View of Automatic Detection of Diabetic Retinopathy in Retinal Images: A Study of Recent Advances

(1)

15277 http://annalsofrscb.ro

Automatic Detection of Diabetic Retinopathy in Retinal Images: A Study of Recent Advances

Telagarapu Prabhakar¹, Gurram Sunitha², Gudavalli Madhavi³, J. Avanija⁴, K.Reddy Madhavi⁵*

1Associate. Professor, ECE,GMR Institute of Technology, GMR Nagar, Rajam,Andhra Pradesh, [email protected].

2Professor, CSE, Sree Vidyanikethan Engineering College, Tirupati, India,[email protected]

3Assistant Professor, CSE, JNTUK University College of Engineering, Narasaraopet Guntur, [email protected].

4 Associate Professor, CSE, Sree Vidyanikethan Engineering College, Tirupati, India,[email protected]

5 Associate Professor, CSE, Sree Vidyanikethan EngineeringCollege,Tirupati, India, [email protected]

ABSTRACT

Most of the people throughout the world are suffering with the disease called diabetes. In severe condition of diabetes, patient may lose vision. This could be prevented by using a method called Diabetic Retinopathy, because diabetes mostly affects retina of the eye.Diabetic Retinopathy could be prevented only by earlier detection. For this detection many research workout has been carried out previously in which they used methods like Convolutional Neural Network, Deep Neural Network, Adaptive thresholding, Gabor wavelet transform method. By using the above methods blood vessel classification can be done, exudates and fundus can be identifiedin the eye. These identified stages are classified as normal proliferative retinopathy, moderate proliferative retinopathy, severe proliferative retinopathy. For the analysis purpose we used data set of images are used like DRIVE AND STARE. By comparing all the studies done previously, best method of diabetic retinopathy is detected,depending on the parameters like accuracy, sensitivity, specificity.

I. Introduction

Diabetic retinopathy (DR) is a rapidly spreading condition caused by diabetes that affects people all over the world. Diabetic patients may experience total vision loss because of the DR. In this situation, early detection of DR is much more important to restore vision and aid with prompt care.

The identification of DR may be achieved manually by ophthalmologists or automatically by a computerized machine.In the manual method, ophthalmologists are used to analyze and explain retinal fundus images, which is a time-consuming and costly process; however, in the automated system, artificial intelligence is used to play a critical role in ophthalmology, especially in the early identification of diabetic retinopathy, which is a time-consuming and expensive task.Several advanced studies on the detection of DR have recently been published. This paper examines the diagnosis of DR in depth, focusing on three main aspects: retinal datasets, DR detection techniques, and performance assessment metrics.

Rest of the paper is organized as follows. Section II focuses on relevant study, section III presents methodologies, section IV shows the comparison of performance metrics such as specificity, sensitivity and accuracy on different datasets, finally section V gives conclusion of our study.

(2)

II. Relevant Study

Ronnie D.caytiles[1] proposed a model to detect the rigorousness of Diabetic retinopathy.it contains three different types of neural networks like back propagation neural network, convolutional neural network, and Deep Neural Network to detect the severity of retinal problems. In testing and training of raw images taken the DNN outperforms CNN. By this process 72.5% of accuracy is obtained.KishoreBalasubramanian et.al proposed [2] that two different technologies like SVM and CNN the patterns and characteristics of the sample images can be extracted efficiently. It is observed that there is 4% improvement when compared to other existing systems and classification systems.

For performing the segmentation, the required images are taken from two data sets like DRIVE AND STARE which means digital retinal image for vessel extraction and structured analysis of the retina. The taken raw images are segmented using orientation based segmentation.Maria c Garcia et al proposed a system in this paper [3] to detect the hard exudates that are formed due to lesions.

Exudate is the clear white fluid which comes from the damaged part of the skin.These hard exudates can be visible in retinal images.In this paper they used three neural network classifierslike multilayer perceptron(MLP), radial basis function(RBF) and support vector machine (SVM)to detect those exudates. Forthis classifier they used 117 images in data set among which 50 are from DR patients.

M Usman Akram et al[4]proposed a new method in this paper to enhance and segment the blood vessels in retina.By using this method we can easily detect the diabetic retinopathy.Here, they used 2D Gablor Wavelet and multilayered thresholdingtechnique.This method is used for illuminate the blood vessels and divide the thinnest vessels.In this method also they used DRIVE AND STARE in which many images are available publicly for analysis.This blood vessel segmentation is mainly used to find whether any patient is having diabetic retinopathy are not.Dilip Singh Sisodiaet al [5]

proposed thatKaggle Diabetic Retinopathy dataset, histogram equalization and by calculation of mean value and standard deviation from which the desired 14 featured results can be obtained. As the analysis of different raw retinal fundus images it is very hard to process by using machine learning algorithms,they used the above methods.To find the Diabetic Retinopathy in early stages they used three different types of images of patients with disease in stages like mild,severe, and normal.Subhasis Chaudhuri et al [6] proposed two dimensional matched filters for the diagnosis of blood vessel segmentation for detecting piecewise linear segments in the retinal images.They have used edge detection methods like soble and prewitt as done in the paper [4]. These operators are widely used in the field of digital image processing.They used an operator in order to detect optical and spatial properties.The obtained results are compared with other methods also. The convolutional algorithm used in this paper takes more time to run on an ordinary computer. So, this drawback can be solved by using more efficient and compact algorithm.

RishabGargeya et al [7] proposed in this paper that the diabetic retinopathy can be done by dynamic deep learning method called artificial intelligence-based grading method. In this process they analyzed many processed color fundusimages. They also used second level gradient boosting

(3)

classifier which is also called as tree-basedclassifiers. VarunGulshan et al [8] suggested a new algorithm in this paper that detects diabetic retinal diseases and diabetic macular edemaautomatically. More than 1 lakh photographs were used in this study, all of which were graded by ophthalmologists.

Michael david Abram off et al [9] compared different papers in which they used deep learning algorithms to detect the diabetic retinopathy. Some important parameters like sensitivity,specificity, accuracy, AUC,negative predictive value, and their confidence intervals were calculated. Feng li et al [10] analyzed 19,233 retinal color images obtained from many adult patients in this paper. By this process feng li want to detect the diabetic retinopathy. In this paper they used inception V3 network for deep learning approach. To differentiate the disease severity, he divided the obtained images as no apparent DR, mild non proliferative DR, moderate non proliferative DR, and severe non proliferative DR.

T. Jemima Jebaseeli et al [11] proposed a method in this paper in which blood vessel can be segmented by using deep learning based SVM and tandem PCNN model. Generally diabetic retinopathy occurs for type2 diabetes.The Contrast Limited AdaptiveHistogram Equalization can improve the contrast of the image and it also removes the noise from the images captured at different illuminations. Tandem Pulse Coupled Neural Network model combines two original images and gives the feature vectors to segment the retinal vasculature.Dr. Kulwinder S Mann et al [12] here images are taken from data sets like DRIVE AND STARE and analyzed for different preprocessing features like image cropping, contrast enhancement, Gabor filtering, intensitymeasure, intensity, gray level-based features.

Jaakosahstlen et al [13] proposed a method for deep learning analysis of fundus images and macular edema grading. This method is specially designed with grading systems by many ophthalmologists.

Firstly, the patients who are taken for diabetic retinopathy assessment are graded as PIRC and PIMEC.PIRC means proposed international clinical diabetic retinopathy and macular edema scales.M. Elena Martinez-Pkrez et al [14] proposed a method for blood vessel segmentation by using second order derivative and region growing process. Generally, region segmentation, region growing, and region merging are important concepts in the field of digital image processing. The minimum eigenvalue and the peakvalueofits gradient refers to the characteristics fora region growingmethod which is classified in two stages. In the first stage, growth is restricted to regions with low gradients, allowing vessels to grow where the valuesofthe minimum eigenvalue lie within a wide interval and allowing rapid growth ofbackgroundregionsoutsideof the vessel boundaries. For the second stage, in which the borders between classes are defined, the algorithm grows vessel and background classes simultaneously without the gradient restriction.

Daniel ShuWei Ting et al [15] compared many patients from different nations with different food habits, different lifestyles in this paper. By comparison with other studies author allocated specificity, sensitivity, and accuracy to 2 different categorized people. One of them are people

(4)

having high threat of vision loss due to diabetes and another one is people with mild diabetes.K.A.Vermeeret al [16] gave a method to detect the blood vessel by laplace and thresholding segmentation step followed by classification step.The images taken for analysis are taken from device named as GDx based on scanning laser polarimetry (SLP) by which two images can be produced like retardation images and reflectance images. This method provides an excellent way of detecting blood vessels, especially for images with specular reflection.

Songyuan Tang et al [17] proposed an automated method for the segmentation of retina vessel. It consists of feature extraction,support vector machine classification and refinement of vessels.The Gabor wavelet transform is used in many studies of image processing as it can represent the signal in both time and frequency domains.For obtaining accuracy values true positive, true negative, false positive and false negative values are required.Xiaoyi Jiang et al [18] proposed a novel method which the blood vessel detection of retinal images can be obtained by adaptive local thresholding based multi threshold probing.In this paper we cannot use global thresholding because the object pixel intensity and background pixel intensity is different. So adaptive thresholdingis used. For the taken retinal images for assessment threshold values are determined. Wong Li Yun et al [19] gave method in which the retinal diabetes is identified by using different image processing techniques on 124 sample retinal images taken from patients. The raw images taken are classified in two groups like normal retina, moderate proliferative retinopathy, severe proliferative retinopathy, and proliferative retinopathy. In this method back propagation algorithm is used to classify the four stages of diabetic retinopathy and also used some important operations like opening, closing, dilation, erosion with original image denoted as A and structuring element B.Lei Zhang et al [20]

proposed a method in which the blood and non blood vessel pixels are separated by using. In method we used Gabor filter for getting image texture.Here multi-scale framework allows us to extract features relating to different vessel widths and these are used to automatically generate textons at salient scales. In the test phase image is filtered with the Gabor filter to generate corresponding responses for every pixel. Pixels are then labelled as vessel/non-vessel by a 1-NN classifier which assigns them to the nearest textoncluster in the dictionary. The paper is organized as follows. Section 1 describes as Introduction; Section 2 gives an outline of the various strategies for detecting DR.

Finally, Section 3 we Concluded the paper.

III. Methodology a.) Neural Networks with Multilayer Perceptron:

The neural network model considered is NN having multilayer perceptron. It’s a basic NN model having multiple layer of computational units inter connected via feed forward way [23]. The neurons for Perceptron modeling is being calculated through. The epoches has being selected based on multiple runs of the NN model, and as the model trained 268 epoches are considered as the approximately accurate.

(5)

b.) Deep Neural Network (DNN):

After validating the accuracy result of NN model with testing samples, the next DNN models have been implemented. The idea of apply this model is to check the accuracy with the NN model. Model accuracy definitely will differ because of utilizing with multiple hidden layer concerning input plus output. The multiple layer extracts feature from lower level of layers and increases the potential of a small network. In DNN model as algorithm “Back Propagation” is being used. Some of the study suggested that even for Sparse network “Back Propagation” performs well. The weight updated in back propagation via “stochastic gradient descent” algorithm, Here, η is assigned as learning rate, C is considered as cost function, and ξ(t) a type of stochastic function. The method of DNN having three Fully Connected (FC) layer. Two activation layers with “ReLU” as activation function, for building model “Tanh” and “Sigmoid” activation functions also but “ReLU” performs better than other two approaches. “ReLU” uses smooth rough calculation to the rectifier. The main activation function defined as. The ReLU considered as the most popular activation function as for Computer Vision [28]. The DNN model uses one softmax layer or normalized exponential function as, International Journal of Grid and Distributed Computing Vol. 11, No. 1 (2018) Copyright ⓒ 2018 SERSC Australia 99 where, Pj represents the class probability (output of the unit j) and xj and xk represent the total input to units j and k of the same level respectively. Cost function defined as, here, dj represents the target probability for output unit j and Pj is the probability output for j after applying the activation function. Both NN and DNN has been carried out on both statistical data of diabetic images and also processed retinopathy images. The difference is coming for the training time of the of both techniques. During work out sessions with DNN, statistical data taken less time then image data.

RETINAL FUNDUS IMAGE DATA SET PREPROCESSING

FEATURE EXTRACTION

EXTRACTE D FEATURES BLOOD

VESSEL AREA

BIFURC ATION POINTS

MICRO ANEURY SMS AREA SHANO

N ENTROP

Y OPTICAL DISTANC

E AREA

HEMO RGAES

AREA EXUDA

TES AREA

QUANTITATIVE ANALYSIS

FEATURE RANKING

(6)

Fig.1: Flow Diagram of Diabetic Retenopathy Segmentation Process

c.) Convolutional Neural Network (CNN):

Only image data is being trained for CNN model. Processed images with single band have been given as a input of the network with given levels. The CNN model has been considered is the VGGnet model. The VGGnet model structured of CONV layers which performs 3x3 convolution alongside stride:1 and pad:1, and of POOL layers which performs 2x2 maxpooling with stride:2.

There is no padding existed in the network. The network trained with CPU support. As activation function ReLU has been used. All convolutional layers are followed with Maxpool used in the pooling layed for extracting the most significant feature between the image pixel. VGGnet works very well with densed featured images. As per model there is no normalizaton layers used here, because either way it does not improve the accuracy of the model. Figure 8 giving the model view.

(7)

Figure 2.Model Architecture of Diabetic Retinopathy Analysis

The simple PCNN model is inadequate to complete the segmentation procedure. In multispectral PCNN (m-PCNN), m-channels are fused to create the outcomes. The resultant image provides more details, but it doubles the execution time.

Figure 3. Representation of DR with increasing severity: (a) no apparent DR images (b)mildnonproliferative DR (NPDR) image (c) moderate NPDR image (d) Severe NPDR image(e)

proliferative DR (PDR) image

Hence, there is a need of Table 1 Fundus Image Databases. Database No. of Images Image size (pixels) DRIVE 40 584 × 584 STARE 20 700 × 605 REVIEW 16 1360 × 1024 HRF 15 3504 × 2336 DRIONS 110 600 × 400 Fig. 1. Preprocessing results of the fundus image (DRIONS). T.J. Jebaseeli, et al. Tandem PCNN model, which fuses the data from two source input images. Tandem PCNN (TPCNN) model is shown in Fig. 3 which triggers the inter and intra channel linking of the input neurons.

IV. Comparison and Assessment Parameters

In this section, we have presented various methodologies and analysis of performance assessment metrics by various authors on different datasets and features extracted.

Table 1. Comparison of retinal datasets, DR detection techniques, and performance assessment metrics.

S.No Author Feature

extraction Methodology Accuracy

%

Sensitivity

%

Specificity

% AUC DATA

SET

1 Ronnie D Caytles

Severity of retinal problems

CNN, DNN, BACK

PROPAGATIO N NN

72.50% Fundus

images

(8)

2

Kishore Balasubrama nian

Pattern and features are extracted

CNN, SVM 74% DRIVE &

STARE

3 Maria C

Garcia

Find exudates due to lesions

MLP, RBF,

SVM 92.54% 100% 92.57%

117 PATIENT

S RAW

IMAGES

4 M

UsmanAkra m

Enhance and segment blood vessels

2D Gabor

wavelet Method and Multilayer thresholding

95.02%(ST ARE) 94.6%

(DRIVE)

DRIVE &

STARE

5 Dilipsinghsis odia

14 feautres are extracted and find exudates

Kaggle Diabetic Retinopathy,His togramEqualizti on ,mean value

6 Subhasis Chaudhuri

detect blood vessel segmentati on

Edge detection perators&Conco lutional algorithm

7 RishabGarge ya

Processing color fundus images

Tree based

Classifiers 0.97

MESS IDOR 2 EOPHTH A

8

Michael David Abraham off

Detect Diabetic Retinopath y

Idx DR X2 98.01 96.80% 87%

9 Varun Gulshan

Diabetic retinal diseases &

Macular Edema

Idx DR X2 0.99%

99.1%

87%

90.3%

98.5%

98.1% 0.98 Messidor 2 Eyepacs 1

10 Feng li

Detect Diabetic Retinopath y

Inception

Network 93.49% 96.93% 93.45% 99 19,233

images

11 T. Jemima Jebaseeli

Blood vessel segmentati on

SVM AND

Tandem PCNN 99.49% 80.61% 99.54%

DRIVE, STARE,H RF,REVIE W &

DRIONS

12

Dr.

Kulwinder S Mann

Early detection of Diabetic Retinopath y

94.71%

95.05%

78.08%

78.37%

98.32 %

98.35% DRIVE STARE

13 Jaakosahstle n

Deep learning analysis

89.60% 97.40% 0.99

(9)

and macular edema

14

M. Elena Martinez- Pkrez

Blood vessel segmentati on

second order derivative and region growing process.

15

Daniel ShuWei Ting

Analyze Diabetic retinopathy

comapred

different studies 100 91.1 0.96

49466 1 RETINAL IMAGES

16 K.A.Vermee r

detect blood vessel

laplace and

thresholding 92.40% 92.10%

GDx based on

scanning laseer polarimetr y

17 Songyuan Tang

segmentati

on of

retina vessel

SVM and Gabor wavelet

transform method

95.05%

18 Xiaoyi Jiang blood vessel detection

adaptive local thresholding&m ultithreshold probing

19 Wong Li Yun

identify retinal images

Back propagation algorithm

84% 90% 100% 124 images

20 Lei Zhang

identify blood and non blood vessel pixels

Multi scale frame network and Gabor filter

95.05% 78.12% 96.68%

From the Table 1 it is observed that Ronnie D.caytiles proposed three different types of neural networks like back propagation neural network, convolutional neural network, and Deep Neural Network to detect the severity of retinal problems. By this process 72.5% of accuracy is obtained.

Kishore Balasubramanian et.al proposed SVM and CNN. It is observed that there is 4%

improvement when compared to other existing systems and classification systems. For performing the segmentation, the required images are taken from two data sets like DRIVE AND STARE which means digital retinal image for vessel extraction and structured analysis of the retina. Maria c Garcia et al proposed to detect the hard exudates that are formed due to lesions. Here they used three neural network classifiers like multilayer perceptron (MLP), radial basis function (RBF) and support vector machine (SVM)to detect those exudates. For this classifier they used 117 images in data set among which 50 are from DR patients. After detection of three neural network classifiers, the multilayer perceptron among all the other three classifiers delivered the better results with mean sensitivity, positive predictive value and accuracy are 100%, 92.57%, 92.54%. M Usman Akram et al proposed

(10)

to enhance and segment the blood vessels in retina. Here, they used 2D Gablor Wavelet and multilayered thresholding technique. This method is used for illuminating the blood vessels and divide the thinnest vessels. In this method also they used DRIVE AND STARE in which many images are available publicly for analysis. This blood vessel segmentation is mainly used to find whether any patient is having diabetic retinopathy are not. They also used edge detection for blood vessels. When compared to other papers our new method gives accuracy of 0.946 through DRIVE and by STARE database images, we got accuracy of 0.9502.Dilip Singh Sisodia et al proposed that Kaggle Diabetic Retinopathy dataset, histogram equalization and by calculation of mean value and standard deviation from which the desired 14 featured results can be obtained. As the analysis of different raw retinal fundus images, it is very hard to process by using machine learning algorithms, they used the above methods. To find the Diabetic Retinopathy in early stages they used three different types of images of patients with disease in stages like mild, severe, and normal. Subhasis Chaudhuri proposed two dimensional matched filters for the diagnosis of blood vessel segmentation for detecting piecewise linear segments in the retinal images. They have used edge detection methods like soble and prewitt as done in the paper. They used an operator to detect optical and spatial properties. The convolutional algorithm used in this paper takes more time to run on an ordinary computer. So, this drawback can be solved by using more efficient and compact algorithm.

Also two of the 12 different kernels that have been used to detect vessel segments along different directions.RishabGargeya et al proposed in this paper that the diabetic retinopathy can be done by dynamic deep learning method called artificial intelligence-based grading method. In this process they analyzed many processed color fundus images. They also used second level gradient boosting classifier which is also called as tree-based classifiers. By using this method, they tested databases like MESSIDOR2 and E-ophtha and achieved 0. 97AUC.Varun Gulshan et al suggested a new algorithm in this paper that detects diabetic retinal diseases and diabetic macular edema automatically. The algorithm had a region under the receiver operating curve of 0.991 for EyePACS- 1 and 0.990 for Messidor-2 when it came to detecting RDR.EyePACS-1 had a sensitivity of 90.3 percent and a precision of 98.1 percent by using the first working cut point with high specificity. The sensitivity of Messidor-2 was 87.0 percent, while the specificity was 98.5 percent.Michael David Abr`amoff et al proved that sensitivity is 96.8 percent, specificity is 87 percent, AUC is 0.980, negative predictive value is 99.9 percent. Feng li et al analyzed 19,233 retinal color images obtained from many adult patients in this paper. By this process feng li want to detect the diabetic retinopathy.

Here they used inception V3 network for deep learning approach. To differentiate the disease severity, he divided the obtained images as no apparent DR, mild non proliferative DR, moderate non proliferative DR, and severe non proliferative DR. In this analysis parameters calculated are 96.93% sensitivity,93.45%specificity,0.9905area under receiver curve (AUC),93.49% accuracy, k value 0.919 is the best value obtained.

T. Jemima Jebaseeli et al proposed a method in which blood vessel can be segmented by using deep learning based SVM and tandem PCNN model. Generally diabetic retinopathy occurs for type2 diabetes. By this proposed model 80.61% of sensitivity,99.54% of specificity,99.49% accuracy is

(11)

achieved. Dr. Kulwinder S Mann et al images are taken from data sets like DRIVE AND STARE and analyzed for different preprocessing features like image cropping, contrast enhancement, Gabor filtering, intensity measure, intensity, gray level-based features. For DRIVE datasets on an average for taken images specificity of 0.9832, sensitivity of 0.7808, accuracy of 0.9471 is obtained, whereas in the case of STARE data set of images on an average specificity of0.9835, sensitivity of 0.7837, accuracy of 0.9505 is obtained.Jaakosahstlen et al proposed a method for deep learning analysis of fundus images and macular edema grading. This method is specially designed with grading systems by many ophthalmologists. Firstly, the patients who are taken for diabetic retinopathy assessment are graded as PIRC and PIMEC.PIRC means proposed international clinical diabetic retinopathy and macular edema scales. They are further classified into three grading systems like NRDR/RDR, NRDME/RDME, three class system of ungradable/NRDR/RDR. In this paper sensitivity of 0.89six is achieved for taken 2097-pixel sized images, 0.974 specificity is taken for five12 pixel sized images and AUC of 0.987 is obtained for 299-pixel sized images. M. Elena Martinez-Pkrez et al proposed a method for blood vessel segmentation by using second order derivative and region growing process. Daniel ShuWei Ting et al compared many patients from different nations with different food habits, different lifestyles. From the results it is observed that 91.1%, sensitivity is 100

% and AUC is0.958. For people with mild diabetes the specificity is 91.6%, sensitivity is 90.5% and AUC is 0.936. K.A.Vermeeret al [16] gave a method to detect the blood vessel by Laplace and thresholding segmentation step followed by classification step. In this paper they achieved sensitivity of 92 .4 percent and specificity of 92.1 percent. The images taken for analysis are taken from device named as GDx based on scanning laser polarimetry (SLP) by which two images can be produced like retardation images and reflectance images. Songyuan Tang et al proposed an automated method for the segmentation of retina vessel. It consists of feature extraction,support vector machine classification and refinement of vessels. By this process we can obtain the accuracy more than 95%.The Gabor wavelet transform is used in many studies of image processing as it can represent the signal in both time and frequency domains. The accuracy and ROC of this method obtained are 0.95 and 0.9620, respectively. Xiaoyi Jiang et al proposed a novel method which the blood vessel detection of retinal images can be obtained by adaptive local thresholding based multi threshold probing. Wong Li Yun et al gave method in which the retinal diabetes is identified by using different image processing techniques on 124 sample retinal images taken from patients. The raw images taken are classified in two groups like normal retina, moderate proliferative retinopathy, severe proliferative retinopathy, and proliferative retinopathy. By using this method specificity of 100 %, sensitivity of 90%and accuracy of 84% is obtained. B.Lei Zhang et al proposed a method in which the blood and non-blood vessel pixels are separated by using Gabor filter for getting image texture. By this process sensitivity of 78.12%, accuracy of 95.05%, specificity of 96.68% is obtained.

V. Conclusion

Many eye disorders require the use of blood vessels for diagnosis. As opposed to previous methods, the proposed approach accurately separates the blood vessels from retinal images. For

(12)

medical diagnosis, blood vessel segmentation can be performed with greater precision and reliability. By analyzing different methods proposed by research specialists in this survey paper we came to know that the method proposed by T. Jemima Jebaseeli to segment the blood vessels by using the SVM AND Tandem PCNN methods is the best method among the analyzed papers. By this method, the accuracy rate is 99.99% and specificity is 99.54% and sensitivity is 80.61% which is more when compare to another research. Another method which stands in second place is the method proposed by Michael David Abraham off to detect diabetic retinopathy with accuracy of 98.01%,sensitivity of 96.80% and specificity of 87%. DRIVE, STARE, HRF, REVIEW& DRIONS data base images are used to find out these parameters.

References

[1] Suvajit Dutta, Bonthala CS Manideep, Syed MuzamilBasha,Ronnie D. Caytiles1 and N. Ch. S.

N. Iyengar2, “Classification of Diabetic Retinopathy Images by Using Deep Learning Models”,http://dx.doi.org/10.14257/ijgdc.2018.11.1.09.

[2] Kishore Balasubramanian1 · N. P. Ananthamoorthy2, “Robust retinal blood vessel segmentation using convolutional neural network and support vector machine”,

https://doi.org/10.1007/s12652-019-01559-w,2019.

[3] MaríaGarcía a,∗, Clara I. Sáncheza, María I. Lópezb, Daniel Abásoloa, Roberto Horneroa,

“Neural network based detection of hard exudates in retinal image”,www.intl.elsevierhealth.com/journals/cmpb.

[4] M. Usman Akram • Shoab A. Khan, “Multilayered thresholding-based blood vessel segmentation for screening of diabetic retinopathy”, DOI 10.1007/s00366-011-0253-7

[5] DILIP SINGH SISODIA*, SHRUTI NAIR and POOJA KHOBRAGADE, “Diabetic Retinal Fundus Images: Preprocessing and Feature Extraction For Early Detection of Diabetic

Retinopathy”,http://dx.doi.org/10.13005/bpj/1148,2017.

[6] SUBHASIS CHAUDHUR , SHANKAR CHATTERJEE, NORMAN KATZ. MARK NELSON, AND MICHAEL GOLDBAUM, “Detection of Blood Vessels in Retinal Images Using Two-

Dimensional Matched Filters”.

[7] Rishab Gargeya,1 Theodore Leng, MD, MS2, “Automated Identification of Diabetic Retinopathy Using Deep Learning” ,

[8] Varun Gulshan, PhD; Lily Peng,MD, PhD; Marc Coram, PhD; Martin C. Stumpe, PhD;

DerekWu, BS; ArunachalamNarayanaswamy, PhD, “Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs”

doi:10.1001/jama.2016.17216.

(13)

[9] Michael David Abr`amoff,1–3 Yiyue Lou,4 Ali Erginay,5 Warren Clarida,3 Ryan Amelon,3 James C. Folk,1,3 and Meindert Niemeijer3, “Improved Automated Detection of Diabetic Retinopathy on a Publicly Available Dataset Through Integration of Deep Learning”, DOI:10.1167/iovs.16-19964.

[10] Feng Li1, Zheng Liu1, Hua Chen1, Minshan Jiang1, Xuedian Zhang1, and Zhizheng Wu,

“Automatic Detection of Diabetic Retinopathy in Retinal Fundus Photographs Based on Deep Learning Algorithm”, https://doi.org/10.1167/tvst.8.6.4.

[11] T. Jemima Jebaseelia,⁎, C. Anand Deva Duraib, J. Dinesh Petera , “Retinal blood vessel segmentation from diabetic retinopathy images using tandem PCNN model and deep learning based SVM”, https://doi.org/10.1016/j.ijleo.2019.163328.

[12] Kulwinder S. Mann and SukhpreetKaur , “Segmentation of retinal blood vessels using artificial neural networks for early detection of diabetic retinopathy”http://dx.doi.org/10.1063/

4981966.

[13] Jaakko Sahlsten1, Joel Jaskari1, Jyri Kivinen1, Lauri Turunen2, Esa Jaanio2, Kustaa Hietala3

&Kimmo Kaski1 , “Deep Learning Fundus Image Analysis for Diabetic Retinopathy and Macular Edema Grading, https://doi.org/10.1038/s41598-019-47181-w , 2019.

[14] M. Elena Martinez-Pkrez'; Alun D. Hughes2, Alice V. Stanton2, Simon A. Thorn2, Ani1 A.

Bharath' and Kim H. Parker' , “Segmentation of Retinal Blood Vessels Based on the Second Directional Derivative and Region Growing .”1999.

[15]Daniel ShuWei Ting, MD, PhD; Carol Yim-Lui Cheung, PhD; Gilbert Lim, PhD; Gavin SiewWei Tan, FRCSEd; Nguyen D. Quang, BEng, “Development and Validation of a Deep Learning Systemfor Diabetic Retinopathy and Related Eye Diseases UsingRetinal Images From Multiethnic Populations With Diabetes .”10.1001/jama.2017.18152.

[16] K.A. Vermeera;b;∗, F.M. Vosa; c, H.G. Lemijb, A.M. Vossepoela, “A model based method for retinal blood vessel detection ,”

[17]Songyuan Tang, Tong Lin, Jian Yang∗, Jingfan Fan, Danni Ai, and Yongtian Wang , “Retinal Vessel Segmentation Using Supervised Classification Based on Multi-Scale Vessel Filtering and Gabor Wavelet .” doi:10.1166/jmihi.2015.1565,2015.

[18] Xiaoyi Jiang, “Adaptive Local Thresholding by Verification-Based Multithreshold Probing with Application to Vessel Detection in Retinal Images”, 2003 .

19] Wong Li Yun a, U. Rajendra Acharya b,*, Y.V. Venkatesh a, Caroline Chee c,Lim Choo Min b, E.Y.K. Ng , “Identification of different stages of diabetic retinopathy using retinal optical

images .”doi:10.1016/j.ins.2007.07.020 , 2007 .

[20] Lei Zhang Mark Fisher Wenjia. Wang , “Retinal vessel segmentation using multi-scale textons derived from keypoints .”http://dx.doi.org/doi:10.1016/j.compmedimag.2015.07.006 , 2015.