Embed Size (px)
Citation preview
IMAGE & SIGNAL PROCESSING
An Intelligent Model for Blood Vessel Segmentation in Diagnosing DRUsing CNN
S. N. Sangeethaa1 & P. Uma Maheswari2
Received: 13 June 2018 /Accepted: 3 August 2018# Springer Science+Business Media, LLC, part of Springer Nature 2018
AbstractDiabetic retinopathy (DR) is an eye disease, which affects the people who are all having the diabetes for more than 10 years. Theophthalmologist identifies when the dilated eye exam causes severe in any one of the following deviations in the retina: changesin blood vessels, leaking blood vessels, newly grown blood vessels, swelling of the macula, changes in the lens, and damages tothe nerve tissue. It can eventually lead to vision loss. The early detection of DR prevents the cause of blindness. In this paper, wepropose the retinal image segmentation and extraction of blood vessels by morphological processing, thresholding, edge detec-tion, and adaptive histogram equalization. For the automatic diagnosis of DR from the fundus image, we also developed anetwork with the convolutional neural network architecture for accurately classifying its severity. By using high-end graphicalprocessor unit (GPU), we trained this network on the publicly available dataset such as DRIVE, DIARETDB0, andDIARETDB1_v1, and the images collected from the Aravind Eye Hospital, Coimbatore, India. Our proposed CNN achievesa sensitivity of 98%, a specificity of 93%, and an accuracy of 96.9% containing a database of 854 images.
Keywords Diabetic retinopathy . Macula . Convolutional neural network . Morphological processing
Introduction
DIABETES is also called as diabetes mellitus, which causeshigh glucose (sugar) in the blood for a prolonged period. If thediabetes is left untreated, it causes severe complications suchas eye, foot, heart problems, skin problems, hearing problems,stroke, diabetic neuropathy, and kidney problems. The diabet-ic retinal diseases include diabetic retinopathy (DR), diabeticmacular edema, and glaucoma. This paper describes onlyabout the severity of the DR which is diagnosed by usingconvolutional neural networks (CNNs).
Diabetic retinopathy is an eye disease which affects theblood vessels in the light-sensitive tissue called the retinawhich is found at the back of the eye. It is the well-knownproblem of visual deficiency and visual impairment amongworking-age grown-ups.
The different signs and indications of DR incorporatemicroaneurysms, leaking blood vessels, retinal swellings, de-velopment of new abnormal blood vessels, and damagednerve tissues. DR can be classified into five different stages:mild, moderate, severe, proliferative, or no disease; amongthat, it starts from mild non-proliferative abnormalities, inwhich the walls of the blood vessels are weakened and allowswelling. It progresses to moderate and severe non-proliferative diabetic retinopathy, which causes vessel abnor-malities. Later, it moves onto proliferative diabetic retinopa-thy, in which the growth of the new blood vessels is charac-terized in and around the retina. This paper proposes the de-tection of blood vessels by using various algorithms.
Besides, DR is the fundamental cause of visual impairmentin the mid-age population. Early detection of DR is a tediousprocedure for not withstanding a well-prepared clinician,which may bring about deferred treatment, miscommunica-tion, and so on. According to Lundquist et al. [1], in India,70% of the population staying in rural areas, and 10,000
This article is part of the Topical Collection on Image & SignalProcessing
* S. N. [email protected]
P. Uma [email protected]
1 Department of CSE, Velalar College of Engineering and Technology,Erode, India
2 Department of CSE, College of Engineering, Anna University,Chennai, India
Journal of Medical Systems (2018) 42:175 https://doi.org/10.1007/s10916-018-1030-6
ophthalmologists are available to care the entire population(i.e., the ratio of 1:100,000 people). To simplify the work ofthe ophthalmologist, neural network systems were introducedin most of the eye clinical laboratories.
Deep Learning (DL) is a subset of machine learning, whichuses the cascade of multiple layers for feature extraction andtransformation. The input of each layer uses the output fromprevious layer. In every layer, it will learn the features byitself. DL achieves higher level of accuracy ever than before.On the contrary, DL requires large amount of classified dataand high performance GPUs that has an ability to reducetraining time to hours or even less. DL models are trained bylarge amount of data, and the neural network directly learnsthe features from the data without any prior knowledge aboutdata from the feature extraction. In such case, we are using oneof the most popular neural network such as ConvolutionalNeural Networks, CNN. CNN consists of input, hidden andthe output layer. In hidden layer consists of alternating convo-lution and pooling operations to reduce the computation timeand build up spatial and configuration invariance; the last fewlayers (close to the outputs) will be fully connected one-dimensional layers and normalization layers. Each hiddenlayers are used to find the different feature of an image.
The current research in diagnosing DR has been explicitlybased on segmenting the features such as detection of bloodvessels and identifies the severity through classifier. Sincethere are more number of detection algorithms and classifiersare used. Specifically we have concentrated on CNN, inwhich, it does not need any prior knowledge about the featuresof the image, it can learn by itself at the time of training.
The contribution of this paper is fully exposed to the earlydetection of DR by preprocessing the fundus retinal image inwhich the blood vessels are segregated. We attempt to quan-tify retinal vessel damage through various image processingtechniques. To confirm our estimation, we introduced deeplearning algorithms to classify the image as normal or dis-eased. The preprocessed image is then feed into the trainedCNN to display the result. CNN can learn the features at thetime of training itself. The benefit of using trained CNN is toget the fast diagnosis and instant report than the expert toidentify the disease. Yearly retinal examination and early de-tection of DR can impressively decrease the risk of vision lossin diabetic people.
Priya and Aruna [2], Lim et al. [3], and Roychowdhury etal. [4] exposed by using machine learning techniques to clas-sify the detection of DR.
Normal Image Mild DR Moderate DR Severe DR Proliferative DR
a b c d e
Fig. 1 a–e Fundography images for various stages of diabetic retinopathy disease
Table 1 Overview of the existingmethod Author Segmentation Method Classifier
Roychowdhury et al. [9] Morphology-based method GMM classifier
Staal et al. [10] Ridge-based Knn classifier
Marín et al. [11] Moment invariants-based method Neural network classifiers
Soares et al. [12] Gabor wavelets GMM classifier
Fraz et al. [13] Gabor filters –
Franklin and Rajan [14] Gabor and moment invariants-based fea-tures
Feedforward network
Kavitha et al. [15] Levenberg–Marquardt based NN
Lupascu et al. [16] – Adaboost classifier
Liskowski et al. [17] – Deep neural network
Ramlugun et al. [18] Gabor filters and CLAHE –
Vasanthi and Wahida Banu[19]
Gray level and moment invariants method ELM
Bala [20] Morphology-based method McNN
Proposed Method CLAHE and Morphology-based method Convolutional NeuralNetwork
175 Page 2 of 10 J Med Syst (2018) 42:175
This paper aims at proposing a DR diagnosis model thatautomatically learns features, which are pivotal in diagnosingthe stage of the disease by training the neural network.
Fundus photography is the tool, which is used to take theback portion of the eye, including retina, optic disc, and mac-ula as specified in Singh et al. [5]. Fundus photography maybe utilized by trained medical experts for observing and ana-lyzing the severity of the disease. Figure 1 illustrates examplefundography images from a sample dataset for various stagesof DR disease.
Progresses in CNNs have aggravated in a modern tech-nique in image classification tasks, and also its variants startedto dominate numerous fields in computer vision which in-cludes object detection, image classification, object tracking,and edge detection. Xu et al. [6] and Doshi et al. [7] exposedthat CNN could gain a progressive structure from claimingfeatures, which might be utilized to image classification pur-poses. As well, the hierarchy approach will be accessible ongaining additional complex features, and in addition, transla-tion and distortion features in higher layers, the exactness ofthe CNN-based image classification method could a chance tobe higher.
Dependent upon this assumption, we investigate the utili-zation of the CNN-based strategy for the DR test in this work.In addition, a particular multi-layer CNN architecture is com-posed, and tests are conducted on true retinal information.Furthermore, the outcomes acquired show our suspicion aswe achieve 96.9% accuracy, which ranks as the highest com-pared with the existing feature-based classifiers.
This paper is organized as follows. Section BRelatedWorks^describes the literature reviews on the relatedworks that involvestages of DR and diagonize using neural networks, whileSection BMaterials and Methods^ describes the architecture ofthe CNN and the training methods used in this work. SectionBExperimental Evaluation and Results^ presents the resultsfrom our experiments. The conclusion and future work arediscussed in Section BDiscussion and Conclusion.^
Related works
Ophthalmologists can predict DR by the presence of lesions inthe retina, which is caused by the disease. The proposed ap-proach is effective because of lack of expertise and equipmentin some areas where the incidence of diabetes is high and DRdetection and treatment is almost often needed. The peoplewho affected by diabetes are increased, the ophthalmologistsare hurry to prevent the blindness, but the infrastructure andthe experts for DR become more insufficient.
In the past decade, the issue of automated segmentation ofretinal blood vessels has received under huge consideration.
Input Image
Preprocessing
Extracting the Blood
Vessels
Convolutional Neural
Network
Output
Morphological
Operation
Fig. 2 Workflow of proposed technique
Fig. 3 In DRIVE Database a Normal b Mild c Moderate d Severe. In Aravind Eye Hospital Database e Normal f Mild g Moderate h Severe
Fig. 4 Different scales in the training and testing dataset
J Med Syst (2018) 42:175 Page 3 of 10 175
Fraz et al. [8] have developed the methods of segmentation bycategorized into unsupervised and supervised. Supervisedsegmentation algorithms use a prior knowledge about theground truth of a training set of images, while unsupervisedalgorithms are trained online during segmentation. In the un-supervised method, related works use the algorithms such asmatched filtering, vessel tracking, morphological transforma-tions, contour model, Laplacian operator, and perceptivetransformation approach. The proposed work is based on thesupervised learning algorithms, by training the network usingthe collected datasets from public databases and the hospitalsby providing labels such as normal, mild, moderate, andsevere.
Roychowdhury et al. [9] describes that the blood vesselsand background areas are detected by the morphology-basedmethod, whereas the Gaussian mixture model (GMM) clas-sifies more number of pixels.
Staal et al. [10] implemented by using ridge-based vesseldetection, in which, every pixel is assigned to its nearest ridgeelement, so that we can partition the image. For each pixel, a27-feature set is evaluated and is implemented by using a kNNclassifier. This method works well in small size of the featureset whereas for large size of the feature set, this method slowsdown the process. The disadvantage of this method is that thetraining data are dependent and it is sensitive to false edges.
Marín et al. [11] proposes not a purely supervised algo-rithm, in which the network learns the features from a humanexpert. It also attains a good AUC. He introduced the inde-pendent of the training dataset and implements neural networkclassifiers using a seven-feature set which is extracted byneighborhood parameters as well as the moment invariants-based method.
Soares et al. [12] exposed to extract the six-feature set byusing Gabor wavelets and implemented using the GMM clas-sifier. Here, training data are also dependent, and it requiresmore time to train GMM models with a mixture of 20Gaussians.
For vessel classification, Fraz et al. [13] uses boostingand bagging strategies with 200 decision trees and a nine-feature set is extracted by using Gabor filters. The draw-back of using this strategy is that it attains high computa-tional complexity.
The proposed work for the vessel segmentation method byusing CNN in tensorflow is motivated by the method describedby Roychowdhury et al. [9] in which the blood vessels andbackground areas are detected by the morphology-based meth-od, whereas more number of pixels are classified by the GMM.It is proposed to increase the accuracy of the classifier accuracy,increase the number of features used for classification, and alsoincrease the number of training and testing images.
Franklin and Rajan [14] proposed a computer-based diag-nosis technique such as a feedforward network and uses backpropagation algorithm for the effective weight change in theabove neural network. Retinal blood vessels are detected byusing a multilayer perceptron neural network in which theinputs are extracted from Gabor and moment invariants-based features. Also, he detected the exudates present in theretinal images.
Kavitha et al. [15] investigated from the preprocessing step,which is processed by using the modified Gaussian filter forenhancing the input retinal image for normal/abnormal andexudates’ detection. Segmentation of region is performed byblood vessel extraction, optic disc extraction, and damage areaextraction. Features are extracted and trained by theLevenberg–Marquardt-based neural network and classify thenormal/abnormal, soft/hard exudates from the abnormal reti-nal images and microaneurysm/non-microaneurysm from theabnormal image. The proposed approach obtained a classifi-cation accuracy of 97.5% for normal/abnormal detection and94.12% for soft/hard exudates’ detection.
Lupascu et al. [16] applied an AdaBoost classifier and a 41-D feature vector, which consists based on the informationabout local intensity structure, spatial properties, and geome-try at different scales.
Fig. 5 Proposed workflow forblood vessel segmentation
Green
Channel
Extraction
CLAHE
Fig. 6 Green channel and CLAHE
175 Page 4 of 10 J Med Syst (2018) 42:175
Liskowski et al. [17] developed a deep neural network byusing unsupervised methods, to extract the vessel pixels fromthe fundus images.
Ramlugun et al. [18] employed a systematic approach fordiagnosing proliferative DR (PDR) by segmenting the retinalblood vessels from the digital fundus images. 2D-Match(Gabor) filters and contrast limited adaptive histogram equal-ization (CLAHE) are used to enhance the blood vessels and adouble-sided thresholding scheme is used to segment the ves-sels. This achieves high accuracy (93.1%) for the vessel ex-traction technique.
Vasanthi and Wahida Banu [19] extracted the blood vesselby using preprocessing, feature extraction, and classification.This extraction is carried out by extracting (i) nine features toform a feature vector and classification is performed by ELMand (ii) seven features based on gray level and moment invari-ants and ELM is used for classification, which obtained anaccuracy of 97.74%. Based on vessel segmentation, exudates,optic disc, and microaneurysms are detected.
Bala and Vijayachitra [20] used two approaches. First, todiagnosis the earlier stages of NPDR, the lesions are detectedby using morphological segmentation techniques and classi-fied using a new meta-cognitive neural network. It attainsgood classification accuracy. Second, to diagnosis PDR, theretinal images are preprocessed by using CLAHE and seg-mented by using modified local entropy thresholding opera-tion. The ELM classifier is applied to attain high sensitivityand specificity. Table 1 shows the overview of the variousexisting method for blood vessel segmentation.
In this paper, we increased the accuracy of the classifier byimproving the features using preprocessing and increasing thenumber of training and testing images.
Materials and methods
In this section, we described the proposed three-layer CNN forautomatic detection of DR. The workflow of our proposedtechnique is illustrated in Fig. 2.
First, the input fundus retinal image is preprocessed bygreen channel, and CLAHE is used to enhance the output,which comes from the green channel. Then, several morpho-logical operations and binarization thresholding are performedto extract the blood vessels from the retinal fundus image.Finally, the segmented blood vessel is tested under the pro-posed CNN, which is developed in python by usingtensorflow, to automatically detect the severity of the DR.
Datasets
The color retinal fundus images are downloaded from theDRIVE, Diaretdb0, and Diaretdb1_v1 from the website andAravind Eye Hospital, Coimbatore, India. The training datasetcontains 854 high-resolution images under a variety of imagingconditions. These retinal images are obtained from a group ofsubjects, and for each subject, two images were obtained for leftand right eyes, respectively. Of the 854 images in the dataset, 103are of patients with no pathologies (normal) and the rest of theimages are abnormal (contain pathologies such as exudates, cot-ton wool spots, microaneurysms, and hemorrhages). The labelswere provided by clinicians who rated the presence of DR ineach image by a scale of B0, 1, 2, and 3,^ which representBnormal,^ Bmild,^ Bmoderate,^ and Bsevere,^ respectively. Asmentioned in the description of the dataset, the images in thedataset come from different models and types of cameras, whichcan affect the visual appearance of left vs. right. Topcon trc-50dxretinal camera is used in the Aravind Eye Hospital to capture thefundus image. The sample images are shown in Fig. 3.
The training was performed for 562 images in the dataset,and for the testing module, the remaining 292 images (valida-tion data) were used.
Figure 4 shows counts of images for different scales in thetraining and testing dataset.
Blood vessel segmentation
Preprocessing
Preprocessing techniques are implemented to improve theoriginal image so that it can progress the rate of success ofthe proposed work. Before processing the segmentation of
Input Image Structuring
Element
Dilation/Eros
ion
Difference Image Vessel
Detection
Fig. 7 Flowchart for vessel detection
Fig. 8 Sample fundus image
J Med Syst (2018) 42:175 Page 5 of 10 175
blood vessels, the preprocessing techniques are applied so thatthe characteristics of the original image are not changed. Toreduce these imperfections and produce more suitable imagesfor extracting the features, a preprocessing consists of fivesteps: (1) green channel extraction, (2) contrast enhancement,(3) morphological operations, (4) background exclusion, and(5) thresholding and contour detection. The flow diagram forthe blood vessel segmentation technique is shown in Fig. 5.
Based on the trial-and-error method, the methodologies areused to work well for the retinal images. The brief explanationof each step is given below:
& Step 1: Green channel
For the segmentation of blood vessels, the green channel ofthe RGB image is chosen because the red and blue channelssuffer from noise and poor image quality.
To enhance the contrast of the retinal images, the red andblue components of the image are discarded before processing.In order to display the best vessel and greatest contrast betweenthe optic disc and the retinal tissue, the green channel is used inthe preprocessing step. Thus, the green pixel values are extract-ed from the input image and stored in the matrix form.
& Step 2: Contrast enhancement
Contrast-limited adaptive histogram equalization is a con-trast enhancement technique, which splits the entire regioninto a number of tiny sections of similar size and works oneach region such that thecontrast of each small region is im-proved. CLAHE is applied on the green channel, which sup-ports to acquire the important blood vessel information.Figure 6 shows how the CLAHE is applied to a fundus image.
Morphological processing
Mathematical morphology is used to examine the interactionbetween an image and a certain chosen structuring element (SE)using the basic operations of erosion and dilation. The process-es mainly used here are dilation, erosion, opening, and closing.
It is processed by combining two sets of pixels, in whichone set consists of the SE or kernel and the other constitutesthe image being processed. The two main transformations areopening and closing. In addition, dilation and erosion expandan image object and shrink it, respectively. The importantaspect of these two operations is the SE, which is a matrixconsisting of only 0 s and 1 s of any size and shape, used toexplore the input image. Themorphological opening and clos-ing operations are applied to smoothen the edge of the vesseland smoothen the region, eliminate small holes, and fill thegaps in the contours. In the proposed work, algorithms com-bine the above processes, to detect the edge, remove the noiseand background, and find specific shapes in images. The ves-sel detection flowchart is shown in Fig. 7.
Background subtraction
It is used to eliminate the background variations from an im-age so that the foreground objects may be easily analyzed. Inthe proposed work, the background elimination is performedby subtracting the morphological image from the contrast im-age enhancement.
Thresholding
Thresholding is the process of removing unnecessary detailsfrom an image. In the fundus images, the blood vessels are
Fig. 9 a Contrast image enhancement. b After performing morphological operation. c Background subtraction. d Image enhancement. e Thresholdingbinarization
Fig. 10 a After masking. b After binarization. c After erosion. d Contour identification. e After bitwise operation
175 Page 6 of 10 J Med Syst (2018) 42:175
reduced to binary pixels by removing all gray level informa-tion. It is important to differentiate the foreground blood ves-sels and the background information. To bring out hiddendetails, thresholding is used. Consequently, selecting a properthreshold value is important, because a low value may de-crease the size or reduce the number of these objects and ahigh value may include other background information.
It is used to produce a binary image inwhich the pixel value iseither 1 (blood vessel) or 0 (background) so that the numericaldata are then applied to various machine learning algorithms.
Contour identification
As a result of the thresholding, unnecessary pixels are consideredas noise (false positive) in the resultant binary image, and conse-quently, post-processing methods are applied for refining the im-age and retaining the desired objects. In this case, a morphologicaloperation (erosion) is applied to remove the undesired objects. Thebright circle present in the edge of the retina is also removed bysubtracting a mask image from the binary image. Figure 8 repre-sents the retinal fundus image, which is to be preprocessed; Figs. 9and 10 shows each step performed in the preprocessing stage.
Network model and training
Convolutional neural networks
The CNN is a deep learning architecture, in which individualneurons are tiled such that they respond to overlapping regionsin the visual fields. It consists of alternating convolution andpooling operations to reduce the computation time and build up
spatial and configuration invariance; the last few layers (close tothe outputs) will be fully connected one-dimensional layers.Figure 11 shows the architecture of the proposed CNN.
In the proposed work, the CNN architecture consists ofthree convolutional layers followed by rectified linear units(RLUs) and softmax. The final layers of the network consistof a fully connected layer and a final soft-max classificationlayer. The first two convolutional layers use 32 small-sizefilters and the third layer uses 64 filters.
The output which comes from the third convolutional layeris a multi-dimensional tensor. However, the flattened layer isused to convert this into a one-dimensional tensor by using thereshape operation.
The two fully connected layer is used which consists of 128neurons, in which it take all the inputs, perform the standardz =wx + b operation on it and also one of the fully connectedlayer added a non-linearity (RELU) to it and the other fullyconnected added a RELU to it.
Softmax, which uses class entropy to generate the output,ranges between 0 and 1, indicating the probability of the pixelfor the positive class.
A single GTX 1080 Ti GPU has only 3 GB of memory,which limits the maximum size of the networks that can betrained on it. In our experiment, 562 labeled images are usedto train the neural network, while 292 images are used toevaluate the performance of the trained neural network.
Problem formulation
The trained CNN model is chosen to perform prediction onthe unseen test images. The trained model is capable of com-paring the test image with each class and determining its prob-ability value. The class with the highest probability is consid-ered the resultant prediction of the test image.
1st Convolutional LayerFilter Size: 3*3No. of Filter: 32
2nd Convolutional LayerFilter Size: 3*3No. of Filter: 32
3rd Convolutional LayerFilter Size: 3*3No. of Filter: 64
FlattenedFully Connected LayersNeurons: 128
Softmax
Input Image
Fig. 11 CNN architecture
Table 2 Description of DRIVE, DIARETDB0, DIARETDB1_v1, andreal-time datasets
Database Normal Mild Moderate Severe Total
DRIVE 25 38 – – 63
Diaretdb0 – 31 50 9 90
Diaretdb1_v – 23 40 4 67
Real time dataset 78 194 106 256 634
Total 103 286 196 269 854
Table 3 Number oftraining and testingimages
Classes Training Testing
Normal 28 75
Mild 212 74
Moderate 124 72
Severe 198 71
Total 562 292
J Med Syst (2018) 42:175 Page 7 of 10 175
In order to determine the accuracy, A
A ¼ Number of images with correctly predicted classes
Total number of test images
� �� 100
For cost optimization, Tensorflow uses AdamOptimizer forgradient calculation and weight optimization. It is used tominimize the cost with a learning rate of 0.0001.
To calculate the cost, the proposed network (three convolution+ one flattening + two fc layers) will have to execute to producelayer_fc2 which is required to calculate cross_entropy, such ascost. To calculate the cost, we use
opt ¼ tf :train:AdamOptimizer lrate ¼ 1e−4ð Þ:min costð Þ
Since there are few changes in the gradients and weightsand are very computationally expensive. To calculate the ac-curacy on the validation set using true labels and predictedlabels such as (y_true) and (y_pred), respectively,
ct prediction ¼ tf :equal y pred cls; y true clsð Þ
acc ¼ tf :red mean tf :cast ct prediction; tf :flaot32ð Þð Þ
Experimental evaluation and results
Estimation of parameters
The proposed work has been explored in Tensorflow. The pro-posed work uses the publicly available database such as DRIVE,Diaretdb0, and Diaretdb1_v1 from the website, and the real timeimages from Aravind Eye Hospital, Coimbatore, India. In thesedatabases, we classified the images according to normal, mild,moderate, and severe. Consequently, the fundus images aretrained one by one in the CNN as per the classification.
The proposed system’s performance is evaluated by usingthe validation accuracy. The training and testing (validationdata) were performed for four classes of retina images (nor-mal, mild, moderate, and severe) in the collected dataset.
For evaluating the performance of the proposed system, 90images in DIARETDB0 database , 67 images inDIARETDB1_v1, 63 images in DRIVE database, and 634 realtime images are taken. With these databases, we have occurred103 normal images, 286 mild images, 196 moderate images, and269 severe images as shown in Table 2.
According to the classes, the images are separated for train-ing and testing as shown in Table 3. The training and testingimages taken for mild class are 212 and 74, respectively. Fornormal classes, the training and testing images are 28 and 75,respectively. For moderate and severe, the training and testingimages are taken as 124 and 72, 198, and 71, respectively.
Performance evaluation
The proposed system is examined on retinal images ofDRIVE, Diaretdb0, Diaretdb1_v1, and real-time dataset.
Fig. 12 Performance measure of CNN classifiers
Table 5 Evaluationmetrics Evaluation metrics Values
Sensitivity 0.9815
Specificity 0.9333
Accuracy 0.9691
Precision 0.9770
F-Score 0.9793
Table 4 Computation ofstatistical parameters Parameters No. of images
TP 213
TN 70
FP 5
FN 4
Table 6 Comparison of severity level identified by experts and theproposed network
Image No. Severity level(annotated)
Severity level (byproposed system)
DR074 Mild Mild
DR025 Moderate Moderate
DR118 Normal Normal
DR256 Normal Normal
DR569 Severe Severe
DR678 Moderate Moderate
DR476 Mild Normal
DR780 Mild Mild
DR801 Severe Moderate
DR642 Severe Severe
175 Page 8 of 10 J Med Syst (2018) 42:175
Different preprocessing, segmentation techniques, andthresholding operations are applied for vessel extraction.Crucial parameters such as accuracy (Acc), specificity(SP), and sensitivity (SE) play an important role in mea-suring the performance of DR.
Sensitivity is a measure that determines the probabilityof the patients having DR. It is calculated by the ratio ofthe number of correctly classified DR images to the totalnumber of DR patients.
True positive (TP) denotes the number of correctlydetected DR images and false negative (FN) denotesthe number of DR images detected wrongly as non-DR images.
Sens ¼ TP
TPþ FN
Specificity is a measure that determines that the personis not affected by the disease. It is the ratio of the numberof correctly classified normal images to the number ofnormal images. True negative (TN) refers to the numberof normal images identified correctly and false positive(FP) refers to the number of normal images that are iden-tified as DR images.
Spec ¼ TN
TNþ FP
Accuracy is the ratio of the total number of images identi-fied correctly to the total number of images taken for theclassification.
Acc ¼ TPþ TN
TPþ TNþ FPþ FN
Precision is the ratio of correctly classified DR images tothe number of non-DR images detected wrongly as he DRimages.
Precision ¼ TP
TPþ FP
The test performances of the CNN classifier are determinedby the computation of the statistical parameters shown inTable 4.
The evaluation metrics for the proposed classifier areshown in Table 5.
The average accuracy values obtained using CNN classi-fiers for the above-mentioned database 96.9% and the preci-sion values and F-score obtained are 97.7 and 97.9%.Sensitivity and specificity values are achieved as 98 and93%, respectively. The evaluation metrics and graphical rep-resentation of performance measure are shown in Table 4 andFig. 12.
A confusion matrix shown in Table 7 is used to describe theperformance of a classification model on a set of test data forwhich the true values are known. It allows the visualization ofthe performance of an algorithm.
The average learning rate and validation size obtained forall the datasets in the proposed method are 0.01 × 10−4 and0.2, respectively. From the above results, it is observed that theproposed CNN classifier provides a better classification accu-racy and learning rate compared to all existing classifiers usedin the existing work.
Table 6 shows the severity level of the fundus image of theproposed work, which is compared with the severity levelgiven by the experts.
Table 7 Confusion matrix
Class Normal Mild Moderate Severe
Normal 70 4 1 0
Mild 1 72 1 0
Moderate 0 0 71 1
Severe 0 0 1 70
Table 8 Statistical analysis andperformance comparison existingsystem
Author Se Sp Acc Time System
Staal et al. – – 0.94 15 min 1.0GHz, 1-GB RAM
Soares et al. – 0.94 3 min 2.17GHz1-GB RAM
Lupascu et al. 0.72 – 0.95 – –
Marín et al. 0.70 0.98 0.94 90 s 2.13GHz2-GB RAM
Franklin et al. – – 0.95 – –
Roychowdhury et al. 0.84 0.72 0.95 3.11 s 2.6GHz, 2-GB RAM
Kavitha et al. – – 0.93 – –
Bala et al. 100 0.90 0.96 – –
Liskowski 0.88 0.78 0.98 – NVIDIA GTXTian GPU
Proposed method 0.98 0.93 0.96 120 s NVIDIA Geforce GTX 1080Ti
J Med Syst (2018) 42:175 Page 9 of 10 175
Here, the image DR476 is actually affected bymild DR andthe image DR801 is affected by severe DR, but the proposednetwork has wrongly given the result as DR476 is normal andDR801 is moderate.
The statistical analysis and performance comparison of theexisting system and the proposed are shown in Table 8.
Conclusion
Our work has shown that the automated diagnosis of identify-ing DR by segmenting the blood vessels and classified usingCNN. The benefit of using trained CNN is to get the fastdiagnosis and instant report than the expert to identify thedisease. There is no issues in classifying the normal and severeimages, but there is an issue in classifying the DR as mild andmoderate at the earlier stage. The significance of this work isto identify the disease in earlier stage with less time and cost.The proposed CNN has successfully worked for accuratelyidentifying the DR earlier to prevent the vision loss. Itachieves an accuracy of about 96.7%, a sensitivity of about98%, and a specificity of about 93%.
Future work
For any diabetes related eye disease, vessel tortuosity is con-sidered as an important factor. In future work, the changes invessel tortuosity and few texture features such asmicroaneurysms, haemorrhages, exuadates are examined todiagnose the severity of the DR disease. Our future work isalso to enhance the CNN by training with more number ofimages to get the better result.
Compliance with Ethical Standards
Conflict of Interest No potential conflict of interest was reported by theauthors.
Ethical approval This article does not contain any studies with humanparticipants or animals performed by any of the authors.
References
1. Lundquist, M. B., Sharma, N., and Kewalramani, K., Patient per-ceptions of eye disease and treatment in Bihar India. J. Clin. Exp.Ophthal. 3(2):1000213, 2012.
2. Priya, R., and Aruna, P., SVM and neural network based diagnosisof diabetic retinopathy. Int. J. Comput. Appl. 41(1):6–12, 2012.
3. Lim, G., Lee, M. L., Hsu, W., and Wong, T. Y., Transformed rep-resentations for convolutional neural networks in diabetic
retinopathy screening. Québec City: AAAI Workshop on ModernArtificial Intelligence for Health Analytics (MAIHA), 2014.
4. Roychowdhury, S., Koozekanani, D. D., and Parhi, K. K.,DREAM: Diabetic retinopathy analysis using machine learning.IEEE Journal of Biomedical and Health Informatics 18(5):1717–1728.
5. Singh, R., Dubey, S., andMittal, U. Blood vessel segmentation andmicroaneurysm detection for diabetic retinopathy. retina-features/README.md, 2017.
6. Xu, K., Feng, D., and Mi, H., Deep convolutional neural network-based early automated detection of diabetic retinopathy using fun-dus image. Molecules 22:2054, 2017.
7. Doshi, D., Shenoy, A., Sidhpura, D., and Gharpure, P., Diabeticretinopathy detection using deep convolutional neural networks.Pune: International Conference on Computing, Analytics andSecurity Trends (CAST), 2017, 261–266.
8. Fraz, M., Remagnino, P., Hoppe, A., Uyyanonvara, B., Rudnicka,A., Owen, C. et al., Blood vessel segmentation methodologies inretinal images a survey. Comput. Methods Prog. Biomed. 108(1):407–433, 2012a.
9. Roychowdhury, S., Koozekanani, D. D., and Parhi, K. K., Bloodvessel segmentation of fundus images by major vessel extractionand subimage classification. IEEE J. Biomed. Health Information19(3):1118–1128, 2015.
10. Staal, J. J., Abramoff, M. D., Niemeijer, M., Viergever, M. A., andvan Ginneken, B., Ridge based vessel segmentation in color imagesof the retina. IEEE Trans. Med. Imaging 23(4):501–509, 2004.
11. Marín, D., Aquino, A., Gegúndez-Arias, M. E., and Bravo, J. M., Anew supervised method for blood vessel segmentation in retinalimages by using gray-level and moment invariants-based features.IEEE Trans. Med. Imaging 30(1):146–158, 2011.
12. Soares, J., Leandro, J., Cesar, R., Jelinek, H., and Cree, M., Retinalvessel segmentation using the 2-D Gabor wavelet and supervisedclassification. IEEE Trans. Med. Imaging 25(9):1214–1222, 2006.
13. Fraz, M., Remagnino, P., Hoppe, A., Uyyanonvara, B., Rudnicka,A., Owen, C. et al., An ensemble classification-based approachapplied to retinal blood vessel segmentation. IEEE Trans.Biomed. Eng. 50(9):2538–2548, 2012b.
14. Franklin, S. W., and Rajan, S. E., Computerized screening of dia-betic retinopathy employing blood vessel segmentation in retinalimages. Biocybernetics and Biomedical Engineering 34(2):117–124, 2014 (ELSEVIER).
15. Kavitha, M., and Palani, S., Blood Vessel, Optical Disk andDamage Area-Based Features for Diabetic Detection from RetinalImages. Arab. J. Sci. Eng. 39(10):7059–7071, 2014.
16. Lupascu, C. A., Tegolo, D., and Trucco, E., Fabc: Retinal vesselsegmentation using AdaBoost. IEEE Trans. Inf. Technol. Biomed.14(5):1267–1274, 2010.
17. Liskowski, P., and Krawiec, K., Segmenting Retinal Blood VesselsWith Deep Neural Networks. IEEE Trans. Med. Imaging 35(11):2369–2380, 2016.
18. Ramlugun, G. S., Nagarajan, V. K., and Chakraborty, C., Smallretinal vessels extraction towards proliferative diabetic retinopathyscreening Expert Syst. Appl. 39:1141–1146, 2012.
19. Vasanthi, S., and Wahida Banu, R. S., Automatic segmentation andclassification of hard exudates to detect macular edema in fundusimages. J Theor. Appl. Inform. Technol. 66(3):684–690, 2014.
20. Bala, M. P., and Vijayachitra, S., Extraction of retinal blood vesselsand diagnosis of proliferative diabetic retinopathy using extremelearning machine. J. Med. Imag. Heal. Inform. 5(2):248–256, 2015.
175 Page 10 of 10 J Med Syst (2018) 42:175
本文献由“学霸图书馆-文献云下载”收集自网络，仅供学习交流使用。
学霸图书馆（www.xuebalib.com）是一个“整合众多图书馆数据库资源，
提供一站式文献检索和下载服务”的24 小时在线不限IP
图书馆。
图书馆致力于便利、促进学习与科研，提供最强文献下载服务。
图书馆导航：
图书馆首页 文献云下载 图书馆入口 外文数据库大全 疑难文献辅助工具
