AUTOMATIC DETECTION OF SEVERITY GRADING IN DIABETIC RETINOPATHY USING CONVOLUTIONAL NEURAL NETWORK

AUTOMATIC DETECTION OF SEVERITY GRADING IN DIABETIC RETINOPATHY USING CONVOLUTIONAL NEURAL NETWORK

Sharan. M 1, Rithik. S2, Lakshmi Priya. S 3

1Department of computer science and Engineering, Sathyabama Institute of Science and Technology, Tamil Nadu, Chennai

2Department of computer science and Engineering, Sathyabama Institute of Science and Technology, Tamil Nadu, Chennai

3Assistant Professor, Department of computer science and Engineering, Sathyabama Institute of Science and Technology, Tamil Nadu, Chennai ***

Abstract – The primary reason for middle-aged people's eyesight is age is diabetic retinopathy (DR). Early identification of the development of diabetic retinopathy can be very beneficial for clinical treatment. Although several different feature extraction various strategies have been put forth, and the classification job for retinal images is still tedious and time-consuming even for those trained clinicians. Hence, primary screening of DR is to avoid vision loss, it is advised that diabetic patients have this procedure performed at least once a year. Recently, deep convolutional neural networks have manifested superior performance in image classification compared to previous handcrafted feature-based image classification methods. As a result, a Random forest classifier has been developed that can distinguish the intricate elements required for classification, such as micro-aneurysms, exudate, and hemorrhages on the retina, and then automatically deliver a diagnosis without human input. Last but not least, a CNN-based automated DR screening approach for retinal pictures is suggested. This method displays the different phases of DR (Mild, Moderate, and Severe) as well as its attention map for the region that is most affected. It also reduces the workload of ophthalmologists. Thus the proposed system of CNN classifier gives a significant improvement in terms of speed and accuracy when compared to previous methods.

Key Words: Diabetic retinopathy (DR) Fundus Images (FIs),micro aneurysm (MA), Flame-shaped haemorrhages (FSHs), Convolutional Neural Network(CNN)

1. INTRODUCTION

Imageprocessingisaformofprocessingimagesthoseare either captured as pictures or frames for which the input is given as an image and the output of the image processing is also a picture associated with the image[1]. Image processing refers to digital image processing but the visual and analog processing is feasible as well[2]. Medical Image Processing is in which the images generatedfromthehuman bodyformedicalpurposesare subjected to processing. It helps easily to detect and

identifythedisease[3].DiabeticOneofthemainreasonsof retinal degeneration (DR) is sightlessness and there subsist valuable behaviours that hold back the development of the disease provided that it would be identifiedintheearlystage[4].Normalretinalassessment of the diabetic patients guarantees an early identification of DR, which considerably reduces the occurrence of blindness[5].Duetothehighprevalenceofdiabetes,mass screening takes a lot of time and requires a large number of qualified graders to carefully examine the fundus images looking for retinal abnormalities. Diabetes and otherdisorderslinkedtoagingandsocietyareontherise rightnow[6].Theissuesrelatingtotheeyescanbedivided into two main categories. The first is eye disease, such as cataract, conjunctivitis, blepharitis, and glaucoma. The second group is categorised as lifestyle-related diseases, including diabetes, hypertension, and atherosclerosis. Diabetescanharmtheeyesbydamagingtheretinalblood vessels, which can ultimately lead to visual loss. When diabetes is treated using prosthetic retinas, Diabetic retinopathy (DR) is the name used to describe this condition[7].Oneofthetreatmentstoreducetheamount ofvisualmutilationprocessedbyDRhasbeenidentifiedas early detection and diagnosis, with a focus on routine medical examinations for the identification and supervision of this condition. During this method, retina images, also known as fundus images (FIs), are carefully processed using a medical imaging camera and are physically checked for the presence of DR objects by screeners and ophthalmologists. Diabetic Retinopathy is an eye condition that diabetes patients experience to a great extent. If a diabetic patient's blood sugar levels are too high, the blood vessels at the back of their eye will be destroyed,whichpreventstheretina fromgetting enough nutrients to adequately retain their vision [8]. One of the main reasons for visual loss worldwide is diabetic retinopathy, also known as DR [9]. It is one of the main causes of preventable blindness and vision impairment [10].The prevalence of DR among diabetic patients globally was found to be 7.62%–47.1% based on a metaanalysis of 35 studies from 35 different countries. The second category of DR severity is non-proliferative

diabetic retinopathy. NPDR and proliferative diabetic retinopathy (PDR).There are three levels of NPDR: mild, moderate,andserious.Microaneurysm(MA)anddot/blot haemorrhage (HA) are early stages of mild NPDR. As the illness advances, flame-shaped haemorrhage’s (FSHs), cotton-wool patches, and hard exudates (HEs) In the moderateNPDRstage,(CWSs)becomevisible.Manymore MAs, HAs, or venous beading (VB) arise in the severe NPDR stage[11]. The most advanced form of DR is called PDR. Neovascularization (NV), pre-retinal haemorrhage’s (PHs), vitreous haemorrhage’s (VH), and fibrous proliferation(FP),whichisthesourceoftractionalretinal detachment, are the important pathologies[12]. Early screeninganddiagnosisofDRinthesediabeticpeoplecan stop vision loss and blindness. However, there isn't an ophthalmologist nearby in a remote rural region[13]. Consequently, an automation software is developed that can screen and DR with pathology extraction using algorithms for digital picture processing[14]. It is anticipated that this software will be a useful tool for medical professionals with limited expertise in DR diagnosis.

Fig.1.2:(a)NormalRetina (b)DiabeticRetina

Fig.1.1:HumanRetina

Figure1.1illustratesthefundusimageofanormalhuman retina. The retina is made up of a thin layer of lightsensitive tissue that is located close to the optical nerve. Light beams are concentrated onto the retina, where they aresubsequentlysenttothebrainforinterpretationofthe images. The macula, a relatively tiny region, is located at themiddle ofthe retina.Thepossibilityofpinpointvision is due to the presence of this macula that plays a major role in reading, writing or recognition of face[15]. The retinaisinturnsurroundedbyperipheralretina.Without the presence of retina, efficient communication between theeyesandbrainarenotpossiblewhereasonlyvision is possiblethroughit.

Diabetic retinopathy typically affects both eyes. In the early stages of the sickness, those who are frequently affected by the disorder do not notice changes in their eyesight. But, when it worsens, it frequently has irreversible effects; including vision loss. Figure 1.2 illustratesthenormalretinaversusdiabeticretina.

In the stage of diabetic retinopathy, blood vessel fluid leakage into the eye causes scarring of the retina. The onset of is the first sign of diabetic retinopathy. Haemorrhage in the retina[16]. The methods, algorithms, andtechniquesused toidentifyhaemorrhagefromretinal images of diabetic retinopathy are reviewed and explained. A fundus image-based algorithm based on a universal logical approach has been developed to detect the presence of Diabetic Retinopathy (DR) correlated lesions. It can distinguish between red and bright lesions and does not require any special pre- or post-processing. Several actions are carried out, and coloured retinal images are used to assess the various stages of diabetic retinopathy[17]. Micro aneurysms, which resemble small, secular pouches and look as tiny red dots, are brought on by a localised enlargement of the capillary walls. Another idea contends that the walls are brittle and prone to shattering, which might result in haemorrhages. Hard exudates are yellow lipid deposits that appear as vivid yellow lesions. The light, spherical region known as the optic disc is where the blood vessels initially develop. Visual acuity is greatest in the fovea, the central region of the retina. A mixture of interior components of microaneurysm detectors including macular centre and retinopathy-relatedlesiondetectionusingspecificallypreprocessing methods and applicant extractors are proposed[18]. The earliest stage of the illness is nonproliferative diabetic retinal disease, where the retinal blood vessels leak fluid or bleed. In NPDR, the arteries in the retina turn out to be very weak and they tend to be veryminuteanddotlikehaemorrhageswillbeseen.These types of weak blood vessels generally tend to swell or cause edema in the retinal image and it results in decreased vision. The symptoms of this disease will be mild or non-existent. Micro aneurysms, haemorrhages, hardexudates,macularedoema,andmacularischemiaare alterations brought on by NPDR that affect the eyes. Proliferative diabetic retinopathy (PDR) is now present since the illness has advanced to that point. PDR causes circulationproblems,whichmakesomepartsoftheretina ischemic or oxygen-depleted. New blood vessels become part of the circulatory system that helps the retina maintain enough oxygen levels[19]. Neovascularization is thewordforthis.Bloodmayenterthevitreousandretina, causing spots or floaters that are consistent with visual loss.SDRcausesaberrantvasculargrowthandscartissue,

which can be major difficulties for glaucoma, immediate retinaldetachment,andgradualvisionloss.

1.2. Symptoms of Diabetic Retinopathy

Certainsymptomsofdiabeticretinopathyidentifiedbythe research community are the observation of spots, dots or cobweb-like dark strings floating in the vision of the patients. Some patients experience hazy vision and a cyclicalchangeintheireyesightfromblurrytoclear.Some patients may experience black or dark spots in their field ofvisioninadditiontohavingimpairednightvision,which can ultimately lead to visual loss[20]. The retinal vessels are connected for a few more reasons. According to reports, this happens when the blood capillaries in the retina change, impacting diabetes patients and even leadingtoeyesightloss.Oncertaincasesthepatientswith retinal vessels suffer swelling and also observe leak fluid thatcannotbereversedaffectingthepatientsinlarge.

2. METHODOLOGY AND ALGORITHMS

2.1 Modules

2.1.1. Pre-processing

i. Augmentation

Augmentation can add randomized rotations to input images so that a network is invariant to the presence of rotationininputimages.

Input:Retinalfundusimage

Output:Augmentedimages

Figure2.1.1:Augmentationofretinalimage

ii.ResizeandNormalize

Imageresizingincreasesor decreasesthetotal number of pixels.

Normalizationisa processthatchangestherangeofpixel intensityvalues.

Input:Augmentedimages

Output:Resizedimage

2.1.2. Segmentation

The tiny, elongated structures in the retina are blood vessels. By segmenting blood vessels in retinal images, early illness identification is made possible. Automating this process has various advantages, including reducing subjectivityandremovinglabor-intensivesteps.Theoptic disc,which representsthe beginningoftheoptic nerve, is where the fibres of retinal ganglion cells converge. At the opticdisc,theretina'smajorbloodarteriesalsoenter.The fovea, a 1.5 mm broad depression on the internal surface of the photoreceptor layer, is made entirely of cone photoreceptorsand istailoredforthe bestpossiblevisual acuity. The 0.5mm-diameter foveal avascular zone is a region inside the fovea (An area without any blood vessels).

Input:Resizedimage

Output: Image segmented with Blood vessel, Optic disc andFovea

2.1.3. Classification

The final test item class is then chosen by averaging the votes from numerous decision trees from a randomly chosenportionofthetrainingset.

Input:SegmentedImages

Output: Image with diseases (MA, Haemorrhages, Exudates)

3.1 SYSTEM ARCHITECTURE

This chapter discusses the overall system architecture anddetaileddescriptionofallmodules.

2.1.4. CNN Classifier

Convolutional neural networks are one sort of artificial neuralnetwork(CNN).Itemploysperceptron,atechnique for supervised learning, to examine data. Each individual neurontakesinavarietyofinputs,weighsthem,andthen sends the weighted result through an activation function toproduceanoutput.

Input:Imagewithdisease

Output:ImageclassifiedbasedonseverityofMA

Figure3.1SystemArchitecture

Figure3.1describestheoverallsystemarchitectureofthe proposed DR detection with its severity from Retinal fundusimages.Thissystemstartswiththepre-processing stage where augmentation, resizing and normalization of retinal images is done. Several segmentations, including blood vessels, the optic disc, and the fovea, are found in thepre-processedimages.Usingarandomforestclassifier, the DR features are found in these segmented images. Finally for each DR feature detected the severity of the disease is calculated using the CNN classifier for better accuracy.

3.3 ALGORITHM

The Machine Learning system uses test data to assess the predictiveaccuracyofthetrainedmodelandtrainingdata to train models to recognize trends. By comparing predictions on the evaluation data set with actual values (also referred to as ground truth) using a variety of measures, machine learning systems assess their predictiveperformance.

• RandomForest

• ConvolutionalNeuralNetwork(CNN)

Thismodelemploystwocrucialideasthatgiveitthename random rather than averaging the predictions of trees, whicharereferredtoasthe"forest"

• selecting at random from training sets when creatingtrees

• Randomfeaturegroupsaretakeninto accountwhensplittingnodes

3.3.2 Convolutional Neural Network (CNN)

CNNs essentially employ very little pre-processing when in comparison to alternative picture classification methods. CNN is able to learn the rules that other algorithms require to be manually created. Input, output, and hidden levels are all present in CNNs. Convolutional, ReLU, pooling, and completely connected layers typically makeupthehiddenlayers.Animage,whichisamatrixof pixel values, serves as the input. Consider that the input matrix reading begins in the upper left corner of the image.Followingthat,thealgorithmselectsa filter,which is a smaller grid. (Or neuron, or core, or whatever). The filter then produces convolution as it processes the input image. It increases the filter values by the starting values ofthepixels.Thesesegmentsaremultipliedasawhole.At the conclusion, one number is obtained. The filter moves one unit to the right each time it completes an action of a similar nature because it has only scanned the picture in thetopleftcorner.Thefilteracquiresamatrixaftergoing through every position, but it is smaller than the input matrix. The nonlinear layer is added after every convolution operation. Through a process referred to as activation, it introduces nonlinear properties. A network wouldn't be robust enough to symbolize the response variable without this feature. (As a class label). The pooling layer follows the nonlinear layer. Pooling layers would result in fewer parameters when the pictures are too large. Spatial pooling, also referred to as subsampling or down sampling, reduces each grid's complexity while maintaining important data. There are three different kinds of geographic pooling: Max Pooling, Average Pooling, and Sum Pooling. As a result, the pooling layer performs a down sampling process on the image's width and height. The picture volume decreases as a consequence. This means that if some features (such as boundaries)werepreviouslyidentified in the convolution process, a detailed image is compressed into less detailed images. Before adding a fully connected layer, all convolutional, nonlinear, and pooling layers must be completed.Therawinputoftheconvolutionalnetworksis used in this layer. When a fully connected layer connects to the network's endpoint, N classes are created from whichthemodelchoosesthedesiredclass.

4. RESULT AND DISCUSSION

Image detection for effective identification of diabetic retinopathyistodiscovertheproblemrelatedto diabetes that can lead to sightlessness if not cured at its preliminarystages.Eventhoughthisresearchhascomplex recognition of DR lesions from retinal images, the effortless occurrence of any lesion is not sufficient to choose on the requirement for recommendation to a

patient. The computerized transmission for Diabetic Retinopathy (DR), a common complication of diabetes, faces a significant challenge in the identification of micro aneurysms in digital color fundus images. Numerous methods were available in the past to accomplish this recognition, but none ofthemhadever beencompared to one another on the same set of data.. Towards the recognition of DR, since micro aneurysms (MAs) are the earliest stage oftheillness, itiscrucial toclassifythem to determine whether or not they exhibit retinopathy symptoms. A novel supervised algorithm for recognizing bloodvesselsvisible in retinal imagesispresented, which employs a Neural Network (NN) model for pixel classification and evaluation of a 7-D vector made up of features based on moment invariants and grey levels for pixel representation. Another method presents the outcome of the microaneurysm recognition prepared in the circumstance of the Retinopathy Online Challenge (ROC), for different features of DR detection. Scientific interest lies in another unique method known as the regularMArecognitionfromdigitalcolourfundusimages, which acts as an early indicator of diabetic retinopathy and their typical recognition from colour retinal images. Diabetic retinopathy is a condition that affects millions of individual’s worldwide (DR).Thus an automatic mechanism to identify the presence of DR along with its severity by evaluating the photograph of the central field of the retina has been developed. Selected pre-processing techniques are carried out for Micro aneurysm discovery, which is crucial in grading diabetic retinopathy, after digital fundus pictures revealed diabetic retinopathy. All recent works have assumed that Visual Dictionaries for Automatic Retinal Lesion Detection entails the developmentofanautomaticDRscreeningsystemcapable of detecting the presence of many DR-related abnormalities. The previous points of interest and visual dictionaries methods of each specific lesion are identified andtheydetecttheautomaticretinallesion,increasingits accuracy rate. But, if the level of specific lesions is increased, then the detection of automatic retinal lesions will become a complicated process which reduces its accuracy rate. In addition, the detection time becomes complex and increases. The present study is conceived with automatic diabetic retinopathy detection technique from fundus images. In addition, it also smoothens the detection technique to avoid complexity and confirm a higher accuracy. This verifies better accuracy and offers efficient detection at a higher level of sensitivity. Finally, the suggested approach employs deep convolutional neural network models that rank the severity of DR in fundusimagesinordertodiagnosethepresenceofDRand offer pertinent advice to DR patients. The goal is to use a deep convolutional neural network to assess the severity ofdiabeticretinopathy.Thegoalistousetrainingdatasets to train the algorithm. Every day, more people are diagnosed with diabetic retinopathy. Despite mounting proof of the value of routine DR screening and early

intervention, it frequently results in poor vision function and is the main cause of blindness. Due to insufficient medicalcare,ithasfrequentlybeenneglectedinthehealth care system and in many low-income nations. A system thatwillprovidepredictionsaboutdiabeticretinopathyis built because there aren't enough methods to detect the condition.

5. CONCLUSIONS

The proposed system assesses the severity of diabetic retinopathy in a patient using digital image processing techniques on fundus images. In the proposed study, a computer-based approach is employed to assess the degree of DR using a CNN classifier, and the results show that DR can be discerned rather well from fundus photographs. It can be used as a substitute or supplemental instrumentforDR screening, particularly in remote locations where ophthalmologists are scarce or in rural areas where ophthalmologists are overburdened with patient cases. In order to increase the DR classification's accuracy, extra digital image processing methodsorotherdeeplearningandartificial intelligencebased techniques may need to be developed in software. Sensitivity, Specificity, and Accuracy performance parametersexhibitbetterperformancewhencomparedto values determined by human observers for these parameters. The outcome demonstrates unequivocally that the suggested approach is successful in identifying severity in DR images. The proposed method has an accuracyof81%.

REFERENCES

[1] Zhentao Gao, Jie Li, Jixiang Guo, Yuanyuan Chen , Zhang Yi , Jie Zhong, ‘Diagnosis of Diabetic Retinopathy using Deep Neural Networks’ IEEE Access,Vol.7,pp:3360-3370,2018.

[2] Yi-Peng Liua, Zhanqing Lib, Cong Xuc, Jing Lid, Ronghua Lianga, ‘Referable diabetic retinopathy identification from eye fundus images with weighted path for convolutional neural network’, Artificial IntelligenceInMedicine,Vol.9,pp:0933-3657,2018.

[3] YogeshKumaran,ChandrashekarM.Patil,’ABrief Review of the Detection of Diabetic Retinopathy in Human Eyes Using Pre-Processing & Segmentation Techniques’, International Journal of Recent Technology and Engineering, Vol. 7, pp : 2277-3878, Issue-4S2,2018.

[4] U. Budak, A. Şengür, Y. Guo, and Y. Akbulut, ‘A novel microaneurysms detection approach based on convolutional neural networks with reinforcement sample learning algorithm’ , Health Information ScienceandSystem,Vol.5,pp:2367-2377,2017.

[5] G.Quellec,K.Charrière,Y.Boudi,B.Cochener,and M. Lamard, ‘Deep image mining for diabetic retinopathy screening’, Medical Image Analysis, Vol. 39,pp:178-193,2017.

[6] M. R. K. Mookiah, T. Lin, J. Yang, J. Fan, ‘Evolutionary algorithm based classier parameter tuning for automatic diabetic retinopathy grading: A hybrid feature extraction approach’, KnowledgeBasedSystem,Vol.39,pp:9-22,2013.

[7] BehdadDashtbozorg,JiongZhang,FanHuang,and Bart M. ter Haar Romeny, ‘Retinal Microaneurysms Detection using Local Convergence Index Features’, IEEE Transactions on Image Processing, Vol. 27, Issue:7,pp.3300-3315,2018.

[8] Nathan Silberman., et.al.., ’Review of Automated Detection for Diabetes Retinopathy Using Fundus Images’, International Journal of Advanced Research in Computer Science and Software Engineering, Volume5,Issue3,March2010.

[9] Su Wang, Hongying Lilian Tang, Lutfiah Ismail Al turk, Yin Hu, Saeid Sanei,George Michael Saleh and Tunde Peto, ‘Localizing Microaneurysms in Fundus Images Through Singular Spectrum Analysis’, IEEE Transactions on Biomedical Engineering , Vol.64 , Issue:5,pp.990-1002,2016.

[10]A.Alaimahal, et al., “Identification of diabetic retinopathy stages in human retinal image”, Luca Giancardoa, et al., “Microaneurysms detection with Radon Cliff operator in retinal fundus image”, IEEE TransactionsonMedicalImaging,Proc.OfSPIEVol.7, 2010.

[11]S. Tang, T. Lin, J. Yang, J. Fan, ‘Retinal Vessel Segmentation using Supervised Classification based on Multi-scale vessel filtering and Gabor Wavelet’, Jour.ofMed.Imag.&HealthInfo.,Vol.5,pp.1571-1574, 2015.

[12]ShraddhaJalan,etal.,‘ReviewpaperonDiagnosis of Diabetic Retinopathy using KNN and SVM Algorithms’, International Journal of Advance Research in Computer Science and Management StudiesVolume3,Issue1,January2015.

[13]Marco Russo ‘Genetic fuzzy learning’, IEEE Transactions on Evolutionary Computation”, Vol. 4, No.3,pp.259-273,2000.

[14]Meindert Niemeijer, et al., ‘Automated Detection and Differentiation of Drusen, Exudates, and CottonWool Spots in Digital Color Fundus Photographs for Diabetic Retinopathy Diagnosis’, IOVS, Vol.48, No. 5, May2007.

[15]C.G.Ravichandran, ‘Blood vessel segmentation for High Resolution Retinal images’, IJCSI International JournalofComputerScienceIssues,Vol.8,Issue6,No. 2,pp.389-393,2011.

[16]Lam,GaoandLiew,(2010)‘Generalretinalvessel segmentation using Regularization based multiconcavity modeling’, IEEE Trans. Med. Imag, pp. 1369–1381.

[17]L.XuandS.Luo,(2010)‘Anovelmethodforblood vessel detection from Retinal images’, Biomed. Eng. Online,Vol.9,No.1,p.14.

[18]Ana Salazar-Gonzalez, Djibril Kaba, Yongmin Li, and Xiaohui Liu, (2014) ‘Segmentation of the Blood VesselsandOpticDiskinRetinalImages’,IEEEJournal ofbiomedicalandhealthinformatics,Vol.18,No.6.

[19]Clara I. Sanchez, et al., (2006) ‘Automatic Image Processing Algorithm to Detect Hard Exudates based on Mixture Models’, Proceedings of the 28th IEEEEMBSAnnualInternationalConferenceNewYork City,USA,Aug30-Sept3.

[20]A.M. Mendonca and A. Campilho (2006), ‘Segmentation of retinal bloodvessels by combining the detection of centerlines and morphological reconstruction’, IEEE Transactions on Medical Imaging,Vol.25,No.9,pp.1200-1213.