Disc And Fovea In Color Fundus Photographs” Med Image Anal. Dec 2009;13(6): 859–870.17. Conor Heneghan, John Flynn, Michael O Keefe, Mark Cahill (2002), ‘Characterization ofchanges in blood vessel width and tortuosity in retinopathy of prematuriy using image analysis’,Medical Image Analysis, Vol.
6, pp. 407- 429.18.
Agostino Accardo P. and Stefano Pensiero (2003), ‘Neural network based system for earlykeratoconus detection from corneal topography’, Journal of Biomedical Informatics, Vol. 35,pp. 151-159.19 H.
Narasimha-Iyer et al ., “Robust detection and classification of longitudinal changes incolor retinal fundus images for monitoring diabetic retinopathy,” in IEEE Transactions onBiomedical Engineering , vol. 53, no. 6, pp.
1084-1098, June 2006.20 T. Stosic and B. D.
Stosic, “Multifractal analysis of human retinal vessels,” in IEEETransactions on Medical Imaging , vol. 25, no. 8, pp. 1101-1107, Aug. 2006.