Research ArticleComputer-Assisted Diagnosis for Diabetic Retinopathy Based onFundus Images Using Deep Convolutional Neural Network

Yung-Hui Li 1 Nai-Ning Yeh1 Shih-Jen Chen 2 and Yu-Chien Chung3

1Department of Computer Science amp Information Engineering National Central University Taoyuan 32001 Taiwan2Department of Ophthalmology Taipei Veterans General Hospital School of Medicine National Yang-Ming University BeitouTaipei Taiwan3Department of Ophthalmology Fu Jen Catholic University Hospital New Taipei City Taiwan

Correspondence should be addressed to Yung-Hui Li liyunghuiicloudcom

Received 27 April 2018 Revised 28 June 2018 Accepted 27 November 2018 Published 2 January 2019

Guest Editor Antonio Celesti

Copyright copy 2019 Yung-Hui Li et al is is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Diabetic retinopathy (DR) is a complication of long-standing diabetes which is hard to detect in its early stage because it onlyshows a few symptoms Nowadays the diagnosis of DR usually requires taking digital fundus images as well as images usingoptical coherence tomography (OCT) Since OCT equipment is very expensive it will benefit both the patients and the oph-thalmologists if an accurate diagnosis can be made based solely on reading digital fundus images In the paper we present a novelalgorithm based on deep convolutional neural network (DCNN) Unlike the traditional DCNN approach we replace thecommonly used max-pooling layers with fractional max-pooling Two of these DCNNs with a different number of layers aretrained to derive more discriminative features for classification After combining features from metadata of the image andDCNNs we train a support vector machine (SVM) classifier to learn the underlying boundary of distributions of each class Forthe experiments we used the publicly available DR detection database provided by Kaggle We used 34124 training images and1000 validation images to build ourmodel and tested with 53572 testing imagese proposed DR classifier classifies the stages ofDR into five categories labeled with an integer ranging between zero and four e experimental results show that the proposedmethod can achieve a recognition rate up to 8617 which is higher than previously reported in the literature In addition todesigning amachine learning algorithm we also develop an app called ldquoDeep Retinardquo Equipped with a handheld ophthalmoscopethe average person can take fundus images by themselves and obtain an immediate result calculated by our algorithm It isbeneficial for home care remote medical care and self-examination

1 Introduction

e global cost of treating adult diabetes and its inducedchronic complications is USD 850 billion in 2017 Diabeticretinopathy (DR) is one of the most common and seriouscomplications of diabetes mellitus and is a leading cause oflow vision and blindness in working-age adults [1 2] eInternational Diabetes Foundation (IDF) estimated that theglobal population with diabetes in 2017 was 451 million andover one-third of the population had DR [3] representing atremendous population at risk of visual impairment orblindness By 2045 the worldwide prevalence of diabetes isexpected to increase to 693 million people [3] In additionalmost half (497) of all people living with diabetes remain

undiagnosed for years because of silent symptoms [3]However long-term high blood sugar levels ultimatelydestroy blood vessels and nerves leading to complicationssuch as cardiovascular disease and blindness Detection andtreatment of DR in the early stage will prevent its devel-opment or progression

e diagnosis and severity of DR are based on retinalexamination Clinically the classification of DR can be di-vided into two categories (1) nonproliferative diabeticretinopathy (NPDR) with exudation and ischemia in dif-ferent severity but without retinal neovascularization and(2) proliferative diabetic retinopathy (PDR) which ischaracterized by neovascularization with or without itscomplications of traditional retinal detachment and the

HindawiMobile Information SystemsVolume 2019 Article ID 6142839 14 pageshttpsdoiorg10115520196142839

initial appearance of vitreous hemorrhage Microvasculardiseases of NPDR include microaneurysms retinal dot andblot hemorrhages lipid exudates venous beading changeand intraretinal microvascular abnormalities (IRMA) Basedon the degree and extent of these lesions NPDR can bedivided into three levels mild NPDR presents withmicroaneurysms or few retinal hemorrhages moderateNPDR shows more severe microaneurysms hemorrhage orsoft exudate but not reaching the level of severe NPDRwhich is associated with marked retinal hemorrhage in 4quadrants venous beading in at least 2 quadrants and IRMAin at least 1 quadrant Table 1 summarizes the DR categorywith its manifestation

Manual grading by ophthalmologists has been themainstay of DR screening in the past decades However dueto the expanding population with diabetes and the recentadvances in technology automated detection of DR offersthe potential to provide an efficient and cost-effective ap-proach to screening Current commercialized automatedretinal image analysis systems (ARIAs) such as iGradingMRetmarker and EyeArt focus on differentiating diseasednodisease or detection of referable DR [5 6] NonethelessARIAs are currently not sufficiently sophisticated to classifydifferent levels of DR which means that identifying thesubtle change between levels is still a challenging task for thetechnique of medical image analysis Figure 1 shows examplefundus images for each lesion

In addition to the accuracy of medical image pro-cessing the mobility and portability of medical examina-tion equipment are of equal importance Currently theacquisition of digital fundus images requires the cooper-ating patient to sit in front of the fundus camera in theroom with ambient lighting minimized or turned off epatient needs to look forward at the camera at a fixed lightand use infrared fundus imaging to focus on the area ofinterest Many nonmydriatic cameras have software thatautomatically detects the posterior pole of the eye and takesa picture when it is focused behind the eye e RGB imagesensor still requires a flash to capture images in the visiblelight spectrum However the digital fundus imagers mostpopularly used in the clinics are bulky and expensive asshown in Figure 2 which limit its capability for large-scalescreening

One of the major goals for this study besides increasingthe classification accuracy using artificial intelligence is tocome up with a new system framework for DR screeninge new framework combines the advantages of mobilecomputing cloud computing big data and artificial in-telligence e components of the proposed framework canbe described as following

(i) Mobile block e fundus image acquisition isachieved using a hand-held fundus imager coupledwith a self-developed iPhone APP e imager issmall and light-weighted It can be carried inside abackpack e deployment of such devices is ex-tremely convenient e portable nature due to itssmall form factors and light-weight can benefit themedical service for remote rural areas

(ii) Cloud blocke proposed system does not sacrificeits computational performance for its portabilityanks to the architecture of cloud computing thecore of the computational resources is moved to thecloud and can be scaled up flexibly as the requestincreases We developed highly efficient deeplearning algorithm which runs on cloud server andis able to respond to the diagnosis request within 10seconds

(iii) Big data block e cloud-based architecture alsohelps to collect big data As more and more enddevices (hand-held fundus camera) are being usedthe number of fundus images that passed into thecloud will increase accordingly By storing all ofthese fundus image data we are able to make gooduse of such big dataset such as machine learningmodel retraining new feature exploration or cross-domain data-mining for different types of ophal-mological diseases

In summary in this paper we propose a new systemframework for DR screening based on artificial intelligencemobile computing cloud computing and big data analyticsFigure 3 shows an illustration of the proposed system Suchsystem is a new paradigm for telemedical service and willbenefit rural areas where the medical resources areinsufficient

In the following sections we will gradually unveil ourideas and show the experimental results In Section 2 weperformed related literature review for important algorithmsfor the foundation of DR classification which is retinal vesselsegmentation In Section 3 we performed related literaturereview for DR detection In Section 4 we illustrated theproposed deep learning and machine learning algorithms infull details We show the experimental results in Section 5and discuss some important findings in Section 6 Finally aconclusion is given in Section 7

2 Literature Review of RetinaVessels Segmentation

In the process of identifying DR it is pivotal to locate theretinal vessels If the vessels position can be correctly knownwe can determine whether the patient is suffering from DRbased on information about the precise location andthickness of the vessels However vessel tracking is acomplex process because of the many other substancesbesides vessels in fundus images Numerous vessel seg-mentation methods have been proposed which can bebroadly divided into five categories vascular trackingmatched filtering morphological processing deformationmodels and machine learning

21 Methods of Vascular Tracking Methods of vasculartracking are based on the continuous structure of vessels bystarting at an initial point and following the vessels until nofurther vessels are founde critical factor in this procedureis the setting of the initial point as this will affect the

2 Mobile Information Systems

accuracy of vessel segmentation Currently setting the initialpoint can be done either artificially or automatically

e earliest adaptive vascular tracking method wasproposed by Liu and Sun [7] in 1993 which extracts thevasculature from X-ray angiograms First given an initialpoint and direction within a vessel the authors apply anldquoextrapolation-updaterdquo scheme that involves estimatinglocal vessel trajectories Once a vessel fragment has beentracked it is removed from the image is procedure isrepeated until the vascular tree has been extracted edrawbacks of this strategy are that due to the algorithm usedthe user must set the vessel starting points and that theapproach does not seem adaptable to three-dimensionalextraction In 1999 an automatic vascular trackingmethod was developed by Can et al [8] is strategy mainlycollects pixel wide vascular local minimum points (usually inthe middle of a vessel) to perform tracking Vlachos andDermatas [9] suggested a multiscale line tracking methodwith morphological postprocessing Yin et al [10] proposeda retinal vascular tree extraction based on iterative trackingand Bayesian method

e advantage of vascular tracking is that it can providelocal information about characteristics such as the

diameterwidth and direction of vessels However thevascular tracking performance can be easily affected bycrossing or branching of vessels which reduces the iden-tification efficiency

22 Methods of Matched Filtering Matched filteringmethods employ multiple matched filters for extraction sodesigning proper filters is essential to detect vessels Since thegray-scale distribution of fundus vessels is in keeping withGaussian an intuitive method exists that uses the maximumresponse of images after filtering to find vessel points As thediameterwidth of vessels is diverse a multiscale Gaussianfilter method is often used for vessel tracking

In 1989 Chaudhuri et al [11] pioneered the applicationof Gaussian filters in vessel tracking by using some vascularcharacteristics such as the fact that vessels are darker thanthe background the width of the vessels ranges from 2 to 10pixels and the vessels grow from the optic disc into a radialshape erefore Chaudhuri et al [11] designed two-dimensional Gaussian filters that can detect vessels in 12different directions However this method needs largecomputation and some of the dark lesions are similar to the

Table 1 Classification of diabetic retinopathy [4]

Category Level Manifestation

Nonproliferative diabetic retinopathy (NPDR)

1 Few microaneurysms or hemorrhage

2 Microaneurysms hemorrhages soft exudates venousbeading

3Severe retinal hemorrhage in ge4 quadrants or venousbeading change in ge2 quadrants or mild intraretinal

microvascular abnormality in ge1 quadrantProliferative diabetic retinopathy (PDR) 4 Neovascularization of retina or optic disc

(a) (b) (c)

(d) (e)

Figure 1 Examples of fundus images showing different lesions (a) 0 level (normal) (b) level 1 (c) level 2 (d) level 3 and (e) level 4

Mobile Information Systems 3

characteristic of vessels causing tracking errors Hooveret al [12] described an improvedmethod that considers localand regional characteristics of vessels to separate bloodvessels in retinal images and iteratively determine whetherthe current point is a vessel point

After such improvement a large number of studies ofreformed filters have been developed Jiang and Mojon [13]promoted a generalized threshold method based on amultithreshold detection Zhang et al [14] improved thematching filtering method by applying a local vessel crosssection analysis using local bilateral thresholding Li et al[15] suggested a multiscale production of the matched filterto enhance the extraction of tiny vessels

23 Methods of Morphological Processing Morphologicalprocessing facilitates the segmentation and identification oftarget objects by analyzing and processing structural ele-ments in a binary image us linear and circular elementsof blood vessels can be selected isolating the desiredstructure instead of the background image In addition

morphological processing can also smooth and fill the imagecontour with the advantage of antinoise However thismethod overrelies on structural elements and does not makegood use of characteristics of vessels

According to vessel characteristics Zana and Klein [16]introduced a mathematical morphology-based algorithmthat allows separating the vessels from all possible un-desirable patterns Building on this approach Ayala et al[17] proposed using different average fuzzy sets In Miri andMahloojifar [18] fundus images were analyzed by the use ofcurvelet transform and morphological reconstruction ofmultistructural elements to enhance the boundaries anddetermine the vascular ridge Karthika and Marimuthu [19]combined curvelet transform and morphological re-construction of multistructural elements with stronglyconnected component analysis (SCCA) to segment andidentify vessels

24 Methods of Deformation Models First introduced byKass et al [20] in 1988 the key benefit of deformation

Level 0Level 1Level 2Level 3Level 4

Level 4

Big datacollection

Onlinetraining

Deep learningalgorithm on

cloud computingplatform

Image acquisitionon mobile device

Figure 3 e proposed framework It is composed of mobile block (image acquisition) cloud block (deep learning algorithm on cloudcomputing platform) and big data block (big data collection and training)

Figure 2 Commercialized fundus camera used in the clinics e machine is bulky and expensive making it difficult for large-scalescreening

4 Mobile Information Systems

models is the ability to produce smooth parametric curves orsurfaces Two categories of deformation models are iden-tified parametric deformation and geometric deformationParametric deformation models are also called active con-tour or snake models (set of points each with an associatedenergy) rough the external and internal forces acting onthe snake the snake model can change its shape andsmoothness toward the desired structure In 2007 Esponaet al [21] used a parametric deformation model method onfundus images and promoted an improved method withmorphological segmentation With the assistance of mor-phological vessel segmentation the snake model expands tothe contour of the obtained vessels until the local energyfunction is minimal Another deformation model methodcalled ribbon of twins (ROT) which combines ribbon snakesand double snakes was proposed by Al-Diri and Hunter[22] Each twin consists of two snakes one inside and oneoutside the vessel edges e double snake model then at-tempts to integrate the pairs of twins on the vessel bordersinto a single ribbon and calculate the vessel width

ere are several shortcomings in parametric de-formation models For instance the segmentation resultsdepend on the initial contour and difficulties arise whenextending from low to high dimensions and in segmentingcomplex objects Geometric deformation can well solve theproblems caused by parametric deformable models Geo-metric deformable models are based on deformation curveevolution theory and have no strict requirement on theposition of the initial contour which increases the robust-ness of the method and allows it to be extended to highdimensions Zhang et al [23] proposed an automatic vesselsegmentation method which uses nonlinear orthogonalprojection to capture the characteristics of retinal bloodvessels and obtained an adaptive local thresholding algo-rithm for blood vessel segmentation Zhao et al [24] sug-gested a retinal vessel segmentation method that employs aregion-based active contour model with a level set imple-mentation and a region growing model

25 Methods of Deformation Model Machine learning is analgorithm that teaches computers to learn to achieve goalsautomatically by building generative or discriminativemodels from accumulated datasets Machine learning can bedivided into supervised learning and unsupervised learninge supervised learning methods learn to achieve goalsbased on ground-truth which means that during thetraining stage the training data used to train the model comewith a ldquolabelrdquo that can be used by the machine learningalgorithm to differentiate the data Applying such paradigmin the problem of DR it means that when using supervisedlearning one needs to mark all of the pixels belonging tovessels in advance whereas the unsupervised learningmethod does not need to mark them beforehand

For supervised learning Cesar and Jelinek [30] andLeandro et al [31] proposed a supervised classification withtwo-dimensional Gabor wavelet Each pixel has a featurevector that consists of the gray-scale feature and responses ofdistinct sizes of two-dimensional Gabor wavelet Ricci and

Perfetti [25] proposed a segmentation method for retinalvessels based online manipulation and support vectorclassification Since the features are extracted by twoorthogonal vertical lines it reduces the features andtraining samples in supervised learning A supervisedmethod using neural network was proposed by Marinet al [26] which has one input layer three hidden layersand one output layer Each pixel in the image is repre-sented by a seven-dimensional feature vector to train thenetwork Shanmugam and Banu [27] used an extremelearning machine (ELM) to detect retinal vessels bycreating a seven-dimensional feature vector based ongray-scale features and invariant moments and using ELMto segment vessels In 2015 Wang et al [28] raised a newhierarchical retinal vascular segmentation includingthree steps preprocessing hierarchical feature extractionand integration classification It involves using simplelinear iterative clustering (SLIC) to perform super-pixelsegmentation and randomly selecting a pixel to representthe entire super-pixel as a more easy and efficient meansof extracting features

For unsupervised learning in 1998 Tolias and Panas[32] created an automatic and unsupervised segmentationmethod based on blurred fundus images which used fuzzyC-means (FCM) to find initial candidate points Xie and Nie[33] proposed a segmentation method based on a geneticalgorithm and FCM Salazar-Gonzalez et al [29] usedmethods of vector flow to segment retinal vessels

Table 2 is a summary about the performance comparisonbetween different existing methods

3 Literature Review of DiabeticRetinopathy Detection

Although extracting vessels before detecting DR with machinelearning can achieve high accuracy it is time-consuming tocreate the marked ground-truth for retinal vessels Anotherparadigm is to train the computer to automatically learn how todistinguish levels of DR by reading retinal images directlywithout performing vessel segmentation In 2000 Ege et al [34]proposed an automatic analysis of DR by different statisticalclassifiers including Bayesian Mahalanobis and k-nearestneighbor Silberman et al [35] introduced an automatic de-tection system for DR and reported an equal error rate of 87Karegowda et al [36] tried to detect exudates in retinal imagesusing back-propagation neural networks (BPN) eir featureswere decided by two methods decision trees and genetic al-gorithmswith correlation-based feature selection (GA-CFS) Intheir experiment the best BPN performance showed 9845accuracy Kavitha and Duraiswamy [37] did some research onautomatic detection of hard and soft exudates in fundus im-ages using color histogram thresholding to classify exudateseir experiments showed 9907 accuracy 89 sensitivityand 99 specificity In 2014 de la Calleja et al [38] used localbinary patterns (LBP) to extract local features and artificialneural networks random forest (RF) and support vectormachines (SVM) for detection In using a dataset containing 71images their best result achieved 9746 accuracy with RF

Mobile Information Systems 5

4 Material and Methods

We propose an automatic DR detection algorithm based onDCNN fractional max-pooling [39] SVM [40] andteaching-learning-based optimization (TLBO) [41] Specif-ically we train two DCNN networks with fractional max-pooling combining their prediction results using SVM andoptimizing the SVM parameters with TLBO e reason fortraining two distinct networks is that different networkarchitectures may have their unique advantages in featurespace representation By training two DCNNs and com-bining their features the prediction accuracy can be furtherenhanced Another important factor impacting the recog-nition rate is the parameter of classifiers We propose tooptimize the SVM parameters using TLBOWe illustrate theimage preprocessing methods in Section 41 and present thefractional max-pooling SVM and TLBO in Sections 4243 and 44 respectively

41 Preprocessing Given the vessels in the original fundusimages are mostly not very clear and the size of eachfundus image may differ it is essential to preprocess imagesso that they have the same size and the visibility of thevessels is improved ere are three steps in preprocessinge first is to rescale images to the same size Since thefundus images are circular we rescale the input images sothat the diameter of the fundus images becomes 540 pixelsAfter rescaling the local average color value is subtractedfrom the rescaled images and another transformation isperformed so that the local average is mapped to 50 gray-scale in order to remove the color divergence caused bydifferent ophthalmoscopes Last but not least becauseboundary effects may occur in some images we remove the

periphery by clipping 10 from the border of the imagesFigure 4 shows the original fundus image and the imageafter preprocessing

42 Fractional Max-Pooling Pooling is a procedure thatturns the input matrix Min times Nin into a smaller Mout times

Nout output matrix e purpose is to divide the inputmatrix into multiple pooling regions (Pij)

pi sub 1 2 3 Min1113864 1113865 for each i isin 1 Mout1113864 1113865

pj sub 1 2 3 Nin1113864 1113865 for each j isin 1 Nout1113864 1113865

Pij pi times pj

(1)

e pooling results are computed according to poolingtype

Outputij Oper

(k l) isin Pij

Inputkl (2)

In equation (2) ldquoOperrdquo refers to a particular mathe-matical operation For example if max-pooling is used theoperation will be to take the maximum of the input regionFor average pooling the average of the input region is takenFor such a network that requires tremendous learning it ispreferable to use as many hidden layers as possible In thiswork the pooling layer used in our networks is fractionalmax-pooling instead of general max-pooling

Fractional max-pooling is a pooling scheme thatmakes the size of the output matrix equivalent to fractionaltimes that of the input matrix after pooling ieNin αNout αge 0 α notin Z To describe the general

Table 2 Performance comparison

Method Year Reference Sn Sp Acc AUC Database

Vascular tracking 2010 [9] 07468 09551 09285 mdash DRIVE2012 [10] 06887 09562 09290 mdash STARE

Matched filtering2009 [14] 06611 09848 09497 mdash STARE

2012 [15] 07191 09687 09407 mdash STARE07154 09716 09343 mdash DRIVE

Morphological processing 2011 [18] 07352 09795 09458 mdash DRIVE2014 [19] 07862 09815 09598 mdash DRIVE

Deformation model

2007 [21] 06634 09682 09316 mdash DRIVE

2009 [23] mdash 09736 09087 mdash STAREmdash 09772 09610 mdash DRIVE

2014 [24] 07187 09767 09509 mdash STARE07354 09789 09477 mdash DRIVE

Machine learning

2007 [25] mdash mdash 09584 09602 STAREmdash mdash 09563 09558 DRIVE

2011 [26] 06944 09819 09526 mdash STARE07067 09801 09452 mdash DRIVE

2013 [27] 08194 09679 09725 mdash DRIVE

2014 [28] 08104 09791 09813 09751 STARE08173 09733 09767 09475 DRIVE

2014 [29] 07887 09633 09441 mdash STARE07512 09684 09412 mdash DRIVE

Sn sensitivity Sp specificity Acc accuracy AUC area under curve

6 Mobile Information Systems

pooling regions let (xi)Mouti0 and (yj)

Noutj0 be two increasing

integer sequences starting with one and ending with Min orNin ese two sequences are used in pooling steps asdescribed in Figures 5 and 6

Pij ximinus1 xi minus a1113858 1113859 times yjminus1 yj minus b1113960 1113961 (3)

e constants a and b in equation (3) stand for theoverlapping length and the width of the pooling windowrespectively Figure 5 is a simple example of overlappingpooling Figure 6 illustrates different pooling region types

After fractional max-pooling the pooling window size isstill integers but the global pooling size will change Namelyfractional max-pooling does not directly change the poolingwindow into a fractional scale Instead it uses windows ofvariable size to achieve fractional pooling e generation of x

and y sequences can be random or pseudorandom Pseudo-random sequences generate more stable pooling regions thanrandom sequences and can also achieve higher accuracy [39]

43 Support Vector Machine (SVM) SVM is a supervisedlearning method used for classification and regressionanalysis SVM can find the hyperplane or decision boundarydefined by the solution vector w which not only separatesthe training vectors but also works well with unseen testdata To improve its generalization ability SVM selectsdecision boundaries based on maximizing margins betweenclasses

Figure 7 illustrates the idea Suppose there are n points ina binary dataset

x1rarr

y1( 1113857 xnrarr

yn( 1113857 (4)

where yi is the data label which can be 1 or minus1 indicatingthe class to which xi

rarr belongs We need to find the optimizedhyperplane such that the distance between the hyperplaneand its nearest point xi

rarr is maximized A hyperplane can bewritten as equation (5) based on x

rarr

wrarr

middot xrarrminus b 0 (5)

where wrarr is the normal vector of the hyperplane and the

value of bwrarr

decides the margin of hyperplane from thetraining data point along the normal vector w

rarr

(a) (b)

(c) (d)

Figure 4 Image before and after the preprocessing stage (a b) images before preprocessing stage (c d) images after preprocessing stage

x = 1 2 3 4 5 y = 1 2 4 5 a = 0 b = 1P11 = [x1minus1 x1 minus 0] times [y1minus1 y1 minus 1] = [1 2] times [1 2]P12 = [x1minus1 x1 minus 1] times [y2minus1 y2 minus 1] = [1 2] times [2 3]

Figure 5 An overlapping example e blue solid line indicatespooling region P11 while the red dotted line shows the P12

Mobile Information Systems 7

For yi whose value is 1 the data must satisfywrarr

middot xrarrminus bge 1 and for yi whose value is minus1 w

rarrmiddot xrarrminus bleminus1

has to be satisfied Combining these two conditions we get

yi(wrarr

middot xrarrminus b)ge 1 (6)

e goal is to maximize bwrarr

according to the constrainof equation (6) in order to derive the optimized decisionhyperplane for classification

Sometimes the training data might not be able to beperfectly separated using linear boundarieserefore in theSVM formulation we need to introduce the error metric εand the cost parameter C as shown in equation (7) e goalnow becomes to minimize

12

wrarr

middot wrarr

+ C 1113944N

i1εi (7)

Subject to

wrarr

middot xirarr

+ bge 1minus εi if wrarr

middot xirarrgt b

wrarr

middot xirarr

+ bleminus1 + εi if wrarr

middot xirarrlt b

εi gt 0

(8)

e performance of SVM is influenced by two mainparameters the first one is C which is a tunable parameter inequation (7) e other one is c which is used in the radialbasis function (RBF) kernel to map data into a higher di-mensional space before training and classification e RBFkernel can be defined as

K xi xj1113872 1113873 eminusc ximinusxj

111386811138681113868111386811138681113868111386811138682

(9)

where c denotes the width of the Gaussian envelope in ahigh-dimensional feature space

44 Teaching-Learning-Based Optimization (TLBO)TLBO an evolution-based optimization algorithm wasproposed by Rao et al [41] in 2011 e concept of TLBO isinspired by the evolution of the learning process when agroup or a class of learners learn a target task ere are two

ways of learning in groups or classes (1) learning from theguidance of the teacher and (2) learning from other learnerse procedure of TLBO can be divided into two phases asdescribed below in Sections 441 and 442

441 Teacher Phase In the whole population the teacher(Xteacher) can be considered as the best solution Namelylearners learn from the teacher in the teacher phase In thisphase the teacher strives for enhancing the results of otherindividuals (Xi) by increasing the mean result of theclassroom is can be described as adjusting Xmean toapproximate Xteacher In order to maintain a stochasticnature during the optimization process two randomlygenerated parameters r and TF are applied in each iterationfor the solution Xi as

Xnew Xi + ri Xteacher minusTFXmean( 1113857 (10)

TF round[1 + rand(0 1)] (11)

In equation (10) ri is a randomly selected number in therange of 0 and 1 Moreover Xnew and Xi are the new andexisting solutions at iteration i respectively TF in equation(11) is a teaching factor which can be either 1 or 2

a = 1 b = 1

x = 1 2 hellip y = 1 2 hellip

x = 1 3 hellip y = 1 2 hellip

x = 1 2 hellip y = 1 3 hellip

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 1] times [1 1]

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 2] times [1 1]

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 1] times [1 2]

Figure 6 Different pooling types when both a and b equal to onee left side shows the equations for pooling condition and computationof pooling region Pictures on the right side indicate regions before and after pooling

w middot x ndash

b = 1

w middot x ndash

b = 0

w middot x ndash

b = ndash1

Figure 7 Example of decision boundary hyperplane with twoclasses of samples

8 Mobile Information Systems

442 Learner Phase e learners gain their knowledge byinteracting with each other erefore an individual learnsnew information if other individuals have more knowledgethan him or her In this phase the student Xi interactsrandomly with another student Xj (ine j) in order to en-hance his or her knowledge Equation (12) shows that if Xi isbetter than Xj (ie f(Xj)gtf(Xi) for minimizationproblems) Xj is moved toward Xi Otherwise it is movedaway from Xi

Xnew Xi + ri Xi minusXj1113872 1113873 f Xj1113872 1113873gtf Xi( 1113857

Xi + ri Xj minusXi1113872 1113873 f Xi( 1113857gtf Xj1113872 1113873

⎧⎪⎨

⎪⎩(12)

If the new solution Xnew to the problem is better than theold ones the new solution Xnew will be recorded as the bestsolution After updating the status of each learner a newiteration begins A stop criterion based on the iterationnumber or the difference of the cost function can be set tostop the iteration properly e flowchart of TLBO is shownin Figure 8

5 Result

Our fundus image data is from the database provided by oneof the Kaggle contests entitled ldquoIdentify signs of diabeticretinopathy in eye imagesrdquo [42] In this database there areabout 90000 images We separate 1000 images from thetraining dataset to be the validation dataset e detailedinformation of each dataset is shown in Table 3 and our twonetwork architectures are shown in Figure 9

Our proposed method uses two DCNNs with fractionalmax-pooling layers For every input fundus image the twoDCNN will output a vector of size 1 times 5 representing theprobability distribution of the prediction for each lesion(category) e probability distribution together with othervalues forms a feature with dimensionality 24 e 24features are described as follows

(i) DCNN probabilities of each lesion respectively (5features)

(ii) Averages of R G and B channel values within 50 lowast50 center cropped image (3 features)

(iii) Widths and heights of 50 lowast 50 center croppedimage (2 features)

(iv) Overall standard deviation of the original image and50 lowast 50 centered cropped image Laplacian-filtered image (2 features)

(v) In total there are 12 features for one fundusimage We then append another 12 features fromthe fundus image of the other eye of the samesubject erefore the overall length of the featurevector is 24 for one fundus image e 24 featurevectors of dimensionality are used as input vec-tors of SVM

e 24-dimensional vector is used to train a multiclassSVM (five classes) whose parameters are optimized usingthe TLBO method We implemented the method described

in [39] and used it as the baseline e baseline system usessimilar features with a scheme of ensemble classifier (RF)

We used the validation set data to optimize the pa-rameter set (C c) in SVM using TLBO e upper andlower bounds of the parameter are set within [0 100] Weran 50 iterations with 50 students

Our final accuracy for five-class classification task ofDR is 8617 and the accuracy for the binary classclassification task is 9105 Labels for five-class classi-fication are normal NPDR level 1 NPDR level 2 NPDRlevel 3 and PDR while labels for binary class classifi-cation are normal and abnormal For binary classifica-tion its sensitivity is 08930 while the specificity is09089 Except counting accuracy we also do a T-test forour binary class classification e T-test is also called theStudentrsquos t-test It is a statistical hypothesis test in whichthe test statistic follows a Studentrsquos t-distributionUsually the t-test is used to compare whether there is asignificant difference between two groups of data andassists in judging the data divergence In doing a pairedsamples t-test with results from binary class classificationand random judgment its outcome is 1 for the hypothesistest result zero for the p value and [03934 04033] forthe confidence interval under null hypothesis at the 5significant level

e hypothesis test result is an index that tells whethertwo data come from the same distribution or not If the datacome from the same distribution the value of the hypothesistest result will be close to 0 On the contrary if the dataresources are distinct the result will be close to 1 whichmeans there is a differentiation between the datae p valueis the probability of accepting the assumption that there is adifference between two data may be wronge smaller the p

value the more reason that there is a disparity between dataAlso we designed an app called ldquoDeep Retinardquo pro-

viding personal examination remote medical care and earlyscreening Figure 10 shows our app interfaces Afterchoosing a fundus image that the user wants to check it willsend the image to our server and use our designed machinelearning method It takes about 10 s (depends on networkspeed) to get the result which will be presented as theprobability of each lesion With a handheld device in-dividuals can do the initial examination at the district officeor even at home More importantly it can benefit someremote areas that lack medical resources

6 Discussion

61 Accuracy Improvement Table 4 shows the accuracycomparison when using different classifiers and pa-rameter optimization methods on each dataset Using thedefault parameters with SVM (without optimization)accuracies in both validation and test sets are higher thanthat of the RF [39] If we optimize the parameters usingthe default parameter searching method provided in theLIBSVM software package though it achieves very highaccuracy in the five-fold cross validation experiment the

Mobile Information Systems 9

validation and test accuracies are even lower than thedefault one From this result we believe that overttingarises when optimizing parameters in SVM

Table 5 shows the confusion matrix of the classicationresults from the two DCNN networks (before performingSVM classication) Network 1 is the architecture shown on

Yes

No

Reject

Yes

No

No

Yes

Yes

Start

Initial students terminal criterion

i = 1

Calculate mean of each design variables

Identify best teachersolution

Based on best teachersolution to modifyXnewi = Xoldi + differencei = Xoldi +

ri (Xteacher ndash (TF)mean)

Randomly select any solution Xj (Xj ne Xi)

i = i + 1

Final value of solution and end

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is Xi better than Xjf (Xi) gt f (Xj)

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is termination criteria satisfiedf (Xnewi) gt f (Xoldi)

No

Yes

No

Xnewi = Xoldi + ri (Xi ndash Xj)

Xi = Xnewi

Xnewi = Xoldi + ri (Xj ndash Xi)

i le N

Xi = Xnewi

Figure 8 TLBO owchart

10 Mobile Information Systems

the left side of Figure 9 and network 2 is the one on the rightside From Table 5 it shows that the lesion classications of 0and 2 are better than the other categories For lesion 1 mostof the prediction results are incorrect Also for lesions 3 and4 the majority of the results are misclassications that areclassied into lesion 2

Table 6 shows the confusion matrix of the classicationresults using the full procedure of the proposed method(using SVM with TLBO) Table 7 displays the dierencebetween Tables 5 and 6 which serves as a performancecomparison between the two methods (using DCNN onlyand DCNN + SVM + TLBO) From the table every classexcept class 0 and overall accuracy is increased in network 1For network 2 each accuracy except class 3 is increasede decline in accuracy of class 3 is mainly caused by

misclassication of class 2 Table 8 shows the confusionmatrix of the classication results using the baseline methodas reported in [39] for comparison purposes

62 Deep Learning vs Traditional Classication MethodsMany traditional classication methods try to solve theproblem of DR detection by (1) using image processing tocapture symptoms in fundus images and then (2) building aclassier to make decisions based on the detected symptoms(1) e shortcoming of image processing methods is that themanifestations of the symptoms are random across dierentimages therefore it is extremely time-consuming and re-quires intense eorts to label the locations of the symptomsAbiding by the new philosophy that comes with the emer-gence of the deep learning technology our proposed methodis trying to learn how to make decisions directly from theimage data itself Dierent than the former approaches ourimages only need to be labeled with lesion number instead oflabeling symptom locations Consequently it saves consid-erable time during the database preprocessing stage On top ofthe classication results by the two DCNN networks we useSVM optimized by TLBO to generate an improved outcomeand we achieve 8617 accuracy Our result is better than therst-place winner in the Kaggle competition It shows that ourresearch result is the state-of-the-art

63 Limitation In our current datasets the number ofimages of lesions 3 and 4 is not suiexclcient to train a networkwhich is a limitation of the proposed methoderefore oneof our future works is to develop deeper collaborative re-lations with hospitals and clinics to acquire more data oflesions 3 and 4 With more data we believe the classicationaccuracy will be further increased In addition from ourresult we found that it is hard to dierentiate the imagesbetween lesions 0 and 1 erefore when we collect newdata it is desirable to collect more images belonging tolesions 0 and 1 Also we can attempt to use a dierentnetwork architecture for this problem

7 Conclusion

It is feasible to train a deep learning model for automaticdiagnosis of DR as long as we have enough data for sta-tistical model training Furthermore the database prepa-ration stage only needs a categorical label for each trainingimage It does not require detailed annotation for retinalvessel tracking in every image Hence it is time-eiexclcientcompared to the traditional machine learning-based methodfor automatic diagnosis of DR e nal accuracy canachieve 8617 and 9105 for ve-class and binary classclassications respectively

e sensitivity and specicity of binary classication are08930 and 09089 respectively which is a satisfactory resultFurthermore we developed an automatic inspection appthat can be used in both personal examination and remotemedical care With more image data collected we expect theaccuracy can be even more enhanced further improving oursystem

Table 3 Detailed information of each dataset

Dataset of imagesDR lesions

0 1 2 3 4Train 34124 73 6 15 3 3Validation 1000 70 8 15 4 3Test 53572 74 7 15 2 2

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

3 times 3 conv 320

3 times 3 conv 352

3 times 3 conv 384

2 times 2 conv 416

1 times 1 conv 448

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

Softmax

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

2 times 2 conv 320

1 times 1 conv 356

Softmax

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

Figure 9 Architecture of the two DCNN networks that we used

Mobile Information Systems 11

Table 4 Accuracy comparison when using different classifiers and different parameter optimization methods on each dataset

DatasetMethod

SVM RF [39]Without optimization Default parameter searching TLBO

Five-fold cross validation (using training data) 88744 965724 mdash mdashValidation data 854 796 mdash 847Test data 861177 810573 8617 8602

Table 5 Confusion matrix of the classification results from the two DCNN networks before performing SVM classification

Determined levelGround-truth level

Network 1 Network 20 1 2 3 4 0 1 2 3 4

0 39031 2693 2361 75 160 38310 2233 1617 51 781 118 438 185 0 1 327 735 290 1 32 339 626 5058 738 392 790 786 5509 620 3453 0 0 170 332 102 4 1 338 486 1854 43 5 85 69 551 100 7 105 56 595Accuracy () 9874 1164 6436 2735 4569 9691 1954 7001 4003 4934Overall accuracy () 8476 8518

(a) (b) (c) (d)

Figure 10 App interfaces

Table 6 Confusion matrix of the classification results using the full procedure of the proposed method (using SVM with TLBO)

Determined levelGround-truth level

SVM optimized with TLBO0 1 2 3 4

0 38611 2233 1513 44 801 312 831 349 1 12 570 696 5706 701 3933 3 0 215 408 1234 35 2 76 60 609Accuracy () 9767 2209 7260 3361 5050Overall accuracy () 8617

12 Mobile Information Systems

Data Availability

All experiment data come from a Kaggle contest calledldquoIdentify signs of diabetic retinopathy in eye imagesrdquo(website httpswwwkagglecomcdiabetic-retinopathy-detection)

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e authors would like to thank Mr Chia-Ming Hu for hishelp in preparing computational instruments and per-forming parts of the experiments in this paper

References

[1] N H Cho J E Shaw S Karuranga et al ldquoIDF Diabetes Atlasglobal estimates of diabetes prevalence for 2017 and pro-jections for 2045rdquo Diabetes Research and Clinical Practicevol 138 pp 271ndash281 2018

[2] American Diabetes Association Eye Complications AmericanDiabetes Association Arlington VA USA 2013 httpwwwdiabetesorgliving-with-diabetescomplicationseye-complicationsreferrerhttpswwwgooglecomtw

[3] International Diabetes Foundation (IDF) Diabetes AtlasInternational Diabetes Federation Brussels Belgium 7thedition 2015

[4] C P Wilkinson F L Ferris R E Klein et al ldquoProposedinternational clinical diabetic retinopathy and diabetic

macular edema disease severity scalesrdquo Ophthalmologyvol 110 no 9 pp 1677ndash1682 2003

[5] A Tufail C Rudisill C Egan et al ldquoAutomated diabeticretinopathy image assessment software diagnostic accuracyand cost-effectiveness compared with human gradersrdquoOphthalmology vol 124 no 3 pp 343ndash351 2017

[6] A Tufail V V Kapetanakis S Salas-Vega et al ldquoAn ob-servational study to assess if automated diabetic retinopathyimage assessment software can replace one or more steps ofmanual imaging grading and to determine their cost-effec-tivenessrdquo Health Technology Assessment vol 20 no 92pp 1ndash72 2016

[7] I Liu and Y Sun ldquoRecursive tracking of vascular networks inangiograms based on the detection-deletion schemerdquo IEEETransactions on Medical Imaging vol 12 no 2 pp 334ndash3411993

[8] A Can H Shen J N Turner H L Tanenbaum andB Roysam ldquoRapid automated tracing and feature extractionfrom retinal fundus images using direct exploratory algo-rithmsrdquo IEEE Transactions on Information Technology inBiomedicine vol 3 no 2 pp 125ndash138 1999

[9] M Vlachos and E Dermatas ldquoMulti-scale retinal vesselsegmentation using line trackingrdquo Computerized MedicalImaging and Graphics vol 34 no 3 pp 213ndash227 2010

[10] Y Yin M Adel and S Bourennane ldquoAn automatic trackingmethod for retinal vascular tree extractionrdquo in Proceedings ofthe IEEE International Conference on Acoustics Speech andSignal Processing IEEE Kyoto Japan March 2012

[11] S Chaudhuri S Chatterjee N Katz M Nelson andM Goldbaum ldquoDetection of blood vessels in retinal imagesusing two-dimensional matched filtersrdquo IEEE Transactions onMedical Imaging vol 8 no 3 pp 263ndash269 1989

Table 8 Confusion matrix of the classification using the baseline method [39] and its overall accuracy

DeterminedOriginal

Random forest0 1 2 3 4

0 38228 2001 1331 38 641 546 1026 434 3 22 710 732 5726 635 3583 2 1 279 465 1464 45 2 89 73 636Accuracy () 9670 2727 7286 3830 5274Overall accuracy () 8601

Table 7 Differences between Tables 5 and 6

Determined lesionGround-truth lesion

SVM-network 1 SVM-network 20 1 2 3 4 0 1 2 3 4

0 +420 minus460 minus848 minus31 minus80 +310 0 minus104 minus43 +21 +194 +393 +164 +1 0 minus15 +96 +59 0 minus22 +231 +70 +648 minus37 +1 minus220 minus90 +197 +81 +483 +3 0 +45 +76 +21 +1 minus1 minus123 minus78 minus624 +3 minus3 +45 +76 +58 minus65 minus5 minus28 +4 +14Accuracy increment () minus108 8977 1280 2288 1052 078 1305 369 minus1603 23Overall accuracy increment () 16 116

Mobile Information Systems 13

accuracy of vessel segmentation Currently setting the initialpoint can be done either artificially or automatically

e earliest adaptive vascular tracking method wasproposed by Liu and Sun [7] in 1993 which extracts thevasculature from X-ray angiograms First given an initialpoint and direction within a vessel the authors apply anldquoextrapolation-updaterdquo scheme that involves estimatinglocal vessel trajectories Once a vessel fragment has beentracked it is removed from the image is procedure isrepeated until the vascular tree has been extracted edrawbacks of this strategy are that due to the algorithm usedthe user must set the vessel starting points and that theapproach does not seem adaptable to three-dimensionalextraction In 1999 an automatic vascular trackingmethod was developed by Can et al [8] is strategy mainlycollects pixel wide vascular local minimum points (usually inthe middle of a vessel) to perform tracking Vlachos andDermatas [9] suggested a multiscale line tracking methodwith morphological postprocessing Yin et al [10] proposeda retinal vascular tree extraction based on iterative trackingand Bayesian method

e advantage of vascular tracking is that it can providelocal information about characteristics such as the

diameterwidth and direction of vessels However thevascular tracking performance can be easily affected bycrossing or branching of vessels which reduces the iden-tification efficiency

22 Methods of Matched Filtering Matched filteringmethods employ multiple matched filters for extraction sodesigning proper filters is essential to detect vessels Since thegray-scale distribution of fundus vessels is in keeping withGaussian an intuitive method exists that uses the maximumresponse of images after filtering to find vessel points As thediameterwidth of vessels is diverse a multiscale Gaussianfilter method is often used for vessel tracking

In 1989 Chaudhuri et al [11] pioneered the applicationof Gaussian filters in vessel tracking by using some vascularcharacteristics such as the fact that vessels are darker thanthe background the width of the vessels ranges from 2 to 10pixels and the vessels grow from the optic disc into a radialshape erefore Chaudhuri et al [11] designed two-dimensional Gaussian filters that can detect vessels in 12different directions However this method needs largecomputation and some of the dark lesions are similar to the

Table 1 Classification of diabetic retinopathy [4]

Category Level Manifestation

Nonproliferative diabetic retinopathy (NPDR)

1 Few microaneurysms or hemorrhage

2 Microaneurysms hemorrhages soft exudates venousbeading

3Severe retinal hemorrhage in ge4 quadrants or venousbeading change in ge2 quadrants or mild intraretinal

microvascular abnormality in ge1 quadrantProliferative diabetic retinopathy (PDR) 4 Neovascularization of retina or optic disc

(a) (b) (c)

(d) (e)

Figure 1 Examples of fundus images showing different lesions (a) 0 level (normal) (b) level 1 (c) level 2 (d) level 3 and (e) level 4

Mobile Information Systems 3

characteristic of vessels causing tracking errors Hooveret al [12] described an improvedmethod that considers localand regional characteristics of vessels to separate bloodvessels in retinal images and iteratively determine whetherthe current point is a vessel point

After such improvement a large number of studies ofreformed filters have been developed Jiang and Mojon [13]promoted a generalized threshold method based on amultithreshold detection Zhang et al [14] improved thematching filtering method by applying a local vessel crosssection analysis using local bilateral thresholding Li et al[15] suggested a multiscale production of the matched filterto enhance the extraction of tiny vessels

23 Methods of Morphological Processing Morphologicalprocessing facilitates the segmentation and identification oftarget objects by analyzing and processing structural ele-ments in a binary image us linear and circular elementsof blood vessels can be selected isolating the desiredstructure instead of the background image In addition

morphological processing can also smooth and fill the imagecontour with the advantage of antinoise However thismethod overrelies on structural elements and does not makegood use of characteristics of vessels

According to vessel characteristics Zana and Klein [16]introduced a mathematical morphology-based algorithmthat allows separating the vessels from all possible un-desirable patterns Building on this approach Ayala et al[17] proposed using different average fuzzy sets In Miri andMahloojifar [18] fundus images were analyzed by the use ofcurvelet transform and morphological reconstruction ofmultistructural elements to enhance the boundaries anddetermine the vascular ridge Karthika and Marimuthu [19]combined curvelet transform and morphological re-construction of multistructural elements with stronglyconnected component analysis (SCCA) to segment andidentify vessels

24 Methods of Deformation Models First introduced byKass et al [20] in 1988 the key benefit of deformation

Level 0Level 1Level 2Level 3Level 4

Level 4

Big datacollection

Onlinetraining

Deep learningalgorithm on

cloud computingplatform

Image acquisitionon mobile device

Figure 3 e proposed framework It is composed of mobile block (image acquisition) cloud block (deep learning algorithm on cloudcomputing platform) and big data block (big data collection and training)

Figure 2 Commercialized fundus camera used in the clinics e machine is bulky and expensive making it difficult for large-scalescreening

4 Mobile Information Systems

models is the ability to produce smooth parametric curves orsurfaces Two categories of deformation models are iden-tified parametric deformation and geometric deformationParametric deformation models are also called active con-tour or snake models (set of points each with an associatedenergy) rough the external and internal forces acting onthe snake the snake model can change its shape andsmoothness toward the desired structure In 2007 Esponaet al [21] used a parametric deformation model method onfundus images and promoted an improved method withmorphological segmentation With the assistance of mor-phological vessel segmentation the snake model expands tothe contour of the obtained vessels until the local energyfunction is minimal Another deformation model methodcalled ribbon of twins (ROT) which combines ribbon snakesand double snakes was proposed by Al-Diri and Hunter[22] Each twin consists of two snakes one inside and oneoutside the vessel edges e double snake model then at-tempts to integrate the pairs of twins on the vessel bordersinto a single ribbon and calculate the vessel width

ere are several shortcomings in parametric de-formation models For instance the segmentation resultsdepend on the initial contour and difficulties arise whenextending from low to high dimensions and in segmentingcomplex objects Geometric deformation can well solve theproblems caused by parametric deformable models Geo-metric deformable models are based on deformation curveevolution theory and have no strict requirement on theposition of the initial contour which increases the robust-ness of the method and allows it to be extended to highdimensions Zhang et al [23] proposed an automatic vesselsegmentation method which uses nonlinear orthogonalprojection to capture the characteristics of retinal bloodvessels and obtained an adaptive local thresholding algo-rithm for blood vessel segmentation Zhao et al [24] sug-gested a retinal vessel segmentation method that employs aregion-based active contour model with a level set imple-mentation and a region growing model

25 Methods of Deformation Model Machine learning is analgorithm that teaches computers to learn to achieve goalsautomatically by building generative or discriminativemodels from accumulated datasets Machine learning can bedivided into supervised learning and unsupervised learninge supervised learning methods learn to achieve goalsbased on ground-truth which means that during thetraining stage the training data used to train the model comewith a ldquolabelrdquo that can be used by the machine learningalgorithm to differentiate the data Applying such paradigmin the problem of DR it means that when using supervisedlearning one needs to mark all of the pixels belonging tovessels in advance whereas the unsupervised learningmethod does not need to mark them beforehand

For supervised learning Cesar and Jelinek [30] andLeandro et al [31] proposed a supervised classification withtwo-dimensional Gabor wavelet Each pixel has a featurevector that consists of the gray-scale feature and responses ofdistinct sizes of two-dimensional Gabor wavelet Ricci and

Perfetti [25] proposed a segmentation method for retinalvessels based online manipulation and support vectorclassification Since the features are extracted by twoorthogonal vertical lines it reduces the features andtraining samples in supervised learning A supervisedmethod using neural network was proposed by Marinet al [26] which has one input layer three hidden layersand one output layer Each pixel in the image is repre-sented by a seven-dimensional feature vector to train thenetwork Shanmugam and Banu [27] used an extremelearning machine (ELM) to detect retinal vessels bycreating a seven-dimensional feature vector based ongray-scale features and invariant moments and using ELMto segment vessels In 2015 Wang et al [28] raised a newhierarchical retinal vascular segmentation includingthree steps preprocessing hierarchical feature extractionand integration classification It involves using simplelinear iterative clustering (SLIC) to perform super-pixelsegmentation and randomly selecting a pixel to representthe entire super-pixel as a more easy and efficient meansof extracting features

For unsupervised learning in 1998 Tolias and Panas[32] created an automatic and unsupervised segmentationmethod based on blurred fundus images which used fuzzyC-means (FCM) to find initial candidate points Xie and Nie[33] proposed a segmentation method based on a geneticalgorithm and FCM Salazar-Gonzalez et al [29] usedmethods of vector flow to segment retinal vessels

Table 2 is a summary about the performance comparisonbetween different existing methods

3 Literature Review of DiabeticRetinopathy Detection

Although extracting vessels before detecting DR with machinelearning can achieve high accuracy it is time-consuming tocreate the marked ground-truth for retinal vessels Anotherparadigm is to train the computer to automatically learn how todistinguish levels of DR by reading retinal images directlywithout performing vessel segmentation In 2000 Ege et al [34]proposed an automatic analysis of DR by different statisticalclassifiers including Bayesian Mahalanobis and k-nearestneighbor Silberman et al [35] introduced an automatic de-tection system for DR and reported an equal error rate of 87Karegowda et al [36] tried to detect exudates in retinal imagesusing back-propagation neural networks (BPN) eir featureswere decided by two methods decision trees and genetic al-gorithmswith correlation-based feature selection (GA-CFS) Intheir experiment the best BPN performance showed 9845accuracy Kavitha and Duraiswamy [37] did some research onautomatic detection of hard and soft exudates in fundus im-ages using color histogram thresholding to classify exudateseir experiments showed 9907 accuracy 89 sensitivityand 99 specificity In 2014 de la Calleja et al [38] used localbinary patterns (LBP) to extract local features and artificialneural networks random forest (RF) and support vectormachines (SVM) for detection In using a dataset containing 71images their best result achieved 9746 accuracy with RF

Mobile Information Systems 5

4 Material and Methods

We propose an automatic DR detection algorithm based onDCNN fractional max-pooling [39] SVM [40] andteaching-learning-based optimization (TLBO) [41] Specif-ically we train two DCNN networks with fractional max-pooling combining their prediction results using SVM andoptimizing the SVM parameters with TLBO e reason fortraining two distinct networks is that different networkarchitectures may have their unique advantages in featurespace representation By training two DCNNs and com-bining their features the prediction accuracy can be furtherenhanced Another important factor impacting the recog-nition rate is the parameter of classifiers We propose tooptimize the SVM parameters using TLBOWe illustrate theimage preprocessing methods in Section 41 and present thefractional max-pooling SVM and TLBO in Sections 4243 and 44 respectively

41 Preprocessing Given the vessels in the original fundusimages are mostly not very clear and the size of eachfundus image may differ it is essential to preprocess imagesso that they have the same size and the visibility of thevessels is improved ere are three steps in preprocessinge first is to rescale images to the same size Since thefundus images are circular we rescale the input images sothat the diameter of the fundus images becomes 540 pixelsAfter rescaling the local average color value is subtractedfrom the rescaled images and another transformation isperformed so that the local average is mapped to 50 gray-scale in order to remove the color divergence caused bydifferent ophthalmoscopes Last but not least becauseboundary effects may occur in some images we remove the

periphery by clipping 10 from the border of the imagesFigure 4 shows the original fundus image and the imageafter preprocessing

42 Fractional Max-Pooling Pooling is a procedure thatturns the input matrix Min times Nin into a smaller Mout times

Nout output matrix e purpose is to divide the inputmatrix into multiple pooling regions (Pij)

pi sub 1 2 3 Min1113864 1113865 for each i isin 1 Mout1113864 1113865

pj sub 1 2 3 Nin1113864 1113865 for each j isin 1 Nout1113864 1113865

Pij pi times pj

(1)

e pooling results are computed according to poolingtype

Outputij Oper

(k l) isin Pij

Inputkl (2)

In equation (2) ldquoOperrdquo refers to a particular mathe-matical operation For example if max-pooling is used theoperation will be to take the maximum of the input regionFor average pooling the average of the input region is takenFor such a network that requires tremendous learning it ispreferable to use as many hidden layers as possible In thiswork the pooling layer used in our networks is fractionalmax-pooling instead of general max-pooling

Fractional max-pooling is a pooling scheme thatmakes the size of the output matrix equivalent to fractionaltimes that of the input matrix after pooling ieNin αNout αge 0 α notin Z To describe the general

Table 2 Performance comparison

Method Year Reference Sn Sp Acc AUC Database

Vascular tracking 2010 [9] 07468 09551 09285 mdash DRIVE2012 [10] 06887 09562 09290 mdash STARE

Matched filtering2009 [14] 06611 09848 09497 mdash STARE

2012 [15] 07191 09687 09407 mdash STARE07154 09716 09343 mdash DRIVE

Morphological processing 2011 [18] 07352 09795 09458 mdash DRIVE2014 [19] 07862 09815 09598 mdash DRIVE

Deformation model

2007 [21] 06634 09682 09316 mdash DRIVE

2009 [23] mdash 09736 09087 mdash STAREmdash 09772 09610 mdash DRIVE

2014 [24] 07187 09767 09509 mdash STARE07354 09789 09477 mdash DRIVE

Machine learning

2007 [25] mdash mdash 09584 09602 STAREmdash mdash 09563 09558 DRIVE

2011 [26] 06944 09819 09526 mdash STARE07067 09801 09452 mdash DRIVE

2013 [27] 08194 09679 09725 mdash DRIVE

2014 [28] 08104 09791 09813 09751 STARE08173 09733 09767 09475 DRIVE

2014 [29] 07887 09633 09441 mdash STARE07512 09684 09412 mdash DRIVE

Sn sensitivity Sp specificity Acc accuracy AUC area under curve

6 Mobile Information Systems

pooling regions let (xi)Mouti0 and (yj)

Noutj0 be two increasing

integer sequences starting with one and ending with Min orNin ese two sequences are used in pooling steps asdescribed in Figures 5 and 6

Pij ximinus1 xi minus a1113858 1113859 times yjminus1 yj minus b1113960 1113961 (3)

e constants a and b in equation (3) stand for theoverlapping length and the width of the pooling windowrespectively Figure 5 is a simple example of overlappingpooling Figure 6 illustrates different pooling region types

After fractional max-pooling the pooling window size isstill integers but the global pooling size will change Namelyfractional max-pooling does not directly change the poolingwindow into a fractional scale Instead it uses windows ofvariable size to achieve fractional pooling e generation of x

and y sequences can be random or pseudorandom Pseudo-random sequences generate more stable pooling regions thanrandom sequences and can also achieve higher accuracy [39]

43 Support Vector Machine (SVM) SVM is a supervisedlearning method used for classification and regressionanalysis SVM can find the hyperplane or decision boundarydefined by the solution vector w which not only separatesthe training vectors but also works well with unseen testdata To improve its generalization ability SVM selectsdecision boundaries based on maximizing margins betweenclasses

Figure 7 illustrates the idea Suppose there are n points ina binary dataset

x1rarr

y1( 1113857 xnrarr

yn( 1113857 (4)

where yi is the data label which can be 1 or minus1 indicatingthe class to which xi

rarr belongs We need to find the optimizedhyperplane such that the distance between the hyperplaneand its nearest point xi

rarr is maximized A hyperplane can bewritten as equation (5) based on x

rarr

wrarr

middot xrarrminus b 0 (5)

where wrarr is the normal vector of the hyperplane and the

value of bwrarr

decides the margin of hyperplane from thetraining data point along the normal vector w

rarr

(a) (b)

(c) (d)

Figure 4 Image before and after the preprocessing stage (a b) images before preprocessing stage (c d) images after preprocessing stage

x = 1 2 3 4 5 y = 1 2 4 5 a = 0 b = 1P11 = [x1minus1 x1 minus 0] times [y1minus1 y1 minus 1] = [1 2] times [1 2]P12 = [x1minus1 x1 minus 1] times [y2minus1 y2 minus 1] = [1 2] times [2 3]

Figure 5 An overlapping example e blue solid line indicatespooling region P11 while the red dotted line shows the P12

Mobile Information Systems 7

For yi whose value is 1 the data must satisfywrarr

middot xrarrminus bge 1 and for yi whose value is minus1 w

rarrmiddot xrarrminus bleminus1

has to be satisfied Combining these two conditions we get

yi(wrarr

middot xrarrminus b)ge 1 (6)

e goal is to maximize bwrarr

according to the constrainof equation (6) in order to derive the optimized decisionhyperplane for classification

Sometimes the training data might not be able to beperfectly separated using linear boundarieserefore in theSVM formulation we need to introduce the error metric εand the cost parameter C as shown in equation (7) e goalnow becomes to minimize

12

wrarr

middot wrarr

+ C 1113944N

i1εi (7)

Subject to

wrarr

middot xirarr

+ bge 1minus εi if wrarr

middot xirarrgt b

wrarr

middot xirarr

+ bleminus1 + εi if wrarr

middot xirarrlt b

εi gt 0

(8)

e performance of SVM is influenced by two mainparameters the first one is C which is a tunable parameter inequation (7) e other one is c which is used in the radialbasis function (RBF) kernel to map data into a higher di-mensional space before training and classification e RBFkernel can be defined as

K xi xj1113872 1113873 eminusc ximinusxj

111386811138681113868111386811138681113868111386811138682

(9)

where c denotes the width of the Gaussian envelope in ahigh-dimensional feature space

44 Teaching-Learning-Based Optimization (TLBO)TLBO an evolution-based optimization algorithm wasproposed by Rao et al [41] in 2011 e concept of TLBO isinspired by the evolution of the learning process when agroup or a class of learners learn a target task ere are two

ways of learning in groups or classes (1) learning from theguidance of the teacher and (2) learning from other learnerse procedure of TLBO can be divided into two phases asdescribed below in Sections 441 and 442

441 Teacher Phase In the whole population the teacher(Xteacher) can be considered as the best solution Namelylearners learn from the teacher in the teacher phase In thisphase the teacher strives for enhancing the results of otherindividuals (Xi) by increasing the mean result of theclassroom is can be described as adjusting Xmean toapproximate Xteacher In order to maintain a stochasticnature during the optimization process two randomlygenerated parameters r and TF are applied in each iterationfor the solution Xi as

Xnew Xi + ri Xteacher minusTFXmean( 1113857 (10)

TF round[1 + rand(0 1)] (11)

In equation (10) ri is a randomly selected number in therange of 0 and 1 Moreover Xnew and Xi are the new andexisting solutions at iteration i respectively TF in equation(11) is a teaching factor which can be either 1 or 2

a = 1 b = 1

x = 1 2 hellip y = 1 2 hellip

x = 1 3 hellip y = 1 2 hellip

x = 1 2 hellip y = 1 3 hellip

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 1] times [1 1]

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 2] times [1 1]

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 1] times [1 2]

Figure 6 Different pooling types when both a and b equal to onee left side shows the equations for pooling condition and computationof pooling region Pictures on the right side indicate regions before and after pooling

w middot x ndash

b = 1

w middot x ndash

b = 0

w middot x ndash

b = ndash1

Figure 7 Example of decision boundary hyperplane with twoclasses of samples

8 Mobile Information Systems

442 Learner Phase e learners gain their knowledge byinteracting with each other erefore an individual learnsnew information if other individuals have more knowledgethan him or her In this phase the student Xi interactsrandomly with another student Xj (ine j) in order to en-hance his or her knowledge Equation (12) shows that if Xi isbetter than Xj (ie f(Xj)gtf(Xi) for minimizationproblems) Xj is moved toward Xi Otherwise it is movedaway from Xi

Xnew Xi + ri Xi minusXj1113872 1113873 f Xj1113872 1113873gtf Xi( 1113857

Xi + ri Xj minusXi1113872 1113873 f Xi( 1113857gtf Xj1113872 1113873

⎧⎪⎨

⎪⎩(12)

If the new solution Xnew to the problem is better than theold ones the new solution Xnew will be recorded as the bestsolution After updating the status of each learner a newiteration begins A stop criterion based on the iterationnumber or the difference of the cost function can be set tostop the iteration properly e flowchart of TLBO is shownin Figure 8

5 Result

Our fundus image data is from the database provided by oneof the Kaggle contests entitled ldquoIdentify signs of diabeticretinopathy in eye imagesrdquo [42] In this database there areabout 90000 images We separate 1000 images from thetraining dataset to be the validation dataset e detailedinformation of each dataset is shown in Table 3 and our twonetwork architectures are shown in Figure 9

Our proposed method uses two DCNNs with fractionalmax-pooling layers For every input fundus image the twoDCNN will output a vector of size 1 times 5 representing theprobability distribution of the prediction for each lesion(category) e probability distribution together with othervalues forms a feature with dimensionality 24 e 24features are described as follows

(i) DCNN probabilities of each lesion respectively (5features)

(ii) Averages of R G and B channel values within 50 lowast50 center cropped image (3 features)

(iii) Widths and heights of 50 lowast 50 center croppedimage (2 features)

(iv) Overall standard deviation of the original image and50 lowast 50 centered cropped image Laplacian-filtered image (2 features)

(v) In total there are 12 features for one fundusimage We then append another 12 features fromthe fundus image of the other eye of the samesubject erefore the overall length of the featurevector is 24 for one fundus image e 24 featurevectors of dimensionality are used as input vec-tors of SVM

e 24-dimensional vector is used to train a multiclassSVM (five classes) whose parameters are optimized usingthe TLBO method We implemented the method described

in [39] and used it as the baseline e baseline system usessimilar features with a scheme of ensemble classifier (RF)

We used the validation set data to optimize the pa-rameter set (C c) in SVM using TLBO e upper andlower bounds of the parameter are set within [0 100] Weran 50 iterations with 50 students

Our final accuracy for five-class classification task ofDR is 8617 and the accuracy for the binary classclassification task is 9105 Labels for five-class classi-fication are normal NPDR level 1 NPDR level 2 NPDRlevel 3 and PDR while labels for binary class classifi-cation are normal and abnormal For binary classifica-tion its sensitivity is 08930 while the specificity is09089 Except counting accuracy we also do a T-test forour binary class classification e T-test is also called theStudentrsquos t-test It is a statistical hypothesis test in whichthe test statistic follows a Studentrsquos t-distributionUsually the t-test is used to compare whether there is asignificant difference between two groups of data andassists in judging the data divergence In doing a pairedsamples t-test with results from binary class classificationand random judgment its outcome is 1 for the hypothesistest result zero for the p value and [03934 04033] forthe confidence interval under null hypothesis at the 5significant level

e hypothesis test result is an index that tells whethertwo data come from the same distribution or not If the datacome from the same distribution the value of the hypothesistest result will be close to 0 On the contrary if the dataresources are distinct the result will be close to 1 whichmeans there is a differentiation between the datae p valueis the probability of accepting the assumption that there is adifference between two data may be wronge smaller the p

value the more reason that there is a disparity between dataAlso we designed an app called ldquoDeep Retinardquo pro-

viding personal examination remote medical care and earlyscreening Figure 10 shows our app interfaces Afterchoosing a fundus image that the user wants to check it willsend the image to our server and use our designed machinelearning method It takes about 10 s (depends on networkspeed) to get the result which will be presented as theprobability of each lesion With a handheld device in-dividuals can do the initial examination at the district officeor even at home More importantly it can benefit someremote areas that lack medical resources

6 Discussion

61 Accuracy Improvement Table 4 shows the accuracycomparison when using different classifiers and pa-rameter optimization methods on each dataset Using thedefault parameters with SVM (without optimization)accuracies in both validation and test sets are higher thanthat of the RF [39] If we optimize the parameters usingthe default parameter searching method provided in theLIBSVM software package though it achieves very highaccuracy in the five-fold cross validation experiment the

Mobile Information Systems 9

validation and test accuracies are even lower than thedefault one From this result we believe that overttingarises when optimizing parameters in SVM

Table 5 shows the confusion matrix of the classicationresults from the two DCNN networks (before performingSVM classication) Network 1 is the architecture shown on

Yes

No

Reject

Yes

No

No

Yes

Yes

Start

Initial students terminal criterion

i = 1

Calculate mean of each design variables

Identify best teachersolution

Based on best teachersolution to modifyXnewi = Xoldi + differencei = Xoldi +

ri (Xteacher ndash (TF)mean)

Randomly select any solution Xj (Xj ne Xi)

i = i + 1

Final value of solution and end

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is Xi better than Xjf (Xi) gt f (Xj)

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is termination criteria satisfiedf (Xnewi) gt f (Xoldi)

No

Yes

No

Xnewi = Xoldi + ri (Xi ndash Xj)

Xi = Xnewi

Xnewi = Xoldi + ri (Xj ndash Xi)

i le N

Xi = Xnewi

Figure 8 TLBO owchart

10 Mobile Information Systems

the left side of Figure 9 and network 2 is the one on the rightside From Table 5 it shows that the lesion classications of 0and 2 are better than the other categories For lesion 1 mostof the prediction results are incorrect Also for lesions 3 and4 the majority of the results are misclassications that areclassied into lesion 2

Table 6 shows the confusion matrix of the classicationresults using the full procedure of the proposed method(using SVM with TLBO) Table 7 displays the dierencebetween Tables 5 and 6 which serves as a performancecomparison between the two methods (using DCNN onlyand DCNN + SVM + TLBO) From the table every classexcept class 0 and overall accuracy is increased in network 1For network 2 each accuracy except class 3 is increasede decline in accuracy of class 3 is mainly caused by

misclassication of class 2 Table 8 shows the confusionmatrix of the classication results using the baseline methodas reported in [39] for comparison purposes

62 Deep Learning vs Traditional Classication MethodsMany traditional classication methods try to solve theproblem of DR detection by (1) using image processing tocapture symptoms in fundus images and then (2) building aclassier to make decisions based on the detected symptoms(1) e shortcoming of image processing methods is that themanifestations of the symptoms are random across dierentimages therefore it is extremely time-consuming and re-quires intense eorts to label the locations of the symptomsAbiding by the new philosophy that comes with the emer-gence of the deep learning technology our proposed methodis trying to learn how to make decisions directly from theimage data itself Dierent than the former approaches ourimages only need to be labeled with lesion number instead oflabeling symptom locations Consequently it saves consid-erable time during the database preprocessing stage On top ofthe classication results by the two DCNN networks we useSVM optimized by TLBO to generate an improved outcomeand we achieve 8617 accuracy Our result is better than therst-place winner in the Kaggle competition It shows that ourresearch result is the state-of-the-art

63 Limitation In our current datasets the number ofimages of lesions 3 and 4 is not suiexclcient to train a networkwhich is a limitation of the proposed methoderefore oneof our future works is to develop deeper collaborative re-lations with hospitals and clinics to acquire more data oflesions 3 and 4 With more data we believe the classicationaccuracy will be further increased In addition from ourresult we found that it is hard to dierentiate the imagesbetween lesions 0 and 1 erefore when we collect newdata it is desirable to collect more images belonging tolesions 0 and 1 Also we can attempt to use a dierentnetwork architecture for this problem

7 Conclusion

It is feasible to train a deep learning model for automaticdiagnosis of DR as long as we have enough data for sta-tistical model training Furthermore the database prepa-ration stage only needs a categorical label for each trainingimage It does not require detailed annotation for retinalvessel tracking in every image Hence it is time-eiexclcientcompared to the traditional machine learning-based methodfor automatic diagnosis of DR e nal accuracy canachieve 8617 and 9105 for ve-class and binary classclassications respectively

e sensitivity and specicity of binary classication are08930 and 09089 respectively which is a satisfactory resultFurthermore we developed an automatic inspection appthat can be used in both personal examination and remotemedical care With more image data collected we expect theaccuracy can be even more enhanced further improving oursystem

Table 3 Detailed information of each dataset

Dataset of imagesDR lesions

0 1 2 3 4Train 34124 73 6 15 3 3Validation 1000 70 8 15 4 3Test 53572 74 7 15 2 2

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

3 times 3 conv 320

3 times 3 conv 352

3 times 3 conv 384

2 times 2 conv 416

1 times 1 conv 448

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

Softmax

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

2 times 2 conv 320

1 times 1 conv 356

Softmax

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

Figure 9 Architecture of the two DCNN networks that we used

Mobile Information Systems 11

Table 4 Accuracy comparison when using different classifiers and different parameter optimization methods on each dataset

DatasetMethod

SVM RF [39]Without optimization Default parameter searching TLBO

Five-fold cross validation (using training data) 88744 965724 mdash mdashValidation data 854 796 mdash 847Test data 861177 810573 8617 8602

Table 5 Confusion matrix of the classification results from the two DCNN networks before performing SVM classification

Determined levelGround-truth level

Network 1 Network 20 1 2 3 4 0 1 2 3 4

0 39031 2693 2361 75 160 38310 2233 1617 51 781 118 438 185 0 1 327 735 290 1 32 339 626 5058 738 392 790 786 5509 620 3453 0 0 170 332 102 4 1 338 486 1854 43 5 85 69 551 100 7 105 56 595Accuracy () 9874 1164 6436 2735 4569 9691 1954 7001 4003 4934Overall accuracy () 8476 8518

(a) (b) (c) (d)

Figure 10 App interfaces

Table 6 Confusion matrix of the classification results using the full procedure of the proposed method (using SVM with TLBO)

Determined levelGround-truth level

SVM optimized with TLBO0 1 2 3 4

0 38611 2233 1513 44 801 312 831 349 1 12 570 696 5706 701 3933 3 0 215 408 1234 35 2 76 60 609Accuracy () 9767 2209 7260 3361 5050Overall accuracy () 8617

12 Mobile Information Systems

Data Availability

All experiment data come from a Kaggle contest calledldquoIdentify signs of diabetic retinopathy in eye imagesrdquo(website httpswwwkagglecomcdiabetic-retinopathy-detection)

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e authors would like to thank Mr Chia-Ming Hu for hishelp in preparing computational instruments and per-forming parts of the experiments in this paper

References

[1] N H Cho J E Shaw S Karuranga et al ldquoIDF Diabetes Atlasglobal estimates of diabetes prevalence for 2017 and pro-jections for 2045rdquo Diabetes Research and Clinical Practicevol 138 pp 271ndash281 2018

[2] American Diabetes Association Eye Complications AmericanDiabetes Association Arlington VA USA 2013 httpwwwdiabetesorgliving-with-diabetescomplicationseye-complicationsreferrerhttpswwwgooglecomtw

[3] International Diabetes Foundation (IDF) Diabetes AtlasInternational Diabetes Federation Brussels Belgium 7thedition 2015

[4] C P Wilkinson F L Ferris R E Klein et al ldquoProposedinternational clinical diabetic retinopathy and diabetic

macular edema disease severity scalesrdquo Ophthalmologyvol 110 no 9 pp 1677ndash1682 2003

[5] A Tufail C Rudisill C Egan et al ldquoAutomated diabeticretinopathy image assessment software diagnostic accuracyand cost-effectiveness compared with human gradersrdquoOphthalmology vol 124 no 3 pp 343ndash351 2017

[6] A Tufail V V Kapetanakis S Salas-Vega et al ldquoAn ob-servational study to assess if automated diabetic retinopathyimage assessment software can replace one or more steps ofmanual imaging grading and to determine their cost-effec-tivenessrdquo Health Technology Assessment vol 20 no 92pp 1ndash72 2016

[7] I Liu and Y Sun ldquoRecursive tracking of vascular networks inangiograms based on the detection-deletion schemerdquo IEEETransactions on Medical Imaging vol 12 no 2 pp 334ndash3411993

[8] A Can H Shen J N Turner H L Tanenbaum andB Roysam ldquoRapid automated tracing and feature extractionfrom retinal fundus images using direct exploratory algo-rithmsrdquo IEEE Transactions on Information Technology inBiomedicine vol 3 no 2 pp 125ndash138 1999

[9] M Vlachos and E Dermatas ldquoMulti-scale retinal vesselsegmentation using line trackingrdquo Computerized MedicalImaging and Graphics vol 34 no 3 pp 213ndash227 2010

[10] Y Yin M Adel and S Bourennane ldquoAn automatic trackingmethod for retinal vascular tree extractionrdquo in Proceedings ofthe IEEE International Conference on Acoustics Speech andSignal Processing IEEE Kyoto Japan March 2012

[11] S Chaudhuri S Chatterjee N Katz M Nelson andM Goldbaum ldquoDetection of blood vessels in retinal imagesusing two-dimensional matched filtersrdquo IEEE Transactions onMedical Imaging vol 8 no 3 pp 263ndash269 1989

Table 8 Confusion matrix of the classification using the baseline method [39] and its overall accuracy

DeterminedOriginal

Random forest0 1 2 3 4

0 38228 2001 1331 38 641 546 1026 434 3 22 710 732 5726 635 3583 2 1 279 465 1464 45 2 89 73 636Accuracy () 9670 2727 7286 3830 5274Overall accuracy () 8601

Table 7 Differences between Tables 5 and 6

Determined lesionGround-truth lesion

SVM-network 1 SVM-network 20 1 2 3 4 0 1 2 3 4

0 +420 minus460 minus848 minus31 minus80 +310 0 minus104 minus43 +21 +194 +393 +164 +1 0 minus15 +96 +59 0 minus22 +231 +70 +648 minus37 +1 minus220 minus90 +197 +81 +483 +3 0 +45 +76 +21 +1 minus1 minus123 minus78 minus624 +3 minus3 +45 +76 +58 minus65 minus5 minus28 +4 +14Accuracy increment () minus108 8977 1280 2288 1052 078 1305 369 minus1603 23Overall accuracy increment () 16 116

Mobile Information Systems 13

characteristic of vessels causing tracking errors Hooveret al [12] described an improvedmethod that considers localand regional characteristics of vessels to separate bloodvessels in retinal images and iteratively determine whetherthe current point is a vessel point

After such improvement a large number of studies ofreformed filters have been developed Jiang and Mojon [13]promoted a generalized threshold method based on amultithreshold detection Zhang et al [14] improved thematching filtering method by applying a local vessel crosssection analysis using local bilateral thresholding Li et al[15] suggested a multiscale production of the matched filterto enhance the extraction of tiny vessels

23 Methods of Morphological Processing Morphologicalprocessing facilitates the segmentation and identification oftarget objects by analyzing and processing structural ele-ments in a binary image us linear and circular elementsof blood vessels can be selected isolating the desiredstructure instead of the background image In addition

morphological processing can also smooth and fill the imagecontour with the advantage of antinoise However thismethod overrelies on structural elements and does not makegood use of characteristics of vessels

According to vessel characteristics Zana and Klein [16]introduced a mathematical morphology-based algorithmthat allows separating the vessels from all possible un-desirable patterns Building on this approach Ayala et al[17] proposed using different average fuzzy sets In Miri andMahloojifar [18] fundus images were analyzed by the use ofcurvelet transform and morphological reconstruction ofmultistructural elements to enhance the boundaries anddetermine the vascular ridge Karthika and Marimuthu [19]combined curvelet transform and morphological re-construction of multistructural elements with stronglyconnected component analysis (SCCA) to segment andidentify vessels

24 Methods of Deformation Models First introduced byKass et al [20] in 1988 the key benefit of deformation

Level 0Level 1Level 2Level 3Level 4

Level 4

Big datacollection

Onlinetraining

Deep learningalgorithm on

cloud computingplatform

Image acquisitionon mobile device

Figure 3 e proposed framework It is composed of mobile block (image acquisition) cloud block (deep learning algorithm on cloudcomputing platform) and big data block (big data collection and training)

Figure 2 Commercialized fundus camera used in the clinics e machine is bulky and expensive making it difficult for large-scalescreening

4 Mobile Information Systems

models is the ability to produce smooth parametric curves orsurfaces Two categories of deformation models are iden-tified parametric deformation and geometric deformationParametric deformation models are also called active con-tour or snake models (set of points each with an associatedenergy) rough the external and internal forces acting onthe snake the snake model can change its shape andsmoothness toward the desired structure In 2007 Esponaet al [21] used a parametric deformation model method onfundus images and promoted an improved method withmorphological segmentation With the assistance of mor-phological vessel segmentation the snake model expands tothe contour of the obtained vessels until the local energyfunction is minimal Another deformation model methodcalled ribbon of twins (ROT) which combines ribbon snakesand double snakes was proposed by Al-Diri and Hunter[22] Each twin consists of two snakes one inside and oneoutside the vessel edges e double snake model then at-tempts to integrate the pairs of twins on the vessel bordersinto a single ribbon and calculate the vessel width

ere are several shortcomings in parametric de-formation models For instance the segmentation resultsdepend on the initial contour and difficulties arise whenextending from low to high dimensions and in segmentingcomplex objects Geometric deformation can well solve theproblems caused by parametric deformable models Geo-metric deformable models are based on deformation curveevolution theory and have no strict requirement on theposition of the initial contour which increases the robust-ness of the method and allows it to be extended to highdimensions Zhang et al [23] proposed an automatic vesselsegmentation method which uses nonlinear orthogonalprojection to capture the characteristics of retinal bloodvessels and obtained an adaptive local thresholding algo-rithm for blood vessel segmentation Zhao et al [24] sug-gested a retinal vessel segmentation method that employs aregion-based active contour model with a level set imple-mentation and a region growing model

25 Methods of Deformation Model Machine learning is analgorithm that teaches computers to learn to achieve goalsautomatically by building generative or discriminativemodels from accumulated datasets Machine learning can bedivided into supervised learning and unsupervised learninge supervised learning methods learn to achieve goalsbased on ground-truth which means that during thetraining stage the training data used to train the model comewith a ldquolabelrdquo that can be used by the machine learningalgorithm to differentiate the data Applying such paradigmin the problem of DR it means that when using supervisedlearning one needs to mark all of the pixels belonging tovessels in advance whereas the unsupervised learningmethod does not need to mark them beforehand

For supervised learning Cesar and Jelinek [30] andLeandro et al [31] proposed a supervised classification withtwo-dimensional Gabor wavelet Each pixel has a featurevector that consists of the gray-scale feature and responses ofdistinct sizes of two-dimensional Gabor wavelet Ricci and

Perfetti [25] proposed a segmentation method for retinalvessels based online manipulation and support vectorclassification Since the features are extracted by twoorthogonal vertical lines it reduces the features andtraining samples in supervised learning A supervisedmethod using neural network was proposed by Marinet al [26] which has one input layer three hidden layersand one output layer Each pixel in the image is repre-sented by a seven-dimensional feature vector to train thenetwork Shanmugam and Banu [27] used an extremelearning machine (ELM) to detect retinal vessels bycreating a seven-dimensional feature vector based ongray-scale features and invariant moments and using ELMto segment vessels In 2015 Wang et al [28] raised a newhierarchical retinal vascular segmentation includingthree steps preprocessing hierarchical feature extractionand integration classification It involves using simplelinear iterative clustering (SLIC) to perform super-pixelsegmentation and randomly selecting a pixel to representthe entire super-pixel as a more easy and efficient meansof extracting features

For unsupervised learning in 1998 Tolias and Panas[32] created an automatic and unsupervised segmentationmethod based on blurred fundus images which used fuzzyC-means (FCM) to find initial candidate points Xie and Nie[33] proposed a segmentation method based on a geneticalgorithm and FCM Salazar-Gonzalez et al [29] usedmethods of vector flow to segment retinal vessels

Table 2 is a summary about the performance comparisonbetween different existing methods

3 Literature Review of DiabeticRetinopathy Detection

Although extracting vessels before detecting DR with machinelearning can achieve high accuracy it is time-consuming tocreate the marked ground-truth for retinal vessels Anotherparadigm is to train the computer to automatically learn how todistinguish levels of DR by reading retinal images directlywithout performing vessel segmentation In 2000 Ege et al [34]proposed an automatic analysis of DR by different statisticalclassifiers including Bayesian Mahalanobis and k-nearestneighbor Silberman et al [35] introduced an automatic de-tection system for DR and reported an equal error rate of 87Karegowda et al [36] tried to detect exudates in retinal imagesusing back-propagation neural networks (BPN) eir featureswere decided by two methods decision trees and genetic al-gorithmswith correlation-based feature selection (GA-CFS) Intheir experiment the best BPN performance showed 9845accuracy Kavitha and Duraiswamy [37] did some research onautomatic detection of hard and soft exudates in fundus im-ages using color histogram thresholding to classify exudateseir experiments showed 9907 accuracy 89 sensitivityand 99 specificity In 2014 de la Calleja et al [38] used localbinary patterns (LBP) to extract local features and artificialneural networks random forest (RF) and support vectormachines (SVM) for detection In using a dataset containing 71images their best result achieved 9746 accuracy with RF

Mobile Information Systems 5

4 Material and Methods

We propose an automatic DR detection algorithm based onDCNN fractional max-pooling [39] SVM [40] andteaching-learning-based optimization (TLBO) [41] Specif-ically we train two DCNN networks with fractional max-pooling combining their prediction results using SVM andoptimizing the SVM parameters with TLBO e reason fortraining two distinct networks is that different networkarchitectures may have their unique advantages in featurespace representation By training two DCNNs and com-bining their features the prediction accuracy can be furtherenhanced Another important factor impacting the recog-nition rate is the parameter of classifiers We propose tooptimize the SVM parameters using TLBOWe illustrate theimage preprocessing methods in Section 41 and present thefractional max-pooling SVM and TLBO in Sections 4243 and 44 respectively

41 Preprocessing Given the vessels in the original fundusimages are mostly not very clear and the size of eachfundus image may differ it is essential to preprocess imagesso that they have the same size and the visibility of thevessels is improved ere are three steps in preprocessinge first is to rescale images to the same size Since thefundus images are circular we rescale the input images sothat the diameter of the fundus images becomes 540 pixelsAfter rescaling the local average color value is subtractedfrom the rescaled images and another transformation isperformed so that the local average is mapped to 50 gray-scale in order to remove the color divergence caused bydifferent ophthalmoscopes Last but not least becauseboundary effects may occur in some images we remove the

periphery by clipping 10 from the border of the imagesFigure 4 shows the original fundus image and the imageafter preprocessing

42 Fractional Max-Pooling Pooling is a procedure thatturns the input matrix Min times Nin into a smaller Mout times

Nout output matrix e purpose is to divide the inputmatrix into multiple pooling regions (Pij)

pi sub 1 2 3 Min1113864 1113865 for each i isin 1 Mout1113864 1113865

pj sub 1 2 3 Nin1113864 1113865 for each j isin 1 Nout1113864 1113865

Pij pi times pj

(1)

e pooling results are computed according to poolingtype

Outputij Oper

(k l) isin Pij

Inputkl (2)

In equation (2) ldquoOperrdquo refers to a particular mathe-matical operation For example if max-pooling is used theoperation will be to take the maximum of the input regionFor average pooling the average of the input region is takenFor such a network that requires tremendous learning it ispreferable to use as many hidden layers as possible In thiswork the pooling layer used in our networks is fractionalmax-pooling instead of general max-pooling

Fractional max-pooling is a pooling scheme thatmakes the size of the output matrix equivalent to fractionaltimes that of the input matrix after pooling ieNin αNout αge 0 α notin Z To describe the general

Table 2 Performance comparison

Method Year Reference Sn Sp Acc AUC Database

Vascular tracking 2010 [9] 07468 09551 09285 mdash DRIVE2012 [10] 06887 09562 09290 mdash STARE

Matched filtering2009 [14] 06611 09848 09497 mdash STARE

2012 [15] 07191 09687 09407 mdash STARE07154 09716 09343 mdash DRIVE

Morphological processing 2011 [18] 07352 09795 09458 mdash DRIVE2014 [19] 07862 09815 09598 mdash DRIVE

Deformation model

2007 [21] 06634 09682 09316 mdash DRIVE

2009 [23] mdash 09736 09087 mdash STAREmdash 09772 09610 mdash DRIVE

2014 [24] 07187 09767 09509 mdash STARE07354 09789 09477 mdash DRIVE

Machine learning

2007 [25] mdash mdash 09584 09602 STAREmdash mdash 09563 09558 DRIVE

2011 [26] 06944 09819 09526 mdash STARE07067 09801 09452 mdash DRIVE

2013 [27] 08194 09679 09725 mdash DRIVE

2014 [28] 08104 09791 09813 09751 STARE08173 09733 09767 09475 DRIVE

2014 [29] 07887 09633 09441 mdash STARE07512 09684 09412 mdash DRIVE

Sn sensitivity Sp specificity Acc accuracy AUC area under curve

6 Mobile Information Systems

pooling regions let (xi)Mouti0 and (yj)

Noutj0 be two increasing

integer sequences starting with one and ending with Min orNin ese two sequences are used in pooling steps asdescribed in Figures 5 and 6

Pij ximinus1 xi minus a1113858 1113859 times yjminus1 yj minus b1113960 1113961 (3)

e constants a and b in equation (3) stand for theoverlapping length and the width of the pooling windowrespectively Figure 5 is a simple example of overlappingpooling Figure 6 illustrates different pooling region types

After fractional max-pooling the pooling window size isstill integers but the global pooling size will change Namelyfractional max-pooling does not directly change the poolingwindow into a fractional scale Instead it uses windows ofvariable size to achieve fractional pooling e generation of x

and y sequences can be random or pseudorandom Pseudo-random sequences generate more stable pooling regions thanrandom sequences and can also achieve higher accuracy [39]

43 Support Vector Machine (SVM) SVM is a supervisedlearning method used for classification and regressionanalysis SVM can find the hyperplane or decision boundarydefined by the solution vector w which not only separatesthe training vectors but also works well with unseen testdata To improve its generalization ability SVM selectsdecision boundaries based on maximizing margins betweenclasses

Figure 7 illustrates the idea Suppose there are n points ina binary dataset

x1rarr

y1( 1113857 xnrarr

yn( 1113857 (4)

where yi is the data label which can be 1 or minus1 indicatingthe class to which xi

rarr belongs We need to find the optimizedhyperplane such that the distance between the hyperplaneand its nearest point xi

rarr is maximized A hyperplane can bewritten as equation (5) based on x

rarr

wrarr

middot xrarrminus b 0 (5)

where wrarr is the normal vector of the hyperplane and the

value of bwrarr

decides the margin of hyperplane from thetraining data point along the normal vector w

rarr

(a) (b)

(c) (d)

Figure 4 Image before and after the preprocessing stage (a b) images before preprocessing stage (c d) images after preprocessing stage

x = 1 2 3 4 5 y = 1 2 4 5 a = 0 b = 1P11 = [x1minus1 x1 minus 0] times [y1minus1 y1 minus 1] = [1 2] times [1 2]P12 = [x1minus1 x1 minus 1] times [y2minus1 y2 minus 1] = [1 2] times [2 3]

Figure 5 An overlapping example e blue solid line indicatespooling region P11 while the red dotted line shows the P12

Mobile Information Systems 7

For yi whose value is 1 the data must satisfywrarr

middot xrarrminus bge 1 and for yi whose value is minus1 w

rarrmiddot xrarrminus bleminus1

has to be satisfied Combining these two conditions we get

yi(wrarr

middot xrarrminus b)ge 1 (6)

e goal is to maximize bwrarr

according to the constrainof equation (6) in order to derive the optimized decisionhyperplane for classification

Sometimes the training data might not be able to beperfectly separated using linear boundarieserefore in theSVM formulation we need to introduce the error metric εand the cost parameter C as shown in equation (7) e goalnow becomes to minimize

12

wrarr

middot wrarr

+ C 1113944N

i1εi (7)

Subject to

wrarr

middot xirarr

+ bge 1minus εi if wrarr

middot xirarrgt b

wrarr

middot xirarr

+ bleminus1 + εi if wrarr

middot xirarrlt b

εi gt 0

(8)

e performance of SVM is influenced by two mainparameters the first one is C which is a tunable parameter inequation (7) e other one is c which is used in the radialbasis function (RBF) kernel to map data into a higher di-mensional space before training and classification e RBFkernel can be defined as

K xi xj1113872 1113873 eminusc ximinusxj

111386811138681113868111386811138681113868111386811138682

(9)

where c denotes the width of the Gaussian envelope in ahigh-dimensional feature space

44 Teaching-Learning-Based Optimization (TLBO)TLBO an evolution-based optimization algorithm wasproposed by Rao et al [41] in 2011 e concept of TLBO isinspired by the evolution of the learning process when agroup or a class of learners learn a target task ere are two

ways of learning in groups or classes (1) learning from theguidance of the teacher and (2) learning from other learnerse procedure of TLBO can be divided into two phases asdescribed below in Sections 441 and 442

441 Teacher Phase In the whole population the teacher(Xteacher) can be considered as the best solution Namelylearners learn from the teacher in the teacher phase In thisphase the teacher strives for enhancing the results of otherindividuals (Xi) by increasing the mean result of theclassroom is can be described as adjusting Xmean toapproximate Xteacher In order to maintain a stochasticnature during the optimization process two randomlygenerated parameters r and TF are applied in each iterationfor the solution Xi as

Xnew Xi + ri Xteacher minusTFXmean( 1113857 (10)

TF round[1 + rand(0 1)] (11)

In equation (10) ri is a randomly selected number in therange of 0 and 1 Moreover Xnew and Xi are the new andexisting solutions at iteration i respectively TF in equation(11) is a teaching factor which can be either 1 or 2

a = 1 b = 1

x = 1 2 hellip y = 1 2 hellip

x = 1 3 hellip y = 1 2 hellip

x = 1 2 hellip y = 1 3 hellip

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 1] times [1 1]

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 2] times [1 1]

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 1] times [1 2]

Figure 6 Different pooling types when both a and b equal to onee left side shows the equations for pooling condition and computationof pooling region Pictures on the right side indicate regions before and after pooling

w middot x ndash

b = 1

w middot x ndash

b = 0

w middot x ndash

b = ndash1

Figure 7 Example of decision boundary hyperplane with twoclasses of samples

8 Mobile Information Systems

442 Learner Phase e learners gain their knowledge byinteracting with each other erefore an individual learnsnew information if other individuals have more knowledgethan him or her In this phase the student Xi interactsrandomly with another student Xj (ine j) in order to en-hance his or her knowledge Equation (12) shows that if Xi isbetter than Xj (ie f(Xj)gtf(Xi) for minimizationproblems) Xj is moved toward Xi Otherwise it is movedaway from Xi

Xnew Xi + ri Xi minusXj1113872 1113873 f Xj1113872 1113873gtf Xi( 1113857

Xi + ri Xj minusXi1113872 1113873 f Xi( 1113857gtf Xj1113872 1113873

⎧⎪⎨

⎪⎩(12)

If the new solution Xnew to the problem is better than theold ones the new solution Xnew will be recorded as the bestsolution After updating the status of each learner a newiteration begins A stop criterion based on the iterationnumber or the difference of the cost function can be set tostop the iteration properly e flowchart of TLBO is shownin Figure 8

5 Result

Our fundus image data is from the database provided by oneof the Kaggle contests entitled ldquoIdentify signs of diabeticretinopathy in eye imagesrdquo [42] In this database there areabout 90000 images We separate 1000 images from thetraining dataset to be the validation dataset e detailedinformation of each dataset is shown in Table 3 and our twonetwork architectures are shown in Figure 9

Our proposed method uses two DCNNs with fractionalmax-pooling layers For every input fundus image the twoDCNN will output a vector of size 1 times 5 representing theprobability distribution of the prediction for each lesion(category) e probability distribution together with othervalues forms a feature with dimensionality 24 e 24features are described as follows

(i) DCNN probabilities of each lesion respectively (5features)

(ii) Averages of R G and B channel values within 50 lowast50 center cropped image (3 features)

(iii) Widths and heights of 50 lowast 50 center croppedimage (2 features)

(iv) Overall standard deviation of the original image and50 lowast 50 centered cropped image Laplacian-filtered image (2 features)

(v) In total there are 12 features for one fundusimage We then append another 12 features fromthe fundus image of the other eye of the samesubject erefore the overall length of the featurevector is 24 for one fundus image e 24 featurevectors of dimensionality are used as input vec-tors of SVM

e 24-dimensional vector is used to train a multiclassSVM (five classes) whose parameters are optimized usingthe TLBO method We implemented the method described

in [39] and used it as the baseline e baseline system usessimilar features with a scheme of ensemble classifier (RF)

We used the validation set data to optimize the pa-rameter set (C c) in SVM using TLBO e upper andlower bounds of the parameter are set within [0 100] Weran 50 iterations with 50 students

Our final accuracy for five-class classification task ofDR is 8617 and the accuracy for the binary classclassification task is 9105 Labels for five-class classi-fication are normal NPDR level 1 NPDR level 2 NPDRlevel 3 and PDR while labels for binary class classifi-cation are normal and abnormal For binary classifica-tion its sensitivity is 08930 while the specificity is09089 Except counting accuracy we also do a T-test forour binary class classification e T-test is also called theStudentrsquos t-test It is a statistical hypothesis test in whichthe test statistic follows a Studentrsquos t-distributionUsually the t-test is used to compare whether there is asignificant difference between two groups of data andassists in judging the data divergence In doing a pairedsamples t-test with results from binary class classificationand random judgment its outcome is 1 for the hypothesistest result zero for the p value and [03934 04033] forthe confidence interval under null hypothesis at the 5significant level

e hypothesis test result is an index that tells whethertwo data come from the same distribution or not If the datacome from the same distribution the value of the hypothesistest result will be close to 0 On the contrary if the dataresources are distinct the result will be close to 1 whichmeans there is a differentiation between the datae p valueis the probability of accepting the assumption that there is adifference between two data may be wronge smaller the p

value the more reason that there is a disparity between dataAlso we designed an app called ldquoDeep Retinardquo pro-

viding personal examination remote medical care and earlyscreening Figure 10 shows our app interfaces Afterchoosing a fundus image that the user wants to check it willsend the image to our server and use our designed machinelearning method It takes about 10 s (depends on networkspeed) to get the result which will be presented as theprobability of each lesion With a handheld device in-dividuals can do the initial examination at the district officeor even at home More importantly it can benefit someremote areas that lack medical resources

6 Discussion

61 Accuracy Improvement Table 4 shows the accuracycomparison when using different classifiers and pa-rameter optimization methods on each dataset Using thedefault parameters with SVM (without optimization)accuracies in both validation and test sets are higher thanthat of the RF [39] If we optimize the parameters usingthe default parameter searching method provided in theLIBSVM software package though it achieves very highaccuracy in the five-fold cross validation experiment the

Mobile Information Systems 9

validation and test accuracies are even lower than thedefault one From this result we believe that overttingarises when optimizing parameters in SVM

Table 5 shows the confusion matrix of the classicationresults from the two DCNN networks (before performingSVM classication) Network 1 is the architecture shown on

Yes

No

Reject

Yes

No

No

Yes

Yes

Start

Initial students terminal criterion

i = 1

Calculate mean of each design variables

Identify best teachersolution

Based on best teachersolution to modifyXnewi = Xoldi + differencei = Xoldi +

ri (Xteacher ndash (TF)mean)

Randomly select any solution Xj (Xj ne Xi)

i = i + 1

Final value of solution and end

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is Xi better than Xjf (Xi) gt f (Xj)

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is termination criteria satisfiedf (Xnewi) gt f (Xoldi)

No

Yes

No

Xnewi = Xoldi + ri (Xi ndash Xj)

Xi = Xnewi

Xnewi = Xoldi + ri (Xj ndash Xi)

i le N

Xi = Xnewi

Figure 8 TLBO owchart

10 Mobile Information Systems

the left side of Figure 9 and network 2 is the one on the rightside From Table 5 it shows that the lesion classications of 0and 2 are better than the other categories For lesion 1 mostof the prediction results are incorrect Also for lesions 3 and4 the majority of the results are misclassications that areclassied into lesion 2

Table 6 shows the confusion matrix of the classicationresults using the full procedure of the proposed method(using SVM with TLBO) Table 7 displays the dierencebetween Tables 5 and 6 which serves as a performancecomparison between the two methods (using DCNN onlyand DCNN + SVM + TLBO) From the table every classexcept class 0 and overall accuracy is increased in network 1For network 2 each accuracy except class 3 is increasede decline in accuracy of class 3 is mainly caused by

misclassication of class 2 Table 8 shows the confusionmatrix of the classication results using the baseline methodas reported in [39] for comparison purposes

62 Deep Learning vs Traditional Classication MethodsMany traditional classication methods try to solve theproblem of DR detection by (1) using image processing tocapture symptoms in fundus images and then (2) building aclassier to make decisions based on the detected symptoms(1) e shortcoming of image processing methods is that themanifestations of the symptoms are random across dierentimages therefore it is extremely time-consuming and re-quires intense eorts to label the locations of the symptomsAbiding by the new philosophy that comes with the emer-gence of the deep learning technology our proposed methodis trying to learn how to make decisions directly from theimage data itself Dierent than the former approaches ourimages only need to be labeled with lesion number instead oflabeling symptom locations Consequently it saves consid-erable time during the database preprocessing stage On top ofthe classication results by the two DCNN networks we useSVM optimized by TLBO to generate an improved outcomeand we achieve 8617 accuracy Our result is better than therst-place winner in the Kaggle competition It shows that ourresearch result is the state-of-the-art

63 Limitation In our current datasets the number ofimages of lesions 3 and 4 is not suiexclcient to train a networkwhich is a limitation of the proposed methoderefore oneof our future works is to develop deeper collaborative re-lations with hospitals and clinics to acquire more data oflesions 3 and 4 With more data we believe the classicationaccuracy will be further increased In addition from ourresult we found that it is hard to dierentiate the imagesbetween lesions 0 and 1 erefore when we collect newdata it is desirable to collect more images belonging tolesions 0 and 1 Also we can attempt to use a dierentnetwork architecture for this problem

7 Conclusion

It is feasible to train a deep learning model for automaticdiagnosis of DR as long as we have enough data for sta-tistical model training Furthermore the database prepa-ration stage only needs a categorical label for each trainingimage It does not require detailed annotation for retinalvessel tracking in every image Hence it is time-eiexclcientcompared to the traditional machine learning-based methodfor automatic diagnosis of DR e nal accuracy canachieve 8617 and 9105 for ve-class and binary classclassications respectively

e sensitivity and specicity of binary classication are08930 and 09089 respectively which is a satisfactory resultFurthermore we developed an automatic inspection appthat can be used in both personal examination and remotemedical care With more image data collected we expect theaccuracy can be even more enhanced further improving oursystem

Table 3 Detailed information of each dataset

Dataset of imagesDR lesions

0 1 2 3 4Train 34124 73 6 15 3 3Validation 1000 70 8 15 4 3Test 53572 74 7 15 2 2

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

3 times 3 conv 320

3 times 3 conv 352

3 times 3 conv 384

2 times 2 conv 416

1 times 1 conv 448

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

Softmax

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

2 times 2 conv 320

1 times 1 conv 356

Softmax

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

Figure 9 Architecture of the two DCNN networks that we used

Mobile Information Systems 11

Table 4 Accuracy comparison when using different classifiers and different parameter optimization methods on each dataset

DatasetMethod

SVM RF [39]Without optimization Default parameter searching TLBO

Five-fold cross validation (using training data) 88744 965724 mdash mdashValidation data 854 796 mdash 847Test data 861177 810573 8617 8602

Table 5 Confusion matrix of the classification results from the two DCNN networks before performing SVM classification

Determined levelGround-truth level

Network 1 Network 20 1 2 3 4 0 1 2 3 4

0 39031 2693 2361 75 160 38310 2233 1617 51 781 118 438 185 0 1 327 735 290 1 32 339 626 5058 738 392 790 786 5509 620 3453 0 0 170 332 102 4 1 338 486 1854 43 5 85 69 551 100 7 105 56 595Accuracy () 9874 1164 6436 2735 4569 9691 1954 7001 4003 4934Overall accuracy () 8476 8518

(a) (b) (c) (d)

Figure 10 App interfaces

Table 6 Confusion matrix of the classification results using the full procedure of the proposed method (using SVM with TLBO)

Determined levelGround-truth level

SVM optimized with TLBO0 1 2 3 4

0 38611 2233 1513 44 801 312 831 349 1 12 570 696 5706 701 3933 3 0 215 408 1234 35 2 76 60 609Accuracy () 9767 2209 7260 3361 5050Overall accuracy () 8617

12 Mobile Information Systems

Data Availability

All experiment data come from a Kaggle contest calledldquoIdentify signs of diabetic retinopathy in eye imagesrdquo(website httpswwwkagglecomcdiabetic-retinopathy-detection)

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e authors would like to thank Mr Chia-Ming Hu for hishelp in preparing computational instruments and per-forming parts of the experiments in this paper

References

[1] N H Cho J E Shaw S Karuranga et al ldquoIDF Diabetes Atlasglobal estimates of diabetes prevalence for 2017 and pro-jections for 2045rdquo Diabetes Research and Clinical Practicevol 138 pp 271ndash281 2018

[2] American Diabetes Association Eye Complications AmericanDiabetes Association Arlington VA USA 2013 httpwwwdiabetesorgliving-with-diabetescomplicationseye-complicationsreferrerhttpswwwgooglecomtw

[3] International Diabetes Foundation (IDF) Diabetes AtlasInternational Diabetes Federation Brussels Belgium 7thedition 2015

[4] C P Wilkinson F L Ferris R E Klein et al ldquoProposedinternational clinical diabetic retinopathy and diabetic

macular edema disease severity scalesrdquo Ophthalmologyvol 110 no 9 pp 1677ndash1682 2003

[5] A Tufail C Rudisill C Egan et al ldquoAutomated diabeticretinopathy image assessment software diagnostic accuracyand cost-effectiveness compared with human gradersrdquoOphthalmology vol 124 no 3 pp 343ndash351 2017

[6] A Tufail V V Kapetanakis S Salas-Vega et al ldquoAn ob-servational study to assess if automated diabetic retinopathyimage assessment software can replace one or more steps ofmanual imaging grading and to determine their cost-effec-tivenessrdquo Health Technology Assessment vol 20 no 92pp 1ndash72 2016

[7] I Liu and Y Sun ldquoRecursive tracking of vascular networks inangiograms based on the detection-deletion schemerdquo IEEETransactions on Medical Imaging vol 12 no 2 pp 334ndash3411993

[8] A Can H Shen J N Turner H L Tanenbaum andB Roysam ldquoRapid automated tracing and feature extractionfrom retinal fundus images using direct exploratory algo-rithmsrdquo IEEE Transactions on Information Technology inBiomedicine vol 3 no 2 pp 125ndash138 1999

[9] M Vlachos and E Dermatas ldquoMulti-scale retinal vesselsegmentation using line trackingrdquo Computerized MedicalImaging and Graphics vol 34 no 3 pp 213ndash227 2010

[10] Y Yin M Adel and S Bourennane ldquoAn automatic trackingmethod for retinal vascular tree extractionrdquo in Proceedings ofthe IEEE International Conference on Acoustics Speech andSignal Processing IEEE Kyoto Japan March 2012

[11] S Chaudhuri S Chatterjee N Katz M Nelson andM Goldbaum ldquoDetection of blood vessels in retinal imagesusing two-dimensional matched filtersrdquo IEEE Transactions onMedical Imaging vol 8 no 3 pp 263ndash269 1989

Table 8 Confusion matrix of the classification using the baseline method [39] and its overall accuracy

DeterminedOriginal

Random forest0 1 2 3 4

0 38228 2001 1331 38 641 546 1026 434 3 22 710 732 5726 635 3583 2 1 279 465 1464 45 2 89 73 636Accuracy () 9670 2727 7286 3830 5274Overall accuracy () 8601

Table 7 Differences between Tables 5 and 6

Determined lesionGround-truth lesion

SVM-network 1 SVM-network 20 1 2 3 4 0 1 2 3 4

0 +420 minus460 minus848 minus31 minus80 +310 0 minus104 minus43 +21 +194 +393 +164 +1 0 minus15 +96 +59 0 minus22 +231 +70 +648 minus37 +1 minus220 minus90 +197 +81 +483 +3 0 +45 +76 +21 +1 minus1 minus123 minus78 minus624 +3 minus3 +45 +76 +58 minus65 minus5 minus28 +4 +14Accuracy increment () minus108 8977 1280 2288 1052 078 1305 369 minus1603 23Overall accuracy increment () 16 116

Mobile Information Systems 13

models is the ability to produce smooth parametric curves orsurfaces Two categories of deformation models are iden-tified parametric deformation and geometric deformationParametric deformation models are also called active con-tour or snake models (set of points each with an associatedenergy) rough the external and internal forces acting onthe snake the snake model can change its shape andsmoothness toward the desired structure In 2007 Esponaet al [21] used a parametric deformation model method onfundus images and promoted an improved method withmorphological segmentation With the assistance of mor-phological vessel segmentation the snake model expands tothe contour of the obtained vessels until the local energyfunction is minimal Another deformation model methodcalled ribbon of twins (ROT) which combines ribbon snakesand double snakes was proposed by Al-Diri and Hunter[22] Each twin consists of two snakes one inside and oneoutside the vessel edges e double snake model then at-tempts to integrate the pairs of twins on the vessel bordersinto a single ribbon and calculate the vessel width

ere are several shortcomings in parametric de-formation models For instance the segmentation resultsdepend on the initial contour and difficulties arise whenextending from low to high dimensions and in segmentingcomplex objects Geometric deformation can well solve theproblems caused by parametric deformable models Geo-metric deformable models are based on deformation curveevolution theory and have no strict requirement on theposition of the initial contour which increases the robust-ness of the method and allows it to be extended to highdimensions Zhang et al [23] proposed an automatic vesselsegmentation method which uses nonlinear orthogonalprojection to capture the characteristics of retinal bloodvessels and obtained an adaptive local thresholding algo-rithm for blood vessel segmentation Zhao et al [24] sug-gested a retinal vessel segmentation method that employs aregion-based active contour model with a level set imple-mentation and a region growing model

25 Methods of Deformation Model Machine learning is analgorithm that teaches computers to learn to achieve goalsautomatically by building generative or discriminativemodels from accumulated datasets Machine learning can bedivided into supervised learning and unsupervised learninge supervised learning methods learn to achieve goalsbased on ground-truth which means that during thetraining stage the training data used to train the model comewith a ldquolabelrdquo that can be used by the machine learningalgorithm to differentiate the data Applying such paradigmin the problem of DR it means that when using supervisedlearning one needs to mark all of the pixels belonging tovessels in advance whereas the unsupervised learningmethod does not need to mark them beforehand

For supervised learning Cesar and Jelinek [30] andLeandro et al [31] proposed a supervised classification withtwo-dimensional Gabor wavelet Each pixel has a featurevector that consists of the gray-scale feature and responses ofdistinct sizes of two-dimensional Gabor wavelet Ricci and

Perfetti [25] proposed a segmentation method for retinalvessels based online manipulation and support vectorclassification Since the features are extracted by twoorthogonal vertical lines it reduces the features andtraining samples in supervised learning A supervisedmethod using neural network was proposed by Marinet al [26] which has one input layer three hidden layersand one output layer Each pixel in the image is repre-sented by a seven-dimensional feature vector to train thenetwork Shanmugam and Banu [27] used an extremelearning machine (ELM) to detect retinal vessels bycreating a seven-dimensional feature vector based ongray-scale features and invariant moments and using ELMto segment vessels In 2015 Wang et al [28] raised a newhierarchical retinal vascular segmentation includingthree steps preprocessing hierarchical feature extractionand integration classification It involves using simplelinear iterative clustering (SLIC) to perform super-pixelsegmentation and randomly selecting a pixel to representthe entire super-pixel as a more easy and efficient meansof extracting features

For unsupervised learning in 1998 Tolias and Panas[32] created an automatic and unsupervised segmentationmethod based on blurred fundus images which used fuzzyC-means (FCM) to find initial candidate points Xie and Nie[33] proposed a segmentation method based on a geneticalgorithm and FCM Salazar-Gonzalez et al [29] usedmethods of vector flow to segment retinal vessels

Table 2 is a summary about the performance comparisonbetween different existing methods

3 Literature Review of DiabeticRetinopathy Detection

Although extracting vessels before detecting DR with machinelearning can achieve high accuracy it is time-consuming tocreate the marked ground-truth for retinal vessels Anotherparadigm is to train the computer to automatically learn how todistinguish levels of DR by reading retinal images directlywithout performing vessel segmentation In 2000 Ege et al [34]proposed an automatic analysis of DR by different statisticalclassifiers including Bayesian Mahalanobis and k-nearestneighbor Silberman et al [35] introduced an automatic de-tection system for DR and reported an equal error rate of 87Karegowda et al [36] tried to detect exudates in retinal imagesusing back-propagation neural networks (BPN) eir featureswere decided by two methods decision trees and genetic al-gorithmswith correlation-based feature selection (GA-CFS) Intheir experiment the best BPN performance showed 9845accuracy Kavitha and Duraiswamy [37] did some research onautomatic detection of hard and soft exudates in fundus im-ages using color histogram thresholding to classify exudateseir experiments showed 9907 accuracy 89 sensitivityand 99 specificity In 2014 de la Calleja et al [38] used localbinary patterns (LBP) to extract local features and artificialneural networks random forest (RF) and support vectormachines (SVM) for detection In using a dataset containing 71images their best result achieved 9746 accuracy with RF

Mobile Information Systems 5

4 Material and Methods

We propose an automatic DR detection algorithm based onDCNN fractional max-pooling [39] SVM [40] andteaching-learning-based optimization (TLBO) [41] Specif-ically we train two DCNN networks with fractional max-pooling combining their prediction results using SVM andoptimizing the SVM parameters with TLBO e reason fortraining two distinct networks is that different networkarchitectures may have their unique advantages in featurespace representation By training two DCNNs and com-bining their features the prediction accuracy can be furtherenhanced Another important factor impacting the recog-nition rate is the parameter of classifiers We propose tooptimize the SVM parameters using TLBOWe illustrate theimage preprocessing methods in Section 41 and present thefractional max-pooling SVM and TLBO in Sections 4243 and 44 respectively

41 Preprocessing Given the vessels in the original fundusimages are mostly not very clear and the size of eachfundus image may differ it is essential to preprocess imagesso that they have the same size and the visibility of thevessels is improved ere are three steps in preprocessinge first is to rescale images to the same size Since thefundus images are circular we rescale the input images sothat the diameter of the fundus images becomes 540 pixelsAfter rescaling the local average color value is subtractedfrom the rescaled images and another transformation isperformed so that the local average is mapped to 50 gray-scale in order to remove the color divergence caused bydifferent ophthalmoscopes Last but not least becauseboundary effects may occur in some images we remove the

periphery by clipping 10 from the border of the imagesFigure 4 shows the original fundus image and the imageafter preprocessing

42 Fractional Max-Pooling Pooling is a procedure thatturns the input matrix Min times Nin into a smaller Mout times

Nout output matrix e purpose is to divide the inputmatrix into multiple pooling regions (Pij)

pi sub 1 2 3 Min1113864 1113865 for each i isin 1 Mout1113864 1113865

pj sub 1 2 3 Nin1113864 1113865 for each j isin 1 Nout1113864 1113865

Pij pi times pj

(1)

e pooling results are computed according to poolingtype

Outputij Oper

(k l) isin Pij

Inputkl (2)

In equation (2) ldquoOperrdquo refers to a particular mathe-matical operation For example if max-pooling is used theoperation will be to take the maximum of the input regionFor average pooling the average of the input region is takenFor such a network that requires tremendous learning it ispreferable to use as many hidden layers as possible In thiswork the pooling layer used in our networks is fractionalmax-pooling instead of general max-pooling

Fractional max-pooling is a pooling scheme thatmakes the size of the output matrix equivalent to fractionaltimes that of the input matrix after pooling ieNin αNout αge 0 α notin Z To describe the general

Table 2 Performance comparison

Method Year Reference Sn Sp Acc AUC Database

Vascular tracking 2010 [9] 07468 09551 09285 mdash DRIVE2012 [10] 06887 09562 09290 mdash STARE

Matched filtering2009 [14] 06611 09848 09497 mdash STARE

2012 [15] 07191 09687 09407 mdash STARE07154 09716 09343 mdash DRIVE

Morphological processing 2011 [18] 07352 09795 09458 mdash DRIVE2014 [19] 07862 09815 09598 mdash DRIVE

Deformation model

2007 [21] 06634 09682 09316 mdash DRIVE

2009 [23] mdash 09736 09087 mdash STAREmdash 09772 09610 mdash DRIVE

2014 [24] 07187 09767 09509 mdash STARE07354 09789 09477 mdash DRIVE

Machine learning

2007 [25] mdash mdash 09584 09602 STAREmdash mdash 09563 09558 DRIVE

2011 [26] 06944 09819 09526 mdash STARE07067 09801 09452 mdash DRIVE

2013 [27] 08194 09679 09725 mdash DRIVE

2014 [28] 08104 09791 09813 09751 STARE08173 09733 09767 09475 DRIVE

2014 [29] 07887 09633 09441 mdash STARE07512 09684 09412 mdash DRIVE

Sn sensitivity Sp specificity Acc accuracy AUC area under curve

6 Mobile Information Systems

pooling regions let (xi)Mouti0 and (yj)

Noutj0 be two increasing

integer sequences starting with one and ending with Min orNin ese two sequences are used in pooling steps asdescribed in Figures 5 and 6

Pij ximinus1 xi minus a1113858 1113859 times yjminus1 yj minus b1113960 1113961 (3)

e constants a and b in equation (3) stand for theoverlapping length and the width of the pooling windowrespectively Figure 5 is a simple example of overlappingpooling Figure 6 illustrates different pooling region types

After fractional max-pooling the pooling window size isstill integers but the global pooling size will change Namelyfractional max-pooling does not directly change the poolingwindow into a fractional scale Instead it uses windows ofvariable size to achieve fractional pooling e generation of x

and y sequences can be random or pseudorandom Pseudo-random sequences generate more stable pooling regions thanrandom sequences and can also achieve higher accuracy [39]

43 Support Vector Machine (SVM) SVM is a supervisedlearning method used for classification and regressionanalysis SVM can find the hyperplane or decision boundarydefined by the solution vector w which not only separatesthe training vectors but also works well with unseen testdata To improve its generalization ability SVM selectsdecision boundaries based on maximizing margins betweenclasses

Figure 7 illustrates the idea Suppose there are n points ina binary dataset

x1rarr

y1( 1113857 xnrarr

yn( 1113857 (4)

where yi is the data label which can be 1 or minus1 indicatingthe class to which xi

rarr belongs We need to find the optimizedhyperplane such that the distance between the hyperplaneand its nearest point xi

rarr is maximized A hyperplane can bewritten as equation (5) based on x

rarr

wrarr

middot xrarrminus b 0 (5)

where wrarr is the normal vector of the hyperplane and the

value of bwrarr

decides the margin of hyperplane from thetraining data point along the normal vector w

rarr

(a) (b)

(c) (d)

Figure 4 Image before and after the preprocessing stage (a b) images before preprocessing stage (c d) images after preprocessing stage

x = 1 2 3 4 5 y = 1 2 4 5 a = 0 b = 1P11 = [x1minus1 x1 minus 0] times [y1minus1 y1 minus 1] = [1 2] times [1 2]P12 = [x1minus1 x1 minus 1] times [y2minus1 y2 minus 1] = [1 2] times [2 3]

Figure 5 An overlapping example e blue solid line indicatespooling region P11 while the red dotted line shows the P12

Mobile Information Systems 7

For yi whose value is 1 the data must satisfywrarr

middot xrarrminus bge 1 and for yi whose value is minus1 w

rarrmiddot xrarrminus bleminus1

has to be satisfied Combining these two conditions we get

yi(wrarr

middot xrarrminus b)ge 1 (6)

e goal is to maximize bwrarr

according to the constrainof equation (6) in order to derive the optimized decisionhyperplane for classification

Sometimes the training data might not be able to beperfectly separated using linear boundarieserefore in theSVM formulation we need to introduce the error metric εand the cost parameter C as shown in equation (7) e goalnow becomes to minimize

12

wrarr

middot wrarr

+ C 1113944N

i1εi (7)

Subject to

wrarr

middot xirarr

+ bge 1minus εi if wrarr

middot xirarrgt b

wrarr

middot xirarr

+ bleminus1 + εi if wrarr

middot xirarrlt b

εi gt 0

(8)

e performance of SVM is influenced by two mainparameters the first one is C which is a tunable parameter inequation (7) e other one is c which is used in the radialbasis function (RBF) kernel to map data into a higher di-mensional space before training and classification e RBFkernel can be defined as

K xi xj1113872 1113873 eminusc ximinusxj

111386811138681113868111386811138681113868111386811138682

(9)

where c denotes the width of the Gaussian envelope in ahigh-dimensional feature space

44 Teaching-Learning-Based Optimization (TLBO)TLBO an evolution-based optimization algorithm wasproposed by Rao et al [41] in 2011 e concept of TLBO isinspired by the evolution of the learning process when agroup or a class of learners learn a target task ere are two

ways of learning in groups or classes (1) learning from theguidance of the teacher and (2) learning from other learnerse procedure of TLBO can be divided into two phases asdescribed below in Sections 441 and 442

441 Teacher Phase In the whole population the teacher(Xteacher) can be considered as the best solution Namelylearners learn from the teacher in the teacher phase In thisphase the teacher strives for enhancing the results of otherindividuals (Xi) by increasing the mean result of theclassroom is can be described as adjusting Xmean toapproximate Xteacher In order to maintain a stochasticnature during the optimization process two randomlygenerated parameters r and TF are applied in each iterationfor the solution Xi as

Xnew Xi + ri Xteacher minusTFXmean( 1113857 (10)

TF round[1 + rand(0 1)] (11)

In equation (10) ri is a randomly selected number in therange of 0 and 1 Moreover Xnew and Xi are the new andexisting solutions at iteration i respectively TF in equation(11) is a teaching factor which can be either 1 or 2

a = 1 b = 1

x = 1 2 hellip y = 1 2 hellip

x = 1 3 hellip y = 1 2 hellip

x = 1 2 hellip y = 1 3 hellip

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 1] times [1 1]

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 2] times [1 1]

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 1] times [1 2]

Figure 6 Different pooling types when both a and b equal to onee left side shows the equations for pooling condition and computationof pooling region Pictures on the right side indicate regions before and after pooling

w middot x ndash

b = 1

w middot x ndash

b = 0

w middot x ndash

b = ndash1

Figure 7 Example of decision boundary hyperplane with twoclasses of samples

8 Mobile Information Systems

442 Learner Phase e learners gain their knowledge byinteracting with each other erefore an individual learnsnew information if other individuals have more knowledgethan him or her In this phase the student Xi interactsrandomly with another student Xj (ine j) in order to en-hance his or her knowledge Equation (12) shows that if Xi isbetter than Xj (ie f(Xj)gtf(Xi) for minimizationproblems) Xj is moved toward Xi Otherwise it is movedaway from Xi

Xnew Xi + ri Xi minusXj1113872 1113873 f Xj1113872 1113873gtf Xi( 1113857

Xi + ri Xj minusXi1113872 1113873 f Xi( 1113857gtf Xj1113872 1113873

⎧⎪⎨

⎪⎩(12)

If the new solution Xnew to the problem is better than theold ones the new solution Xnew will be recorded as the bestsolution After updating the status of each learner a newiteration begins A stop criterion based on the iterationnumber or the difference of the cost function can be set tostop the iteration properly e flowchart of TLBO is shownin Figure 8

5 Result

Our fundus image data is from the database provided by oneof the Kaggle contests entitled ldquoIdentify signs of diabeticretinopathy in eye imagesrdquo [42] In this database there areabout 90000 images We separate 1000 images from thetraining dataset to be the validation dataset e detailedinformation of each dataset is shown in Table 3 and our twonetwork architectures are shown in Figure 9

Our proposed method uses two DCNNs with fractionalmax-pooling layers For every input fundus image the twoDCNN will output a vector of size 1 times 5 representing theprobability distribution of the prediction for each lesion(category) e probability distribution together with othervalues forms a feature with dimensionality 24 e 24features are described as follows

(i) DCNN probabilities of each lesion respectively (5features)

(ii) Averages of R G and B channel values within 50 lowast50 center cropped image (3 features)

(iii) Widths and heights of 50 lowast 50 center croppedimage (2 features)

(iv) Overall standard deviation of the original image and50 lowast 50 centered cropped image Laplacian-filtered image (2 features)

(v) In total there are 12 features for one fundusimage We then append another 12 features fromthe fundus image of the other eye of the samesubject erefore the overall length of the featurevector is 24 for one fundus image e 24 featurevectors of dimensionality are used as input vec-tors of SVM

e 24-dimensional vector is used to train a multiclassSVM (five classes) whose parameters are optimized usingthe TLBO method We implemented the method described

in [39] and used it as the baseline e baseline system usessimilar features with a scheme of ensemble classifier (RF)

We used the validation set data to optimize the pa-rameter set (C c) in SVM using TLBO e upper andlower bounds of the parameter are set within [0 100] Weran 50 iterations with 50 students

Our final accuracy for five-class classification task ofDR is 8617 and the accuracy for the binary classclassification task is 9105 Labels for five-class classi-fication are normal NPDR level 1 NPDR level 2 NPDRlevel 3 and PDR while labels for binary class classifi-cation are normal and abnormal For binary classifica-tion its sensitivity is 08930 while the specificity is09089 Except counting accuracy we also do a T-test forour binary class classification e T-test is also called theStudentrsquos t-test It is a statistical hypothesis test in whichthe test statistic follows a Studentrsquos t-distributionUsually the t-test is used to compare whether there is asignificant difference between two groups of data andassists in judging the data divergence In doing a pairedsamples t-test with results from binary class classificationand random judgment its outcome is 1 for the hypothesistest result zero for the p value and [03934 04033] forthe confidence interval under null hypothesis at the 5significant level

e hypothesis test result is an index that tells whethertwo data come from the same distribution or not If the datacome from the same distribution the value of the hypothesistest result will be close to 0 On the contrary if the dataresources are distinct the result will be close to 1 whichmeans there is a differentiation between the datae p valueis the probability of accepting the assumption that there is adifference between two data may be wronge smaller the p

value the more reason that there is a disparity between dataAlso we designed an app called ldquoDeep Retinardquo pro-

viding personal examination remote medical care and earlyscreening Figure 10 shows our app interfaces Afterchoosing a fundus image that the user wants to check it willsend the image to our server and use our designed machinelearning method It takes about 10 s (depends on networkspeed) to get the result which will be presented as theprobability of each lesion With a handheld device in-dividuals can do the initial examination at the district officeor even at home More importantly it can benefit someremote areas that lack medical resources

6 Discussion

61 Accuracy Improvement Table 4 shows the accuracycomparison when using different classifiers and pa-rameter optimization methods on each dataset Using thedefault parameters with SVM (without optimization)accuracies in both validation and test sets are higher thanthat of the RF [39] If we optimize the parameters usingthe default parameter searching method provided in theLIBSVM software package though it achieves very highaccuracy in the five-fold cross validation experiment the

Mobile Information Systems 9

validation and test accuracies are even lower than thedefault one From this result we believe that overttingarises when optimizing parameters in SVM

Table 5 shows the confusion matrix of the classicationresults from the two DCNN networks (before performingSVM classication) Network 1 is the architecture shown on

Yes

No

Reject

Yes

No

No

Yes

Yes

Start

Initial students terminal criterion

i = 1

Calculate mean of each design variables

Identify best teachersolution

Based on best teachersolution to modifyXnewi = Xoldi + differencei = Xoldi +

ri (Xteacher ndash (TF)mean)

Randomly select any solution Xj (Xj ne Xi)

i = i + 1

Final value of solution and end

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is Xi better than Xjf (Xi) gt f (Xj)

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is termination criteria satisfiedf (Xnewi) gt f (Xoldi)

No

Yes

No

Xnewi = Xoldi + ri (Xi ndash Xj)

Xi = Xnewi

Xnewi = Xoldi + ri (Xj ndash Xi)

i le N

Xi = Xnewi

Figure 8 TLBO owchart

10 Mobile Information Systems

the left side of Figure 9 and network 2 is the one on the rightside From Table 5 it shows that the lesion classications of 0and 2 are better than the other categories For lesion 1 mostof the prediction results are incorrect Also for lesions 3 and4 the majority of the results are misclassications that areclassied into lesion 2

Table 6 shows the confusion matrix of the classicationresults using the full procedure of the proposed method(using SVM with TLBO) Table 7 displays the dierencebetween Tables 5 and 6 which serves as a performancecomparison between the two methods (using DCNN onlyand DCNN + SVM + TLBO) From the table every classexcept class 0 and overall accuracy is increased in network 1For network 2 each accuracy except class 3 is increasede decline in accuracy of class 3 is mainly caused by

misclassication of class 2 Table 8 shows the confusionmatrix of the classication results using the baseline methodas reported in [39] for comparison purposes

62 Deep Learning vs Traditional Classication MethodsMany traditional classication methods try to solve theproblem of DR detection by (1) using image processing tocapture symptoms in fundus images and then (2) building aclassier to make decisions based on the detected symptoms(1) e shortcoming of image processing methods is that themanifestations of the symptoms are random across dierentimages therefore it is extremely time-consuming and re-quires intense eorts to label the locations of the symptomsAbiding by the new philosophy that comes with the emer-gence of the deep learning technology our proposed methodis trying to learn how to make decisions directly from theimage data itself Dierent than the former approaches ourimages only need to be labeled with lesion number instead oflabeling symptom locations Consequently it saves consid-erable time during the database preprocessing stage On top ofthe classication results by the two DCNN networks we useSVM optimized by TLBO to generate an improved outcomeand we achieve 8617 accuracy Our result is better than therst-place winner in the Kaggle competition It shows that ourresearch result is the state-of-the-art

63 Limitation In our current datasets the number ofimages of lesions 3 and 4 is not suiexclcient to train a networkwhich is a limitation of the proposed methoderefore oneof our future works is to develop deeper collaborative re-lations with hospitals and clinics to acquire more data oflesions 3 and 4 With more data we believe the classicationaccuracy will be further increased In addition from ourresult we found that it is hard to dierentiate the imagesbetween lesions 0 and 1 erefore when we collect newdata it is desirable to collect more images belonging tolesions 0 and 1 Also we can attempt to use a dierentnetwork architecture for this problem

7 Conclusion

It is feasible to train a deep learning model for automaticdiagnosis of DR as long as we have enough data for sta-tistical model training Furthermore the database prepa-ration stage only needs a categorical label for each trainingimage It does not require detailed annotation for retinalvessel tracking in every image Hence it is time-eiexclcientcompared to the traditional machine learning-based methodfor automatic diagnosis of DR e nal accuracy canachieve 8617 and 9105 for ve-class and binary classclassications respectively

e sensitivity and specicity of binary classication are08930 and 09089 respectively which is a satisfactory resultFurthermore we developed an automatic inspection appthat can be used in both personal examination and remotemedical care With more image data collected we expect theaccuracy can be even more enhanced further improving oursystem

Table 3 Detailed information of each dataset

Dataset of imagesDR lesions

0 1 2 3 4Train 34124 73 6 15 3 3Validation 1000 70 8 15 4 3Test 53572 74 7 15 2 2

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

3 times 3 conv 320

3 times 3 conv 352

3 times 3 conv 384

2 times 2 conv 416

1 times 1 conv 448

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

Softmax

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

2 times 2 conv 320

1 times 1 conv 356

Softmax

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

Figure 9 Architecture of the two DCNN networks that we used

Mobile Information Systems 11

Table 4 Accuracy comparison when using different classifiers and different parameter optimization methods on each dataset

DatasetMethod

SVM RF [39]Without optimization Default parameter searching TLBO

Five-fold cross validation (using training data) 88744 965724 mdash mdashValidation data 854 796 mdash 847Test data 861177 810573 8617 8602

Table 5 Confusion matrix of the classification results from the two DCNN networks before performing SVM classification

Determined levelGround-truth level

Network 1 Network 20 1 2 3 4 0 1 2 3 4

0 39031 2693 2361 75 160 38310 2233 1617 51 781 118 438 185 0 1 327 735 290 1 32 339 626 5058 738 392 790 786 5509 620 3453 0 0 170 332 102 4 1 338 486 1854 43 5 85 69 551 100 7 105 56 595Accuracy () 9874 1164 6436 2735 4569 9691 1954 7001 4003 4934Overall accuracy () 8476 8518

(a) (b) (c) (d)

Figure 10 App interfaces

Table 6 Confusion matrix of the classification results using the full procedure of the proposed method (using SVM with TLBO)

Determined levelGround-truth level

SVM optimized with TLBO0 1 2 3 4

0 38611 2233 1513 44 801 312 831 349 1 12 570 696 5706 701 3933 3 0 215 408 1234 35 2 76 60 609Accuracy () 9767 2209 7260 3361 5050Overall accuracy () 8617

12 Mobile Information Systems

Data Availability

All experiment data come from a Kaggle contest calledldquoIdentify signs of diabetic retinopathy in eye imagesrdquo(website httpswwwkagglecomcdiabetic-retinopathy-detection)

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e authors would like to thank Mr Chia-Ming Hu for hishelp in preparing computational instruments and per-forming parts of the experiments in this paper

References

[1] N H Cho J E Shaw S Karuranga et al ldquoIDF Diabetes Atlasglobal estimates of diabetes prevalence for 2017 and pro-jections for 2045rdquo Diabetes Research and Clinical Practicevol 138 pp 271ndash281 2018

[2] American Diabetes Association Eye Complications AmericanDiabetes Association Arlington VA USA 2013 httpwwwdiabetesorgliving-with-diabetescomplicationseye-complicationsreferrerhttpswwwgooglecomtw

[3] International Diabetes Foundation (IDF) Diabetes AtlasInternational Diabetes Federation Brussels Belgium 7thedition 2015

[4] C P Wilkinson F L Ferris R E Klein et al ldquoProposedinternational clinical diabetic retinopathy and diabetic

macular edema disease severity scalesrdquo Ophthalmologyvol 110 no 9 pp 1677ndash1682 2003

[5] A Tufail C Rudisill C Egan et al ldquoAutomated diabeticretinopathy image assessment software diagnostic accuracyand cost-effectiveness compared with human gradersrdquoOphthalmology vol 124 no 3 pp 343ndash351 2017

[6] A Tufail V V Kapetanakis S Salas-Vega et al ldquoAn ob-servational study to assess if automated diabetic retinopathyimage assessment software can replace one or more steps ofmanual imaging grading and to determine their cost-effec-tivenessrdquo Health Technology Assessment vol 20 no 92pp 1ndash72 2016

[7] I Liu and Y Sun ldquoRecursive tracking of vascular networks inangiograms based on the detection-deletion schemerdquo IEEETransactions on Medical Imaging vol 12 no 2 pp 334ndash3411993

[8] A Can H Shen J N Turner H L Tanenbaum andB Roysam ldquoRapid automated tracing and feature extractionfrom retinal fundus images using direct exploratory algo-rithmsrdquo IEEE Transactions on Information Technology inBiomedicine vol 3 no 2 pp 125ndash138 1999

[9] M Vlachos and E Dermatas ldquoMulti-scale retinal vesselsegmentation using line trackingrdquo Computerized MedicalImaging and Graphics vol 34 no 3 pp 213ndash227 2010

[10] Y Yin M Adel and S Bourennane ldquoAn automatic trackingmethod for retinal vascular tree extractionrdquo in Proceedings ofthe IEEE International Conference on Acoustics Speech andSignal Processing IEEE Kyoto Japan March 2012

[11] S Chaudhuri S Chatterjee N Katz M Nelson andM Goldbaum ldquoDetection of blood vessels in retinal imagesusing two-dimensional matched filtersrdquo IEEE Transactions onMedical Imaging vol 8 no 3 pp 263ndash269 1989

Table 8 Confusion matrix of the classification using the baseline method [39] and its overall accuracy

DeterminedOriginal

Random forest0 1 2 3 4

0 38228 2001 1331 38 641 546 1026 434 3 22 710 732 5726 635 3583 2 1 279 465 1464 45 2 89 73 636Accuracy () 9670 2727 7286 3830 5274Overall accuracy () 8601

Table 7 Differences between Tables 5 and 6

Determined lesionGround-truth lesion

SVM-network 1 SVM-network 20 1 2 3 4 0 1 2 3 4

0 +420 minus460 minus848 minus31 minus80 +310 0 minus104 minus43 +21 +194 +393 +164 +1 0 minus15 +96 +59 0 minus22 +231 +70 +648 minus37 +1 minus220 minus90 +197 +81 +483 +3 0 +45 +76 +21 +1 minus1 minus123 minus78 minus624 +3 minus3 +45 +76 +58 minus65 minus5 minus28 +4 +14Accuracy increment () minus108 8977 1280 2288 1052 078 1305 369 minus1603 23Overall accuracy increment () 16 116

Mobile Information Systems 13

4 Material and Methods

We propose an automatic DR detection algorithm based onDCNN fractional max-pooling [39] SVM [40] andteaching-learning-based optimization (TLBO) [41] Specif-ically we train two DCNN networks with fractional max-pooling combining their prediction results using SVM andoptimizing the SVM parameters with TLBO e reason fortraining two distinct networks is that different networkarchitectures may have their unique advantages in featurespace representation By training two DCNNs and com-bining their features the prediction accuracy can be furtherenhanced Another important factor impacting the recog-nition rate is the parameter of classifiers We propose tooptimize the SVM parameters using TLBOWe illustrate theimage preprocessing methods in Section 41 and present thefractional max-pooling SVM and TLBO in Sections 4243 and 44 respectively

41 Preprocessing Given the vessels in the original fundusimages are mostly not very clear and the size of eachfundus image may differ it is essential to preprocess imagesso that they have the same size and the visibility of thevessels is improved ere are three steps in preprocessinge first is to rescale images to the same size Since thefundus images are circular we rescale the input images sothat the diameter of the fundus images becomes 540 pixelsAfter rescaling the local average color value is subtractedfrom the rescaled images and another transformation isperformed so that the local average is mapped to 50 gray-scale in order to remove the color divergence caused bydifferent ophthalmoscopes Last but not least becauseboundary effects may occur in some images we remove the

periphery by clipping 10 from the border of the imagesFigure 4 shows the original fundus image and the imageafter preprocessing

42 Fractional Max-Pooling Pooling is a procedure thatturns the input matrix Min times Nin into a smaller Mout times

Nout output matrix e purpose is to divide the inputmatrix into multiple pooling regions (Pij)

pi sub 1 2 3 Min1113864 1113865 for each i isin 1 Mout1113864 1113865

pj sub 1 2 3 Nin1113864 1113865 for each j isin 1 Nout1113864 1113865

Pij pi times pj

(1)

e pooling results are computed according to poolingtype

Outputij Oper

(k l) isin Pij

Inputkl (2)

In equation (2) ldquoOperrdquo refers to a particular mathe-matical operation For example if max-pooling is used theoperation will be to take the maximum of the input regionFor average pooling the average of the input region is takenFor such a network that requires tremendous learning it ispreferable to use as many hidden layers as possible In thiswork the pooling layer used in our networks is fractionalmax-pooling instead of general max-pooling

Fractional max-pooling is a pooling scheme thatmakes the size of the output matrix equivalent to fractionaltimes that of the input matrix after pooling ieNin αNout αge 0 α notin Z To describe the general

Table 2 Performance comparison

Method Year Reference Sn Sp Acc AUC Database

Vascular tracking 2010 [9] 07468 09551 09285 mdash DRIVE2012 [10] 06887 09562 09290 mdash STARE

Matched filtering2009 [14] 06611 09848 09497 mdash STARE

2012 [15] 07191 09687 09407 mdash STARE07154 09716 09343 mdash DRIVE

Morphological processing 2011 [18] 07352 09795 09458 mdash DRIVE2014 [19] 07862 09815 09598 mdash DRIVE

Deformation model

2007 [21] 06634 09682 09316 mdash DRIVE

2009 [23] mdash 09736 09087 mdash STAREmdash 09772 09610 mdash DRIVE

2014 [24] 07187 09767 09509 mdash STARE07354 09789 09477 mdash DRIVE

Machine learning

2007 [25] mdash mdash 09584 09602 STAREmdash mdash 09563 09558 DRIVE

2011 [26] 06944 09819 09526 mdash STARE07067 09801 09452 mdash DRIVE

2013 [27] 08194 09679 09725 mdash DRIVE

2014 [28] 08104 09791 09813 09751 STARE08173 09733 09767 09475 DRIVE

2014 [29] 07887 09633 09441 mdash STARE07512 09684 09412 mdash DRIVE

Sn sensitivity Sp specificity Acc accuracy AUC area under curve

6 Mobile Information Systems

pooling regions let (xi)Mouti0 and (yj)

Noutj0 be two increasing

integer sequences starting with one and ending with Min orNin ese two sequences are used in pooling steps asdescribed in Figures 5 and 6

Pij ximinus1 xi minus a1113858 1113859 times yjminus1 yj minus b1113960 1113961 (3)

e constants a and b in equation (3) stand for theoverlapping length and the width of the pooling windowrespectively Figure 5 is a simple example of overlappingpooling Figure 6 illustrates different pooling region types

After fractional max-pooling the pooling window size isstill integers but the global pooling size will change Namelyfractional max-pooling does not directly change the poolingwindow into a fractional scale Instead it uses windows ofvariable size to achieve fractional pooling e generation of x

and y sequences can be random or pseudorandom Pseudo-random sequences generate more stable pooling regions thanrandom sequences and can also achieve higher accuracy [39]

43 Support Vector Machine (SVM) SVM is a supervisedlearning method used for classification and regressionanalysis SVM can find the hyperplane or decision boundarydefined by the solution vector w which not only separatesthe training vectors but also works well with unseen testdata To improve its generalization ability SVM selectsdecision boundaries based on maximizing margins betweenclasses

Figure 7 illustrates the idea Suppose there are n points ina binary dataset

x1rarr

y1( 1113857 xnrarr

yn( 1113857 (4)

where yi is the data label which can be 1 or minus1 indicatingthe class to which xi

rarr belongs We need to find the optimizedhyperplane such that the distance between the hyperplaneand its nearest point xi

rarr is maximized A hyperplane can bewritten as equation (5) based on x

rarr

wrarr

middot xrarrminus b 0 (5)

where wrarr is the normal vector of the hyperplane and the

value of bwrarr

decides the margin of hyperplane from thetraining data point along the normal vector w

rarr

(a) (b)

(c) (d)

Figure 4 Image before and after the preprocessing stage (a b) images before preprocessing stage (c d) images after preprocessing stage

x = 1 2 3 4 5 y = 1 2 4 5 a = 0 b = 1P11 = [x1minus1 x1 minus 0] times [y1minus1 y1 minus 1] = [1 2] times [1 2]P12 = [x1minus1 x1 minus 1] times [y2minus1 y2 minus 1] = [1 2] times [2 3]

Figure 5 An overlapping example e blue solid line indicatespooling region P11 while the red dotted line shows the P12

Mobile Information Systems 7

For yi whose value is 1 the data must satisfywrarr

middot xrarrminus bge 1 and for yi whose value is minus1 w

rarrmiddot xrarrminus bleminus1

has to be satisfied Combining these two conditions we get

yi(wrarr

middot xrarrminus b)ge 1 (6)

e goal is to maximize bwrarr

according to the constrainof equation (6) in order to derive the optimized decisionhyperplane for classification

Sometimes the training data might not be able to beperfectly separated using linear boundarieserefore in theSVM formulation we need to introduce the error metric εand the cost parameter C as shown in equation (7) e goalnow becomes to minimize

12

wrarr

middot wrarr

+ C 1113944N

i1εi (7)

Subject to

wrarr

middot xirarr

+ bge 1minus εi if wrarr

middot xirarrgt b

wrarr

middot xirarr

+ bleminus1 + εi if wrarr

middot xirarrlt b

εi gt 0

(8)

e performance of SVM is influenced by two mainparameters the first one is C which is a tunable parameter inequation (7) e other one is c which is used in the radialbasis function (RBF) kernel to map data into a higher di-mensional space before training and classification e RBFkernel can be defined as

K xi xj1113872 1113873 eminusc ximinusxj

111386811138681113868111386811138681113868111386811138682

(9)

where c denotes the width of the Gaussian envelope in ahigh-dimensional feature space

44 Teaching-Learning-Based Optimization (TLBO)TLBO an evolution-based optimization algorithm wasproposed by Rao et al [41] in 2011 e concept of TLBO isinspired by the evolution of the learning process when agroup or a class of learners learn a target task ere are two

ways of learning in groups or classes (1) learning from theguidance of the teacher and (2) learning from other learnerse procedure of TLBO can be divided into two phases asdescribed below in Sections 441 and 442

441 Teacher Phase In the whole population the teacher(Xteacher) can be considered as the best solution Namelylearners learn from the teacher in the teacher phase In thisphase the teacher strives for enhancing the results of otherindividuals (Xi) by increasing the mean result of theclassroom is can be described as adjusting Xmean toapproximate Xteacher In order to maintain a stochasticnature during the optimization process two randomlygenerated parameters r and TF are applied in each iterationfor the solution Xi as

Xnew Xi + ri Xteacher minusTFXmean( 1113857 (10)

TF round[1 + rand(0 1)] (11)

In equation (10) ri is a randomly selected number in therange of 0 and 1 Moreover Xnew and Xi are the new andexisting solutions at iteration i respectively TF in equation(11) is a teaching factor which can be either 1 or 2

a = 1 b = 1

x = 1 2 hellip y = 1 2 hellip

x = 1 3 hellip y = 1 2 hellip

x = 1 2 hellip y = 1 3 hellip

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 1] times [1 1]

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 2] times [1 1]

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 1] times [1 2]

Figure 6 Different pooling types when both a and b equal to onee left side shows the equations for pooling condition and computationof pooling region Pictures on the right side indicate regions before and after pooling

w middot x ndash

b = 1

w middot x ndash

b = 0

w middot x ndash

b = ndash1

Figure 7 Example of decision boundary hyperplane with twoclasses of samples

8 Mobile Information Systems

442 Learner Phase e learners gain their knowledge byinteracting with each other erefore an individual learnsnew information if other individuals have more knowledgethan him or her In this phase the student Xi interactsrandomly with another student Xj (ine j) in order to en-hance his or her knowledge Equation (12) shows that if Xi isbetter than Xj (ie f(Xj)gtf(Xi) for minimizationproblems) Xj is moved toward Xi Otherwise it is movedaway from Xi

Xnew Xi + ri Xi minusXj1113872 1113873 f Xj1113872 1113873gtf Xi( 1113857

Xi + ri Xj minusXi1113872 1113873 f Xi( 1113857gtf Xj1113872 1113873

⎧⎪⎨

⎪⎩(12)

If the new solution Xnew to the problem is better than theold ones the new solution Xnew will be recorded as the bestsolution After updating the status of each learner a newiteration begins A stop criterion based on the iterationnumber or the difference of the cost function can be set tostop the iteration properly e flowchart of TLBO is shownin Figure 8

5 Result

Our fundus image data is from the database provided by oneof the Kaggle contests entitled ldquoIdentify signs of diabeticretinopathy in eye imagesrdquo [42] In this database there areabout 90000 images We separate 1000 images from thetraining dataset to be the validation dataset e detailedinformation of each dataset is shown in Table 3 and our twonetwork architectures are shown in Figure 9

Our proposed method uses two DCNNs with fractionalmax-pooling layers For every input fundus image the twoDCNN will output a vector of size 1 times 5 representing theprobability distribution of the prediction for each lesion(category) e probability distribution together with othervalues forms a feature with dimensionality 24 e 24features are described as follows

(i) DCNN probabilities of each lesion respectively (5features)

(ii) Averages of R G and B channel values within 50 lowast50 center cropped image (3 features)

(iii) Widths and heights of 50 lowast 50 center croppedimage (2 features)

(iv) Overall standard deviation of the original image and50 lowast 50 centered cropped image Laplacian-filtered image (2 features)

(v) In total there are 12 features for one fundusimage We then append another 12 features fromthe fundus image of the other eye of the samesubject erefore the overall length of the featurevector is 24 for one fundus image e 24 featurevectors of dimensionality are used as input vec-tors of SVM

e 24-dimensional vector is used to train a multiclassSVM (five classes) whose parameters are optimized usingthe TLBO method We implemented the method described

in [39] and used it as the baseline e baseline system usessimilar features with a scheme of ensemble classifier (RF)

We used the validation set data to optimize the pa-rameter set (C c) in SVM using TLBO e upper andlower bounds of the parameter are set within [0 100] Weran 50 iterations with 50 students

Our final accuracy for five-class classification task ofDR is 8617 and the accuracy for the binary classclassification task is 9105 Labels for five-class classi-fication are normal NPDR level 1 NPDR level 2 NPDRlevel 3 and PDR while labels for binary class classifi-cation are normal and abnormal For binary classifica-tion its sensitivity is 08930 while the specificity is09089 Except counting accuracy we also do a T-test forour binary class classification e T-test is also called theStudentrsquos t-test It is a statistical hypothesis test in whichthe test statistic follows a Studentrsquos t-distributionUsually the t-test is used to compare whether there is asignificant difference between two groups of data andassists in judging the data divergence In doing a pairedsamples t-test with results from binary class classificationand random judgment its outcome is 1 for the hypothesistest result zero for the p value and [03934 04033] forthe confidence interval under null hypothesis at the 5significant level

e hypothesis test result is an index that tells whethertwo data come from the same distribution or not If the datacome from the same distribution the value of the hypothesistest result will be close to 0 On the contrary if the dataresources are distinct the result will be close to 1 whichmeans there is a differentiation between the datae p valueis the probability of accepting the assumption that there is adifference between two data may be wronge smaller the p

value the more reason that there is a disparity between dataAlso we designed an app called ldquoDeep Retinardquo pro-

viding personal examination remote medical care and earlyscreening Figure 10 shows our app interfaces Afterchoosing a fundus image that the user wants to check it willsend the image to our server and use our designed machinelearning method It takes about 10 s (depends on networkspeed) to get the result which will be presented as theprobability of each lesion With a handheld device in-dividuals can do the initial examination at the district officeor even at home More importantly it can benefit someremote areas that lack medical resources

6 Discussion

61 Accuracy Improvement Table 4 shows the accuracycomparison when using different classifiers and pa-rameter optimization methods on each dataset Using thedefault parameters with SVM (without optimization)accuracies in both validation and test sets are higher thanthat of the RF [39] If we optimize the parameters usingthe default parameter searching method provided in theLIBSVM software package though it achieves very highaccuracy in the five-fold cross validation experiment the

Mobile Information Systems 9

validation and test accuracies are even lower than thedefault one From this result we believe that overttingarises when optimizing parameters in SVM

Table 5 shows the confusion matrix of the classicationresults from the two DCNN networks (before performingSVM classication) Network 1 is the architecture shown on

Yes

No

Reject

Yes

No

No

Yes

Yes

Start

Initial students terminal criterion

i = 1

Calculate mean of each design variables

Identify best teachersolution

Based on best teachersolution to modifyXnewi = Xoldi + differencei = Xoldi +

ri (Xteacher ndash (TF)mean)

Randomly select any solution Xj (Xj ne Xi)

i = i + 1

Final value of solution and end

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is Xi better than Xjf (Xi) gt f (Xj)

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is termination criteria satisfiedf (Xnewi) gt f (Xoldi)

No

Yes

No

Xnewi = Xoldi + ri (Xi ndash Xj)

Xi = Xnewi

Xnewi = Xoldi + ri (Xj ndash Xi)

i le N

Xi = Xnewi

Figure 8 TLBO owchart

10 Mobile Information Systems

the left side of Figure 9 and network 2 is the one on the rightside From Table 5 it shows that the lesion classications of 0and 2 are better than the other categories For lesion 1 mostof the prediction results are incorrect Also for lesions 3 and4 the majority of the results are misclassications that areclassied into lesion 2

Table 6 shows the confusion matrix of the classicationresults using the full procedure of the proposed method(using SVM with TLBO) Table 7 displays the dierencebetween Tables 5 and 6 which serves as a performancecomparison between the two methods (using DCNN onlyand DCNN + SVM + TLBO) From the table every classexcept class 0 and overall accuracy is increased in network 1For network 2 each accuracy except class 3 is increasede decline in accuracy of class 3 is mainly caused by

misclassication of class 2 Table 8 shows the confusionmatrix of the classication results using the baseline methodas reported in [39] for comparison purposes

62 Deep Learning vs Traditional Classication MethodsMany traditional classication methods try to solve theproblem of DR detection by (1) using image processing tocapture symptoms in fundus images and then (2) building aclassier to make decisions based on the detected symptoms(1) e shortcoming of image processing methods is that themanifestations of the symptoms are random across dierentimages therefore it is extremely time-consuming and re-quires intense eorts to label the locations of the symptomsAbiding by the new philosophy that comes with the emer-gence of the deep learning technology our proposed methodis trying to learn how to make decisions directly from theimage data itself Dierent than the former approaches ourimages only need to be labeled with lesion number instead oflabeling symptom locations Consequently it saves consid-erable time during the database preprocessing stage On top ofthe classication results by the two DCNN networks we useSVM optimized by TLBO to generate an improved outcomeand we achieve 8617 accuracy Our result is better than therst-place winner in the Kaggle competition It shows that ourresearch result is the state-of-the-art

63 Limitation In our current datasets the number ofimages of lesions 3 and 4 is not suiexclcient to train a networkwhich is a limitation of the proposed methoderefore oneof our future works is to develop deeper collaborative re-lations with hospitals and clinics to acquire more data oflesions 3 and 4 With more data we believe the classicationaccuracy will be further increased In addition from ourresult we found that it is hard to dierentiate the imagesbetween lesions 0 and 1 erefore when we collect newdata it is desirable to collect more images belonging tolesions 0 and 1 Also we can attempt to use a dierentnetwork architecture for this problem

7 Conclusion

It is feasible to train a deep learning model for automaticdiagnosis of DR as long as we have enough data for sta-tistical model training Furthermore the database prepa-ration stage only needs a categorical label for each trainingimage It does not require detailed annotation for retinalvessel tracking in every image Hence it is time-eiexclcientcompared to the traditional machine learning-based methodfor automatic diagnosis of DR e nal accuracy canachieve 8617 and 9105 for ve-class and binary classclassications respectively

e sensitivity and specicity of binary classication are08930 and 09089 respectively which is a satisfactory resultFurthermore we developed an automatic inspection appthat can be used in both personal examination and remotemedical care With more image data collected we expect theaccuracy can be even more enhanced further improving oursystem

Table 3 Detailed information of each dataset

Dataset of imagesDR lesions

0 1 2 3 4Train 34124 73 6 15 3 3Validation 1000 70 8 15 4 3Test 53572 74 7 15 2 2

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

3 times 3 conv 320

3 times 3 conv 352

3 times 3 conv 384

2 times 2 conv 416

1 times 1 conv 448

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

Softmax

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

2 times 2 conv 320

1 times 1 conv 356

Softmax

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

Figure 9 Architecture of the two DCNN networks that we used

Mobile Information Systems 11

Table 4 Accuracy comparison when using different classifiers and different parameter optimization methods on each dataset

DatasetMethod

SVM RF [39]Without optimization Default parameter searching TLBO

Five-fold cross validation (using training data) 88744 965724 mdash mdashValidation data 854 796 mdash 847Test data 861177 810573 8617 8602

Table 5 Confusion matrix of the classification results from the two DCNN networks before performing SVM classification

Determined levelGround-truth level

Network 1 Network 20 1 2 3 4 0 1 2 3 4

0 39031 2693 2361 75 160 38310 2233 1617 51 781 118 438 185 0 1 327 735 290 1 32 339 626 5058 738 392 790 786 5509 620 3453 0 0 170 332 102 4 1 338 486 1854 43 5 85 69 551 100 7 105 56 595Accuracy () 9874 1164 6436 2735 4569 9691 1954 7001 4003 4934Overall accuracy () 8476 8518

(a) (b) (c) (d)

Figure 10 App interfaces

Table 6 Confusion matrix of the classification results using the full procedure of the proposed method (using SVM with TLBO)

Determined levelGround-truth level

SVM optimized with TLBO0 1 2 3 4

0 38611 2233 1513 44 801 312 831 349 1 12 570 696 5706 701 3933 3 0 215 408 1234 35 2 76 60 609Accuracy () 9767 2209 7260 3361 5050Overall accuracy () 8617

12 Mobile Information Systems

Data Availability

All experiment data come from a Kaggle contest calledldquoIdentify signs of diabetic retinopathy in eye imagesrdquo(website httpswwwkagglecomcdiabetic-retinopathy-detection)

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e authors would like to thank Mr Chia-Ming Hu for hishelp in preparing computational instruments and per-forming parts of the experiments in this paper

References

[1] N H Cho J E Shaw S Karuranga et al ldquoIDF Diabetes Atlasglobal estimates of diabetes prevalence for 2017 and pro-jections for 2045rdquo Diabetes Research and Clinical Practicevol 138 pp 271ndash281 2018

[2] American Diabetes Association Eye Complications AmericanDiabetes Association Arlington VA USA 2013 httpwwwdiabetesorgliving-with-diabetescomplicationseye-complicationsreferrerhttpswwwgooglecomtw

[3] International Diabetes Foundation (IDF) Diabetes AtlasInternational Diabetes Federation Brussels Belgium 7thedition 2015

[4] C P Wilkinson F L Ferris R E Klein et al ldquoProposedinternational clinical diabetic retinopathy and diabetic

macular edema disease severity scalesrdquo Ophthalmologyvol 110 no 9 pp 1677ndash1682 2003

[5] A Tufail C Rudisill C Egan et al ldquoAutomated diabeticretinopathy image assessment software diagnostic accuracyand cost-effectiveness compared with human gradersrdquoOphthalmology vol 124 no 3 pp 343ndash351 2017

[6] A Tufail V V Kapetanakis S Salas-Vega et al ldquoAn ob-servational study to assess if automated diabetic retinopathyimage assessment software can replace one or more steps ofmanual imaging grading and to determine their cost-effec-tivenessrdquo Health Technology Assessment vol 20 no 92pp 1ndash72 2016

[7] I Liu and Y Sun ldquoRecursive tracking of vascular networks inangiograms based on the detection-deletion schemerdquo IEEETransactions on Medical Imaging vol 12 no 2 pp 334ndash3411993

[8] A Can H Shen J N Turner H L Tanenbaum andB Roysam ldquoRapid automated tracing and feature extractionfrom retinal fundus images using direct exploratory algo-rithmsrdquo IEEE Transactions on Information Technology inBiomedicine vol 3 no 2 pp 125ndash138 1999

[9] M Vlachos and E Dermatas ldquoMulti-scale retinal vesselsegmentation using line trackingrdquo Computerized MedicalImaging and Graphics vol 34 no 3 pp 213ndash227 2010

[10] Y Yin M Adel and S Bourennane ldquoAn automatic trackingmethod for retinal vascular tree extractionrdquo in Proceedings ofthe IEEE International Conference on Acoustics Speech andSignal Processing IEEE Kyoto Japan March 2012

[11] S Chaudhuri S Chatterjee N Katz M Nelson andM Goldbaum ldquoDetection of blood vessels in retinal imagesusing two-dimensional matched filtersrdquo IEEE Transactions onMedical Imaging vol 8 no 3 pp 263ndash269 1989

Table 8 Confusion matrix of the classification using the baseline method [39] and its overall accuracy

DeterminedOriginal

Random forest0 1 2 3 4

0 38228 2001 1331 38 641 546 1026 434 3 22 710 732 5726 635 3583 2 1 279 465 1464 45 2 89 73 636Accuracy () 9670 2727 7286 3830 5274Overall accuracy () 8601

Table 7 Differences between Tables 5 and 6

Determined lesionGround-truth lesion

SVM-network 1 SVM-network 20 1 2 3 4 0 1 2 3 4

0 +420 minus460 minus848 minus31 minus80 +310 0 minus104 minus43 +21 +194 +393 +164 +1 0 minus15 +96 +59 0 minus22 +231 +70 +648 minus37 +1 minus220 minus90 +197 +81 +483 +3 0 +45 +76 +21 +1 minus1 minus123 minus78 minus624 +3 minus3 +45 +76 +58 minus65 minus5 minus28 +4 +14Accuracy increment () minus108 8977 1280 2288 1052 078 1305 369 minus1603 23Overall accuracy increment () 16 116

Mobile Information Systems 13

pooling regions let (xi)Mouti0 and (yj)

Noutj0 be two increasing

integer sequences starting with one and ending with Min orNin ese two sequences are used in pooling steps asdescribed in Figures 5 and 6

Pij ximinus1 xi minus a1113858 1113859 times yjminus1 yj minus b1113960 1113961 (3)

e constants a and b in equation (3) stand for theoverlapping length and the width of the pooling windowrespectively Figure 5 is a simple example of overlappingpooling Figure 6 illustrates different pooling region types

After fractional max-pooling the pooling window size isstill integers but the global pooling size will change Namelyfractional max-pooling does not directly change the poolingwindow into a fractional scale Instead it uses windows ofvariable size to achieve fractional pooling e generation of x

and y sequences can be random or pseudorandom Pseudo-random sequences generate more stable pooling regions thanrandom sequences and can also achieve higher accuracy [39]

43 Support Vector Machine (SVM) SVM is a supervisedlearning method used for classification and regressionanalysis SVM can find the hyperplane or decision boundarydefined by the solution vector w which not only separatesthe training vectors but also works well with unseen testdata To improve its generalization ability SVM selectsdecision boundaries based on maximizing margins betweenclasses

Figure 7 illustrates the idea Suppose there are n points ina binary dataset

x1rarr

y1( 1113857 xnrarr

yn( 1113857 (4)

where yi is the data label which can be 1 or minus1 indicatingthe class to which xi

rarr belongs We need to find the optimizedhyperplane such that the distance between the hyperplaneand its nearest point xi

rarr is maximized A hyperplane can bewritten as equation (5) based on x

rarr

wrarr

middot xrarrminus b 0 (5)

where wrarr is the normal vector of the hyperplane and the

value of bwrarr

decides the margin of hyperplane from thetraining data point along the normal vector w

rarr

(a) (b)

(c) (d)

Figure 4 Image before and after the preprocessing stage (a b) images before preprocessing stage (c d) images after preprocessing stage

x = 1 2 3 4 5 y = 1 2 4 5 a = 0 b = 1P11 = [x1minus1 x1 minus 0] times [y1minus1 y1 minus 1] = [1 2] times [1 2]P12 = [x1minus1 x1 minus 1] times [y2minus1 y2 minus 1] = [1 2] times [2 3]

Figure 5 An overlapping example e blue solid line indicatespooling region P11 while the red dotted line shows the P12

Mobile Information Systems 7

For yi whose value is 1 the data must satisfywrarr

middot xrarrminus bge 1 and for yi whose value is minus1 w

rarrmiddot xrarrminus bleminus1

has to be satisfied Combining these two conditions we get

yi(wrarr

middot xrarrminus b)ge 1 (6)

e goal is to maximize bwrarr

according to the constrainof equation (6) in order to derive the optimized decisionhyperplane for classification

Sometimes the training data might not be able to beperfectly separated using linear boundarieserefore in theSVM formulation we need to introduce the error metric εand the cost parameter C as shown in equation (7) e goalnow becomes to minimize

12

wrarr

middot wrarr

+ C 1113944N

i1εi (7)

Subject to

wrarr

middot xirarr

+ bge 1minus εi if wrarr

middot xirarrgt b

wrarr

middot xirarr

+ bleminus1 + εi if wrarr

middot xirarrlt b

εi gt 0

(8)

e performance of SVM is influenced by two mainparameters the first one is C which is a tunable parameter inequation (7) e other one is c which is used in the radialbasis function (RBF) kernel to map data into a higher di-mensional space before training and classification e RBFkernel can be defined as

K xi xj1113872 1113873 eminusc ximinusxj

111386811138681113868111386811138681113868111386811138682

(9)

where c denotes the width of the Gaussian envelope in ahigh-dimensional feature space

44 Teaching-Learning-Based Optimization (TLBO)TLBO an evolution-based optimization algorithm wasproposed by Rao et al [41] in 2011 e concept of TLBO isinspired by the evolution of the learning process when agroup or a class of learners learn a target task ere are two

ways of learning in groups or classes (1) learning from theguidance of the teacher and (2) learning from other learnerse procedure of TLBO can be divided into two phases asdescribed below in Sections 441 and 442

441 Teacher Phase In the whole population the teacher(Xteacher) can be considered as the best solution Namelylearners learn from the teacher in the teacher phase In thisphase the teacher strives for enhancing the results of otherindividuals (Xi) by increasing the mean result of theclassroom is can be described as adjusting Xmean toapproximate Xteacher In order to maintain a stochasticnature during the optimization process two randomlygenerated parameters r and TF are applied in each iterationfor the solution Xi as

Xnew Xi + ri Xteacher minusTFXmean( 1113857 (10)

TF round[1 + rand(0 1)] (11)

In equation (10) ri is a randomly selected number in therange of 0 and 1 Moreover Xnew and Xi are the new andexisting solutions at iteration i respectively TF in equation(11) is a teaching factor which can be either 1 or 2

a = 1 b = 1

x = 1 2 hellip y = 1 2 hellip

x = 1 3 hellip y = 1 2 hellip

x = 1 2 hellip y = 1 3 hellip

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 1] times [1 1]

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 2] times [1 1]

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 1] times [1 2]

Figure 6 Different pooling types when both a and b equal to onee left side shows the equations for pooling condition and computationof pooling region Pictures on the right side indicate regions before and after pooling

w middot x ndash

b = 1

w middot x ndash

b = 0

w middot x ndash

b = ndash1

Figure 7 Example of decision boundary hyperplane with twoclasses of samples

8 Mobile Information Systems

442 Learner Phase e learners gain their knowledge byinteracting with each other erefore an individual learnsnew information if other individuals have more knowledgethan him or her In this phase the student Xi interactsrandomly with another student Xj (ine j) in order to en-hance his or her knowledge Equation (12) shows that if Xi isbetter than Xj (ie f(Xj)gtf(Xi) for minimizationproblems) Xj is moved toward Xi Otherwise it is movedaway from Xi

Xnew Xi + ri Xi minusXj1113872 1113873 f Xj1113872 1113873gtf Xi( 1113857

Xi + ri Xj minusXi1113872 1113873 f Xi( 1113857gtf Xj1113872 1113873

⎧⎪⎨

⎪⎩(12)

If the new solution Xnew to the problem is better than theold ones the new solution Xnew will be recorded as the bestsolution After updating the status of each learner a newiteration begins A stop criterion based on the iterationnumber or the difference of the cost function can be set tostop the iteration properly e flowchart of TLBO is shownin Figure 8

5 Result

Our fundus image data is from the database provided by oneof the Kaggle contests entitled ldquoIdentify signs of diabeticretinopathy in eye imagesrdquo [42] In this database there areabout 90000 images We separate 1000 images from thetraining dataset to be the validation dataset e detailedinformation of each dataset is shown in Table 3 and our twonetwork architectures are shown in Figure 9

Our proposed method uses two DCNNs with fractionalmax-pooling layers For every input fundus image the twoDCNN will output a vector of size 1 times 5 representing theprobability distribution of the prediction for each lesion(category) e probability distribution together with othervalues forms a feature with dimensionality 24 e 24features are described as follows

(i) DCNN probabilities of each lesion respectively (5features)

(ii) Averages of R G and B channel values within 50 lowast50 center cropped image (3 features)

(iii) Widths and heights of 50 lowast 50 center croppedimage (2 features)

(iv) Overall standard deviation of the original image and50 lowast 50 centered cropped image Laplacian-filtered image (2 features)

(v) In total there are 12 features for one fundusimage We then append another 12 features fromthe fundus image of the other eye of the samesubject erefore the overall length of the featurevector is 24 for one fundus image e 24 featurevectors of dimensionality are used as input vec-tors of SVM

e 24-dimensional vector is used to train a multiclassSVM (five classes) whose parameters are optimized usingthe TLBO method We implemented the method described

in [39] and used it as the baseline e baseline system usessimilar features with a scheme of ensemble classifier (RF)

We used the validation set data to optimize the pa-rameter set (C c) in SVM using TLBO e upper andlower bounds of the parameter are set within [0 100] Weran 50 iterations with 50 students

Our final accuracy for five-class classification task ofDR is 8617 and the accuracy for the binary classclassification task is 9105 Labels for five-class classi-fication are normal NPDR level 1 NPDR level 2 NPDRlevel 3 and PDR while labels for binary class classifi-cation are normal and abnormal For binary classifica-tion its sensitivity is 08930 while the specificity is09089 Except counting accuracy we also do a T-test forour binary class classification e T-test is also called theStudentrsquos t-test It is a statistical hypothesis test in whichthe test statistic follows a Studentrsquos t-distributionUsually the t-test is used to compare whether there is asignificant difference between two groups of data andassists in judging the data divergence In doing a pairedsamples t-test with results from binary class classificationand random judgment its outcome is 1 for the hypothesistest result zero for the p value and [03934 04033] forthe confidence interval under null hypothesis at the 5significant level

e hypothesis test result is an index that tells whethertwo data come from the same distribution or not If the datacome from the same distribution the value of the hypothesistest result will be close to 0 On the contrary if the dataresources are distinct the result will be close to 1 whichmeans there is a differentiation between the datae p valueis the probability of accepting the assumption that there is adifference between two data may be wronge smaller the p

value the more reason that there is a disparity between dataAlso we designed an app called ldquoDeep Retinardquo pro-

viding personal examination remote medical care and earlyscreening Figure 10 shows our app interfaces Afterchoosing a fundus image that the user wants to check it willsend the image to our server and use our designed machinelearning method It takes about 10 s (depends on networkspeed) to get the result which will be presented as theprobability of each lesion With a handheld device in-dividuals can do the initial examination at the district officeor even at home More importantly it can benefit someremote areas that lack medical resources

6 Discussion

61 Accuracy Improvement Table 4 shows the accuracycomparison when using different classifiers and pa-rameter optimization methods on each dataset Using thedefault parameters with SVM (without optimization)accuracies in both validation and test sets are higher thanthat of the RF [39] If we optimize the parameters usingthe default parameter searching method provided in theLIBSVM software package though it achieves very highaccuracy in the five-fold cross validation experiment the

Mobile Information Systems 9

validation and test accuracies are even lower than thedefault one From this result we believe that overttingarises when optimizing parameters in SVM

Table 5 shows the confusion matrix of the classicationresults from the two DCNN networks (before performingSVM classication) Network 1 is the architecture shown on

Yes

No

Reject

Yes

No

No

Yes

Yes

Start

Initial students terminal criterion

i = 1

Calculate mean of each design variables

Identify best teachersolution

Based on best teachersolution to modifyXnewi = Xoldi + differencei = Xoldi +

ri (Xteacher ndash (TF)mean)

Randomly select any solution Xj (Xj ne Xi)

i = i + 1

Final value of solution and end

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is Xi better than Xjf (Xi) gt f (Xj)

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is termination criteria satisfiedf (Xnewi) gt f (Xoldi)

No

Yes

No

Xnewi = Xoldi + ri (Xi ndash Xj)

Xi = Xnewi

Xnewi = Xoldi + ri (Xj ndash Xi)

i le N

Xi = Xnewi

Figure 8 TLBO owchart

10 Mobile Information Systems

the left side of Figure 9 and network 2 is the one on the rightside From Table 5 it shows that the lesion classications of 0and 2 are better than the other categories For lesion 1 mostof the prediction results are incorrect Also for lesions 3 and4 the majority of the results are misclassications that areclassied into lesion 2

Table 6 shows the confusion matrix of the classicationresults using the full procedure of the proposed method(using SVM with TLBO) Table 7 displays the dierencebetween Tables 5 and 6 which serves as a performancecomparison between the two methods (using DCNN onlyand DCNN + SVM + TLBO) From the table every classexcept class 0 and overall accuracy is increased in network 1For network 2 each accuracy except class 3 is increasede decline in accuracy of class 3 is mainly caused by

misclassication of class 2 Table 8 shows the confusionmatrix of the classication results using the baseline methodas reported in [39] for comparison purposes

62 Deep Learning vs Traditional Classication MethodsMany traditional classication methods try to solve theproblem of DR detection by (1) using image processing tocapture symptoms in fundus images and then (2) building aclassier to make decisions based on the detected symptoms(1) e shortcoming of image processing methods is that themanifestations of the symptoms are random across dierentimages therefore it is extremely time-consuming and re-quires intense eorts to label the locations of the symptomsAbiding by the new philosophy that comes with the emer-gence of the deep learning technology our proposed methodis trying to learn how to make decisions directly from theimage data itself Dierent than the former approaches ourimages only need to be labeled with lesion number instead oflabeling symptom locations Consequently it saves consid-erable time during the database preprocessing stage On top ofthe classication results by the two DCNN networks we useSVM optimized by TLBO to generate an improved outcomeand we achieve 8617 accuracy Our result is better than therst-place winner in the Kaggle competition It shows that ourresearch result is the state-of-the-art

63 Limitation In our current datasets the number ofimages of lesions 3 and 4 is not suiexclcient to train a networkwhich is a limitation of the proposed methoderefore oneof our future works is to develop deeper collaborative re-lations with hospitals and clinics to acquire more data oflesions 3 and 4 With more data we believe the classicationaccuracy will be further increased In addition from ourresult we found that it is hard to dierentiate the imagesbetween lesions 0 and 1 erefore when we collect newdata it is desirable to collect more images belonging tolesions 0 and 1 Also we can attempt to use a dierentnetwork architecture for this problem

7 Conclusion

It is feasible to train a deep learning model for automaticdiagnosis of DR as long as we have enough data for sta-tistical model training Furthermore the database prepa-ration stage only needs a categorical label for each trainingimage It does not require detailed annotation for retinalvessel tracking in every image Hence it is time-eiexclcientcompared to the traditional machine learning-based methodfor automatic diagnosis of DR e nal accuracy canachieve 8617 and 9105 for ve-class and binary classclassications respectively

e sensitivity and specicity of binary classication are08930 and 09089 respectively which is a satisfactory resultFurthermore we developed an automatic inspection appthat can be used in both personal examination and remotemedical care With more image data collected we expect theaccuracy can be even more enhanced further improving oursystem

Table 3 Detailed information of each dataset

Dataset of imagesDR lesions

0 1 2 3 4Train 34124 73 6 15 3 3Validation 1000 70 8 15 4 3Test 53572 74 7 15 2 2

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

3 times 3 conv 320

3 times 3 conv 352

3 times 3 conv 384

2 times 2 conv 416

1 times 1 conv 448

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

Softmax

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

2 times 2 conv 320

1 times 1 conv 356

Softmax

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

Figure 9 Architecture of the two DCNN networks that we used

Mobile Information Systems 11

Table 4 Accuracy comparison when using different classifiers and different parameter optimization methods on each dataset

DatasetMethod

SVM RF [39]Without optimization Default parameter searching TLBO

Five-fold cross validation (using training data) 88744 965724 mdash mdashValidation data 854 796 mdash 847Test data 861177 810573 8617 8602

Table 5 Confusion matrix of the classification results from the two DCNN networks before performing SVM classification

Determined levelGround-truth level

Network 1 Network 20 1 2 3 4 0 1 2 3 4

0 39031 2693 2361 75 160 38310 2233 1617 51 781 118 438 185 0 1 327 735 290 1 32 339 626 5058 738 392 790 786 5509 620 3453 0 0 170 332 102 4 1 338 486 1854 43 5 85 69 551 100 7 105 56 595Accuracy () 9874 1164 6436 2735 4569 9691 1954 7001 4003 4934Overall accuracy () 8476 8518

(a) (b) (c) (d)

Figure 10 App interfaces

Table 6 Confusion matrix of the classification results using the full procedure of the proposed method (using SVM with TLBO)

Determined levelGround-truth level

SVM optimized with TLBO0 1 2 3 4

0 38611 2233 1513 44 801 312 831 349 1 12 570 696 5706 701 3933 3 0 215 408 1234 35 2 76 60 609Accuracy () 9767 2209 7260 3361 5050Overall accuracy () 8617

12 Mobile Information Systems

Data Availability

All experiment data come from a Kaggle contest calledldquoIdentify signs of diabetic retinopathy in eye imagesrdquo(website httpswwwkagglecomcdiabetic-retinopathy-detection)

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e authors would like to thank Mr Chia-Ming Hu for hishelp in preparing computational instruments and per-forming parts of the experiments in this paper

References

[1] N H Cho J E Shaw S Karuranga et al ldquoIDF Diabetes Atlasglobal estimates of diabetes prevalence for 2017 and pro-jections for 2045rdquo Diabetes Research and Clinical Practicevol 138 pp 271ndash281 2018

[2] American Diabetes Association Eye Complications AmericanDiabetes Association Arlington VA USA 2013 httpwwwdiabetesorgliving-with-diabetescomplicationseye-complicationsreferrerhttpswwwgooglecomtw

[3] International Diabetes Foundation (IDF) Diabetes AtlasInternational Diabetes Federation Brussels Belgium 7thedition 2015

[4] C P Wilkinson F L Ferris R E Klein et al ldquoProposedinternational clinical diabetic retinopathy and diabetic

macular edema disease severity scalesrdquo Ophthalmologyvol 110 no 9 pp 1677ndash1682 2003

[5] A Tufail C Rudisill C Egan et al ldquoAutomated diabeticretinopathy image assessment software diagnostic accuracyand cost-effectiveness compared with human gradersrdquoOphthalmology vol 124 no 3 pp 343ndash351 2017

[6] A Tufail V V Kapetanakis S Salas-Vega et al ldquoAn ob-servational study to assess if automated diabetic retinopathyimage assessment software can replace one or more steps ofmanual imaging grading and to determine their cost-effec-tivenessrdquo Health Technology Assessment vol 20 no 92pp 1ndash72 2016

[7] I Liu and Y Sun ldquoRecursive tracking of vascular networks inangiograms based on the detection-deletion schemerdquo IEEETransactions on Medical Imaging vol 12 no 2 pp 334ndash3411993

[8] A Can H Shen J N Turner H L Tanenbaum andB Roysam ldquoRapid automated tracing and feature extractionfrom retinal fundus images using direct exploratory algo-rithmsrdquo IEEE Transactions on Information Technology inBiomedicine vol 3 no 2 pp 125ndash138 1999

[9] M Vlachos and E Dermatas ldquoMulti-scale retinal vesselsegmentation using line trackingrdquo Computerized MedicalImaging and Graphics vol 34 no 3 pp 213ndash227 2010

[10] Y Yin M Adel and S Bourennane ldquoAn automatic trackingmethod for retinal vascular tree extractionrdquo in Proceedings ofthe IEEE International Conference on Acoustics Speech andSignal Processing IEEE Kyoto Japan March 2012

[11] S Chaudhuri S Chatterjee N Katz M Nelson andM Goldbaum ldquoDetection of blood vessels in retinal imagesusing two-dimensional matched filtersrdquo IEEE Transactions onMedical Imaging vol 8 no 3 pp 263ndash269 1989

Table 8 Confusion matrix of the classification using the baseline method [39] and its overall accuracy

DeterminedOriginal

Random forest0 1 2 3 4

0 38228 2001 1331 38 641 546 1026 434 3 22 710 732 5726 635 3583 2 1 279 465 1464 45 2 89 73 636Accuracy () 9670 2727 7286 3830 5274Overall accuracy () 8601

Table 7 Differences between Tables 5 and 6

Determined lesionGround-truth lesion

SVM-network 1 SVM-network 20 1 2 3 4 0 1 2 3 4

0 +420 minus460 minus848 minus31 minus80 +310 0 minus104 minus43 +21 +194 +393 +164 +1 0 minus15 +96 +59 0 minus22 +231 +70 +648 minus37 +1 minus220 minus90 +197 +81 +483 +3 0 +45 +76 +21 +1 minus1 minus123 minus78 minus624 +3 minus3 +45 +76 +58 minus65 minus5 minus28 +4 +14Accuracy increment () minus108 8977 1280 2288 1052 078 1305 369 minus1603 23Overall accuracy increment () 16 116

Mobile Information Systems 13

For yi whose value is 1 the data must satisfywrarr

middot xrarrminus bge 1 and for yi whose value is minus1 w

rarrmiddot xrarrminus bleminus1

has to be satisfied Combining these two conditions we get

yi(wrarr

middot xrarrminus b)ge 1 (6)

e goal is to maximize bwrarr

according to the constrainof equation (6) in order to derive the optimized decisionhyperplane for classification

Sometimes the training data might not be able to beperfectly separated using linear boundarieserefore in theSVM formulation we need to introduce the error metric εand the cost parameter C as shown in equation (7) e goalnow becomes to minimize

12

wrarr

middot wrarr

+ C 1113944N

i1εi (7)

Subject to

wrarr

middot xirarr

+ bge 1minus εi if wrarr

middot xirarrgt b

wrarr

middot xirarr

+ bleminus1 + εi if wrarr

middot xirarrlt b

εi gt 0

(8)

e performance of SVM is influenced by two mainparameters the first one is C which is a tunable parameter inequation (7) e other one is c which is used in the radialbasis function (RBF) kernel to map data into a higher di-mensional space before training and classification e RBFkernel can be defined as

K xi xj1113872 1113873 eminusc ximinusxj

111386811138681113868111386811138681113868111386811138682

(9)

where c denotes the width of the Gaussian envelope in ahigh-dimensional feature space

44 Teaching-Learning-Based Optimization (TLBO)TLBO an evolution-based optimization algorithm wasproposed by Rao et al [41] in 2011 e concept of TLBO isinspired by the evolution of the learning process when agroup or a class of learners learn a target task ere are two

ways of learning in groups or classes (1) learning from theguidance of the teacher and (2) learning from other learnerse procedure of TLBO can be divided into two phases asdescribed below in Sections 441 and 442

441 Teacher Phase In the whole population the teacher(Xteacher) can be considered as the best solution Namelylearners learn from the teacher in the teacher phase In thisphase the teacher strives for enhancing the results of otherindividuals (Xi) by increasing the mean result of theclassroom is can be described as adjusting Xmean toapproximate Xteacher In order to maintain a stochasticnature during the optimization process two randomlygenerated parameters r and TF are applied in each iterationfor the solution Xi as

Xnew Xi + ri Xteacher minusTFXmean( 1113857 (10)

TF round[1 + rand(0 1)] (11)

In equation (10) ri is a randomly selected number in therange of 0 and 1 Moreover Xnew and Xi are the new andexisting solutions at iteration i respectively TF in equation(11) is a teaching factor which can be either 1 or 2

a = 1 b = 1

x = 1 2 hellip y = 1 2 hellip

x = 1 3 hellip y = 1 2 hellip

x = 1 2 hellip y = 1 3 hellip

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 1] times [1 1]

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 2] times [1 1]

P11 = [x1minus1 x1 minus 1] times [y1minus1 y1 minus 1] = [1 1] times [1 2]

Figure 6 Different pooling types when both a and b equal to onee left side shows the equations for pooling condition and computationof pooling region Pictures on the right side indicate regions before and after pooling

w middot x ndash

b = 1

w middot x ndash

b = 0

w middot x ndash

b = ndash1

Figure 7 Example of decision boundary hyperplane with twoclasses of samples

8 Mobile Information Systems

442 Learner Phase e learners gain their knowledge byinteracting with each other erefore an individual learnsnew information if other individuals have more knowledgethan him or her In this phase the student Xi interactsrandomly with another student Xj (ine j) in order to en-hance his or her knowledge Equation (12) shows that if Xi isbetter than Xj (ie f(Xj)gtf(Xi) for minimizationproblems) Xj is moved toward Xi Otherwise it is movedaway from Xi

Xnew Xi + ri Xi minusXj1113872 1113873 f Xj1113872 1113873gtf Xi( 1113857

Xi + ri Xj minusXi1113872 1113873 f Xi( 1113857gtf Xj1113872 1113873

⎧⎪⎨

⎪⎩(12)

If the new solution Xnew to the problem is better than theold ones the new solution Xnew will be recorded as the bestsolution After updating the status of each learner a newiteration begins A stop criterion based on the iterationnumber or the difference of the cost function can be set tostop the iteration properly e flowchart of TLBO is shownin Figure 8

5 Result

Our fundus image data is from the database provided by oneof the Kaggle contests entitled ldquoIdentify signs of diabeticretinopathy in eye imagesrdquo [42] In this database there areabout 90000 images We separate 1000 images from thetraining dataset to be the validation dataset e detailedinformation of each dataset is shown in Table 3 and our twonetwork architectures are shown in Figure 9

Our proposed method uses two DCNNs with fractionalmax-pooling layers For every input fundus image the twoDCNN will output a vector of size 1 times 5 representing theprobability distribution of the prediction for each lesion(category) e probability distribution together with othervalues forms a feature with dimensionality 24 e 24features are described as follows

(i) DCNN probabilities of each lesion respectively (5features)

(ii) Averages of R G and B channel values within 50 lowast50 center cropped image (3 features)

(iii) Widths and heights of 50 lowast 50 center croppedimage (2 features)

(iv) Overall standard deviation of the original image and50 lowast 50 centered cropped image Laplacian-filtered image (2 features)

(v) In total there are 12 features for one fundusimage We then append another 12 features fromthe fundus image of the other eye of the samesubject erefore the overall length of the featurevector is 24 for one fundus image e 24 featurevectors of dimensionality are used as input vec-tors of SVM

e 24-dimensional vector is used to train a multiclassSVM (five classes) whose parameters are optimized usingthe TLBO method We implemented the method described

in [39] and used it as the baseline e baseline system usessimilar features with a scheme of ensemble classifier (RF)

We used the validation set data to optimize the pa-rameter set (C c) in SVM using TLBO e upper andlower bounds of the parameter are set within [0 100] Weran 50 iterations with 50 students

Our final accuracy for five-class classification task ofDR is 8617 and the accuracy for the binary classclassification task is 9105 Labels for five-class classi-fication are normal NPDR level 1 NPDR level 2 NPDRlevel 3 and PDR while labels for binary class classifi-cation are normal and abnormal For binary classifica-tion its sensitivity is 08930 while the specificity is09089 Except counting accuracy we also do a T-test forour binary class classification e T-test is also called theStudentrsquos t-test It is a statistical hypothesis test in whichthe test statistic follows a Studentrsquos t-distributionUsually the t-test is used to compare whether there is asignificant difference between two groups of data andassists in judging the data divergence In doing a pairedsamples t-test with results from binary class classificationand random judgment its outcome is 1 for the hypothesistest result zero for the p value and [03934 04033] forthe confidence interval under null hypothesis at the 5significant level

e hypothesis test result is an index that tells whethertwo data come from the same distribution or not If the datacome from the same distribution the value of the hypothesistest result will be close to 0 On the contrary if the dataresources are distinct the result will be close to 1 whichmeans there is a differentiation between the datae p valueis the probability of accepting the assumption that there is adifference between two data may be wronge smaller the p

value the more reason that there is a disparity between dataAlso we designed an app called ldquoDeep Retinardquo pro-

viding personal examination remote medical care and earlyscreening Figure 10 shows our app interfaces Afterchoosing a fundus image that the user wants to check it willsend the image to our server and use our designed machinelearning method It takes about 10 s (depends on networkspeed) to get the result which will be presented as theprobability of each lesion With a handheld device in-dividuals can do the initial examination at the district officeor even at home More importantly it can benefit someremote areas that lack medical resources

6 Discussion

61 Accuracy Improvement Table 4 shows the accuracycomparison when using different classifiers and pa-rameter optimization methods on each dataset Using thedefault parameters with SVM (without optimization)accuracies in both validation and test sets are higher thanthat of the RF [39] If we optimize the parameters usingthe default parameter searching method provided in theLIBSVM software package though it achieves very highaccuracy in the five-fold cross validation experiment the

Mobile Information Systems 9

validation and test accuracies are even lower than thedefault one From this result we believe that overttingarises when optimizing parameters in SVM

Table 5 shows the confusion matrix of the classicationresults from the two DCNN networks (before performingSVM classication) Network 1 is the architecture shown on

Yes

No

Reject

Yes

No

No

Yes

Yes

Start

Initial students terminal criterion

i = 1

Calculate mean of each design variables

Identify best teachersolution

Based on best teachersolution to modifyXnewi = Xoldi + differencei = Xoldi +

ri (Xteacher ndash (TF)mean)

Randomly select any solution Xj (Xj ne Xi)

i = i + 1

Final value of solution and end

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is Xi better than Xjf (Xi) gt f (Xj)

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is termination criteria satisfiedf (Xnewi) gt f (Xoldi)

No

Yes

No

Xnewi = Xoldi + ri (Xi ndash Xj)

Xi = Xnewi

Xnewi = Xoldi + ri (Xj ndash Xi)

i le N

Xi = Xnewi

Figure 8 TLBO owchart

10 Mobile Information Systems

the left side of Figure 9 and network 2 is the one on the rightside From Table 5 it shows that the lesion classications of 0and 2 are better than the other categories For lesion 1 mostof the prediction results are incorrect Also for lesions 3 and4 the majority of the results are misclassications that areclassied into lesion 2

Table 6 shows the confusion matrix of the classicationresults using the full procedure of the proposed method(using SVM with TLBO) Table 7 displays the dierencebetween Tables 5 and 6 which serves as a performancecomparison between the two methods (using DCNN onlyand DCNN + SVM + TLBO) From the table every classexcept class 0 and overall accuracy is increased in network 1For network 2 each accuracy except class 3 is increasede decline in accuracy of class 3 is mainly caused by

misclassication of class 2 Table 8 shows the confusionmatrix of the classication results using the baseline methodas reported in [39] for comparison purposes

62 Deep Learning vs Traditional Classication MethodsMany traditional classication methods try to solve theproblem of DR detection by (1) using image processing tocapture symptoms in fundus images and then (2) building aclassier to make decisions based on the detected symptoms(1) e shortcoming of image processing methods is that themanifestations of the symptoms are random across dierentimages therefore it is extremely time-consuming and re-quires intense eorts to label the locations of the symptomsAbiding by the new philosophy that comes with the emer-gence of the deep learning technology our proposed methodis trying to learn how to make decisions directly from theimage data itself Dierent than the former approaches ourimages only need to be labeled with lesion number instead oflabeling symptom locations Consequently it saves consid-erable time during the database preprocessing stage On top ofthe classication results by the two DCNN networks we useSVM optimized by TLBO to generate an improved outcomeand we achieve 8617 accuracy Our result is better than therst-place winner in the Kaggle competition It shows that ourresearch result is the state-of-the-art

63 Limitation In our current datasets the number ofimages of lesions 3 and 4 is not suiexclcient to train a networkwhich is a limitation of the proposed methoderefore oneof our future works is to develop deeper collaborative re-lations with hospitals and clinics to acquire more data oflesions 3 and 4 With more data we believe the classicationaccuracy will be further increased In addition from ourresult we found that it is hard to dierentiate the imagesbetween lesions 0 and 1 erefore when we collect newdata it is desirable to collect more images belonging tolesions 0 and 1 Also we can attempt to use a dierentnetwork architecture for this problem

7 Conclusion

It is feasible to train a deep learning model for automaticdiagnosis of DR as long as we have enough data for sta-tistical model training Furthermore the database prepa-ration stage only needs a categorical label for each trainingimage It does not require detailed annotation for retinalvessel tracking in every image Hence it is time-eiexclcientcompared to the traditional machine learning-based methodfor automatic diagnosis of DR e nal accuracy canachieve 8617 and 9105 for ve-class and binary classclassications respectively

e sensitivity and specicity of binary classication are08930 and 09089 respectively which is a satisfactory resultFurthermore we developed an automatic inspection appthat can be used in both personal examination and remotemedical care With more image data collected we expect theaccuracy can be even more enhanced further improving oursystem

Table 3 Detailed information of each dataset

Dataset of imagesDR lesions

0 1 2 3 4Train 34124 73 6 15 3 3Validation 1000 70 8 15 4 3Test 53572 74 7 15 2 2

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

3 times 3 conv 320

3 times 3 conv 352

3 times 3 conv 384

2 times 2 conv 416

1 times 1 conv 448

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

Softmax

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

2 times 2 conv 320

1 times 1 conv 356

Softmax

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

Figure 9 Architecture of the two DCNN networks that we used

Mobile Information Systems 11

Table 4 Accuracy comparison when using different classifiers and different parameter optimization methods on each dataset

DatasetMethod

SVM RF [39]Without optimization Default parameter searching TLBO

Five-fold cross validation (using training data) 88744 965724 mdash mdashValidation data 854 796 mdash 847Test data 861177 810573 8617 8602

Table 5 Confusion matrix of the classification results from the two DCNN networks before performing SVM classification

Determined levelGround-truth level

Network 1 Network 20 1 2 3 4 0 1 2 3 4

0 39031 2693 2361 75 160 38310 2233 1617 51 781 118 438 185 0 1 327 735 290 1 32 339 626 5058 738 392 790 786 5509 620 3453 0 0 170 332 102 4 1 338 486 1854 43 5 85 69 551 100 7 105 56 595Accuracy () 9874 1164 6436 2735 4569 9691 1954 7001 4003 4934Overall accuracy () 8476 8518

(a) (b) (c) (d)

Figure 10 App interfaces

Table 6 Confusion matrix of the classification results using the full procedure of the proposed method (using SVM with TLBO)

Determined levelGround-truth level

SVM optimized with TLBO0 1 2 3 4

0 38611 2233 1513 44 801 312 831 349 1 12 570 696 5706 701 3933 3 0 215 408 1234 35 2 76 60 609Accuracy () 9767 2209 7260 3361 5050Overall accuracy () 8617

12 Mobile Information Systems

Data Availability

All experiment data come from a Kaggle contest calledldquoIdentify signs of diabetic retinopathy in eye imagesrdquo(website httpswwwkagglecomcdiabetic-retinopathy-detection)

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e authors would like to thank Mr Chia-Ming Hu for hishelp in preparing computational instruments and per-forming parts of the experiments in this paper

References

[1] N H Cho J E Shaw S Karuranga et al ldquoIDF Diabetes Atlasglobal estimates of diabetes prevalence for 2017 and pro-jections for 2045rdquo Diabetes Research and Clinical Practicevol 138 pp 271ndash281 2018

[2] American Diabetes Association Eye Complications AmericanDiabetes Association Arlington VA USA 2013 httpwwwdiabetesorgliving-with-diabetescomplicationseye-complicationsreferrerhttpswwwgooglecomtw

[3] International Diabetes Foundation (IDF) Diabetes AtlasInternational Diabetes Federation Brussels Belgium 7thedition 2015

[4] C P Wilkinson F L Ferris R E Klein et al ldquoProposedinternational clinical diabetic retinopathy and diabetic

macular edema disease severity scalesrdquo Ophthalmologyvol 110 no 9 pp 1677ndash1682 2003

[5] A Tufail C Rudisill C Egan et al ldquoAutomated diabeticretinopathy image assessment software diagnostic accuracyand cost-effectiveness compared with human gradersrdquoOphthalmology vol 124 no 3 pp 343ndash351 2017

[6] A Tufail V V Kapetanakis S Salas-Vega et al ldquoAn ob-servational study to assess if automated diabetic retinopathyimage assessment software can replace one or more steps ofmanual imaging grading and to determine their cost-effec-tivenessrdquo Health Technology Assessment vol 20 no 92pp 1ndash72 2016

[7] I Liu and Y Sun ldquoRecursive tracking of vascular networks inangiograms based on the detection-deletion schemerdquo IEEETransactions on Medical Imaging vol 12 no 2 pp 334ndash3411993

[8] A Can H Shen J N Turner H L Tanenbaum andB Roysam ldquoRapid automated tracing and feature extractionfrom retinal fundus images using direct exploratory algo-rithmsrdquo IEEE Transactions on Information Technology inBiomedicine vol 3 no 2 pp 125ndash138 1999

[9] M Vlachos and E Dermatas ldquoMulti-scale retinal vesselsegmentation using line trackingrdquo Computerized MedicalImaging and Graphics vol 34 no 3 pp 213ndash227 2010

[10] Y Yin M Adel and S Bourennane ldquoAn automatic trackingmethod for retinal vascular tree extractionrdquo in Proceedings ofthe IEEE International Conference on Acoustics Speech andSignal Processing IEEE Kyoto Japan March 2012

[11] S Chaudhuri S Chatterjee N Katz M Nelson andM Goldbaum ldquoDetection of blood vessels in retinal imagesusing two-dimensional matched filtersrdquo IEEE Transactions onMedical Imaging vol 8 no 3 pp 263ndash269 1989

Table 8 Confusion matrix of the classification using the baseline method [39] and its overall accuracy

DeterminedOriginal

Random forest0 1 2 3 4

0 38228 2001 1331 38 641 546 1026 434 3 22 710 732 5726 635 3583 2 1 279 465 1464 45 2 89 73 636Accuracy () 9670 2727 7286 3830 5274Overall accuracy () 8601

Table 7 Differences between Tables 5 and 6

Determined lesionGround-truth lesion

SVM-network 1 SVM-network 20 1 2 3 4 0 1 2 3 4

0 +420 minus460 minus848 minus31 minus80 +310 0 minus104 minus43 +21 +194 +393 +164 +1 0 minus15 +96 +59 0 minus22 +231 +70 +648 minus37 +1 minus220 minus90 +197 +81 +483 +3 0 +45 +76 +21 +1 minus1 minus123 minus78 minus624 +3 minus3 +45 +76 +58 minus65 minus5 minus28 +4 +14Accuracy increment () minus108 8977 1280 2288 1052 078 1305 369 minus1603 23Overall accuracy increment () 16 116

Mobile Information Systems 13

442 Learner Phase e learners gain their knowledge byinteracting with each other erefore an individual learnsnew information if other individuals have more knowledgethan him or her In this phase the student Xi interactsrandomly with another student Xj (ine j) in order to en-hance his or her knowledge Equation (12) shows that if Xi isbetter than Xj (ie f(Xj)gtf(Xi) for minimizationproblems) Xj is moved toward Xi Otherwise it is movedaway from Xi

Xnew Xi + ri Xi minusXj1113872 1113873 f Xj1113872 1113873gtf Xi( 1113857

Xi + ri Xj minusXi1113872 1113873 f Xi( 1113857gtf Xj1113872 1113873

⎧⎪⎨

⎪⎩(12)

If the new solution Xnew to the problem is better than theold ones the new solution Xnew will be recorded as the bestsolution After updating the status of each learner a newiteration begins A stop criterion based on the iterationnumber or the difference of the cost function can be set tostop the iteration properly e flowchart of TLBO is shownin Figure 8

5 Result

Our fundus image data is from the database provided by oneof the Kaggle contests entitled ldquoIdentify signs of diabeticretinopathy in eye imagesrdquo [42] In this database there areabout 90000 images We separate 1000 images from thetraining dataset to be the validation dataset e detailedinformation of each dataset is shown in Table 3 and our twonetwork architectures are shown in Figure 9

Our proposed method uses two DCNNs with fractionalmax-pooling layers For every input fundus image the twoDCNN will output a vector of size 1 times 5 representing theprobability distribution of the prediction for each lesion(category) e probability distribution together with othervalues forms a feature with dimensionality 24 e 24features are described as follows

(i) DCNN probabilities of each lesion respectively (5features)

(ii) Averages of R G and B channel values within 50 lowast50 center cropped image (3 features)

(iii) Widths and heights of 50 lowast 50 center croppedimage (2 features)

(iv) Overall standard deviation of the original image and50 lowast 50 centered cropped image Laplacian-filtered image (2 features)

(v) In total there are 12 features for one fundusimage We then append another 12 features fromthe fundus image of the other eye of the samesubject erefore the overall length of the featurevector is 24 for one fundus image e 24 featurevectors of dimensionality are used as input vec-tors of SVM

e 24-dimensional vector is used to train a multiclassSVM (five classes) whose parameters are optimized usingthe TLBO method We implemented the method described

in [39] and used it as the baseline e baseline system usessimilar features with a scheme of ensemble classifier (RF)

We used the validation set data to optimize the pa-rameter set (C c) in SVM using TLBO e upper andlower bounds of the parameter are set within [0 100] Weran 50 iterations with 50 students

Our final accuracy for five-class classification task ofDR is 8617 and the accuracy for the binary classclassification task is 9105 Labels for five-class classi-fication are normal NPDR level 1 NPDR level 2 NPDRlevel 3 and PDR while labels for binary class classifi-cation are normal and abnormal For binary classifica-tion its sensitivity is 08930 while the specificity is09089 Except counting accuracy we also do a T-test forour binary class classification e T-test is also called theStudentrsquos t-test It is a statistical hypothesis test in whichthe test statistic follows a Studentrsquos t-distributionUsually the t-test is used to compare whether there is asignificant difference between two groups of data andassists in judging the data divergence In doing a pairedsamples t-test with results from binary class classificationand random judgment its outcome is 1 for the hypothesistest result zero for the p value and [03934 04033] forthe confidence interval under null hypothesis at the 5significant level

e hypothesis test result is an index that tells whethertwo data come from the same distribution or not If the datacome from the same distribution the value of the hypothesistest result will be close to 0 On the contrary if the dataresources are distinct the result will be close to 1 whichmeans there is a differentiation between the datae p valueis the probability of accepting the assumption that there is adifference between two data may be wronge smaller the p

value the more reason that there is a disparity between dataAlso we designed an app called ldquoDeep Retinardquo pro-

viding personal examination remote medical care and earlyscreening Figure 10 shows our app interfaces Afterchoosing a fundus image that the user wants to check it willsend the image to our server and use our designed machinelearning method It takes about 10 s (depends on networkspeed) to get the result which will be presented as theprobability of each lesion With a handheld device in-dividuals can do the initial examination at the district officeor even at home More importantly it can benefit someremote areas that lack medical resources

6 Discussion

61 Accuracy Improvement Table 4 shows the accuracycomparison when using different classifiers and pa-rameter optimization methods on each dataset Using thedefault parameters with SVM (without optimization)accuracies in both validation and test sets are higher thanthat of the RF [39] If we optimize the parameters usingthe default parameter searching method provided in theLIBSVM software package though it achieves very highaccuracy in the five-fold cross validation experiment the

Mobile Information Systems 9

validation and test accuracies are even lower than thedefault one From this result we believe that overttingarises when optimizing parameters in SVM

Table 5 shows the confusion matrix of the classicationresults from the two DCNN networks (before performingSVM classication) Network 1 is the architecture shown on

Yes

No

Reject

Yes

No

No

Yes

Yes

Start

Initial students terminal criterion

i = 1

Calculate mean of each design variables

Identify best teachersolution

Based on best teachersolution to modifyXnewi = Xoldi + differencei = Xoldi +

ri (Xteacher ndash (TF)mean)

Randomly select any solution Xj (Xj ne Xi)

i = i + 1

Final value of solution and end

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is Xi better than Xjf (Xi) gt f (Xj)

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is termination criteria satisfiedf (Xnewi) gt f (Xoldi)

No

Yes

No

Xnewi = Xoldi + ri (Xi ndash Xj)

Xi = Xnewi

Xnewi = Xoldi + ri (Xj ndash Xi)

i le N

Xi = Xnewi

Figure 8 TLBO owchart

10 Mobile Information Systems

the left side of Figure 9 and network 2 is the one on the rightside From Table 5 it shows that the lesion classications of 0and 2 are better than the other categories For lesion 1 mostof the prediction results are incorrect Also for lesions 3 and4 the majority of the results are misclassications that areclassied into lesion 2

Table 6 shows the confusion matrix of the classicationresults using the full procedure of the proposed method(using SVM with TLBO) Table 7 displays the dierencebetween Tables 5 and 6 which serves as a performancecomparison between the two methods (using DCNN onlyand DCNN + SVM + TLBO) From the table every classexcept class 0 and overall accuracy is increased in network 1For network 2 each accuracy except class 3 is increasede decline in accuracy of class 3 is mainly caused by

misclassication of class 2 Table 8 shows the confusionmatrix of the classication results using the baseline methodas reported in [39] for comparison purposes

62 Deep Learning vs Traditional Classication MethodsMany traditional classication methods try to solve theproblem of DR detection by (1) using image processing tocapture symptoms in fundus images and then (2) building aclassier to make decisions based on the detected symptoms(1) e shortcoming of image processing methods is that themanifestations of the symptoms are random across dierentimages therefore it is extremely time-consuming and re-quires intense eorts to label the locations of the symptomsAbiding by the new philosophy that comes with the emer-gence of the deep learning technology our proposed methodis trying to learn how to make decisions directly from theimage data itself Dierent than the former approaches ourimages only need to be labeled with lesion number instead oflabeling symptom locations Consequently it saves consid-erable time during the database preprocessing stage On top ofthe classication results by the two DCNN networks we useSVM optimized by TLBO to generate an improved outcomeand we achieve 8617 accuracy Our result is better than therst-place winner in the Kaggle competition It shows that ourresearch result is the state-of-the-art

63 Limitation In our current datasets the number ofimages of lesions 3 and 4 is not suiexclcient to train a networkwhich is a limitation of the proposed methoderefore oneof our future works is to develop deeper collaborative re-lations with hospitals and clinics to acquire more data oflesions 3 and 4 With more data we believe the classicationaccuracy will be further increased In addition from ourresult we found that it is hard to dierentiate the imagesbetween lesions 0 and 1 erefore when we collect newdata it is desirable to collect more images belonging tolesions 0 and 1 Also we can attempt to use a dierentnetwork architecture for this problem

7 Conclusion

It is feasible to train a deep learning model for automaticdiagnosis of DR as long as we have enough data for sta-tistical model training Furthermore the database prepa-ration stage only needs a categorical label for each trainingimage It does not require detailed annotation for retinalvessel tracking in every image Hence it is time-eiexclcientcompared to the traditional machine learning-based methodfor automatic diagnosis of DR e nal accuracy canachieve 8617 and 9105 for ve-class and binary classclassications respectively

e sensitivity and specicity of binary classication are08930 and 09089 respectively which is a satisfactory resultFurthermore we developed an automatic inspection appthat can be used in both personal examination and remotemedical care With more image data collected we expect theaccuracy can be even more enhanced further improving oursystem

Table 3 Detailed information of each dataset

Dataset of imagesDR lesions

0 1 2 3 4Train 34124 73 6 15 3 3Validation 1000 70 8 15 4 3Test 53572 74 7 15 2 2

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

3 times 3 conv 320

3 times 3 conv 352

3 times 3 conv 384

2 times 2 conv 416

1 times 1 conv 448

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

Softmax

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

2 times 2 conv 320

1 times 1 conv 356

Softmax

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

Figure 9 Architecture of the two DCNN networks that we used

Mobile Information Systems 11

Table 4 Accuracy comparison when using different classifiers and different parameter optimization methods on each dataset

DatasetMethod

SVM RF [39]Without optimization Default parameter searching TLBO

Five-fold cross validation (using training data) 88744 965724 mdash mdashValidation data 854 796 mdash 847Test data 861177 810573 8617 8602

Table 5 Confusion matrix of the classification results from the two DCNN networks before performing SVM classification

Determined levelGround-truth level

Network 1 Network 20 1 2 3 4 0 1 2 3 4

0 39031 2693 2361 75 160 38310 2233 1617 51 781 118 438 185 0 1 327 735 290 1 32 339 626 5058 738 392 790 786 5509 620 3453 0 0 170 332 102 4 1 338 486 1854 43 5 85 69 551 100 7 105 56 595Accuracy () 9874 1164 6436 2735 4569 9691 1954 7001 4003 4934Overall accuracy () 8476 8518

(a) (b) (c) (d)

Figure 10 App interfaces

Table 6 Confusion matrix of the classification results using the full procedure of the proposed method (using SVM with TLBO)

Determined levelGround-truth level

SVM optimized with TLBO0 1 2 3 4

0 38611 2233 1513 44 801 312 831 349 1 12 570 696 5706 701 3933 3 0 215 408 1234 35 2 76 60 609Accuracy () 9767 2209 7260 3361 5050Overall accuracy () 8617

12 Mobile Information Systems

Data Availability

All experiment data come from a Kaggle contest calledldquoIdentify signs of diabetic retinopathy in eye imagesrdquo(website httpswwwkagglecomcdiabetic-retinopathy-detection)

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e authors would like to thank Mr Chia-Ming Hu for hishelp in preparing computational instruments and per-forming parts of the experiments in this paper

References

[1] N H Cho J E Shaw S Karuranga et al ldquoIDF Diabetes Atlasglobal estimates of diabetes prevalence for 2017 and pro-jections for 2045rdquo Diabetes Research and Clinical Practicevol 138 pp 271ndash281 2018

[2] American Diabetes Association Eye Complications AmericanDiabetes Association Arlington VA USA 2013 httpwwwdiabetesorgliving-with-diabetescomplicationseye-complicationsreferrerhttpswwwgooglecomtw

[3] International Diabetes Foundation (IDF) Diabetes AtlasInternational Diabetes Federation Brussels Belgium 7thedition 2015

[4] C P Wilkinson F L Ferris R E Klein et al ldquoProposedinternational clinical diabetic retinopathy and diabetic

macular edema disease severity scalesrdquo Ophthalmologyvol 110 no 9 pp 1677ndash1682 2003

[5] A Tufail C Rudisill C Egan et al ldquoAutomated diabeticretinopathy image assessment software diagnostic accuracyand cost-effectiveness compared with human gradersrdquoOphthalmology vol 124 no 3 pp 343ndash351 2017

[6] A Tufail V V Kapetanakis S Salas-Vega et al ldquoAn ob-servational study to assess if automated diabetic retinopathyimage assessment software can replace one or more steps ofmanual imaging grading and to determine their cost-effec-tivenessrdquo Health Technology Assessment vol 20 no 92pp 1ndash72 2016

[7] I Liu and Y Sun ldquoRecursive tracking of vascular networks inangiograms based on the detection-deletion schemerdquo IEEETransactions on Medical Imaging vol 12 no 2 pp 334ndash3411993

[8] A Can H Shen J N Turner H L Tanenbaum andB Roysam ldquoRapid automated tracing and feature extractionfrom retinal fundus images using direct exploratory algo-rithmsrdquo IEEE Transactions on Information Technology inBiomedicine vol 3 no 2 pp 125ndash138 1999

[9] M Vlachos and E Dermatas ldquoMulti-scale retinal vesselsegmentation using line trackingrdquo Computerized MedicalImaging and Graphics vol 34 no 3 pp 213ndash227 2010

[10] Y Yin M Adel and S Bourennane ldquoAn automatic trackingmethod for retinal vascular tree extractionrdquo in Proceedings ofthe IEEE International Conference on Acoustics Speech andSignal Processing IEEE Kyoto Japan March 2012

[11] S Chaudhuri S Chatterjee N Katz M Nelson andM Goldbaum ldquoDetection of blood vessels in retinal imagesusing two-dimensional matched filtersrdquo IEEE Transactions onMedical Imaging vol 8 no 3 pp 263ndash269 1989

Table 8 Confusion matrix of the classification using the baseline method [39] and its overall accuracy

DeterminedOriginal

Random forest0 1 2 3 4

0 38228 2001 1331 38 641 546 1026 434 3 22 710 732 5726 635 3583 2 1 279 465 1464 45 2 89 73 636Accuracy () 9670 2727 7286 3830 5274Overall accuracy () 8601

Table 7 Differences between Tables 5 and 6

Determined lesionGround-truth lesion

SVM-network 1 SVM-network 20 1 2 3 4 0 1 2 3 4

0 +420 minus460 minus848 minus31 minus80 +310 0 minus104 minus43 +21 +194 +393 +164 +1 0 minus15 +96 +59 0 minus22 +231 +70 +648 minus37 +1 minus220 minus90 +197 +81 +483 +3 0 +45 +76 +21 +1 minus1 minus123 minus78 minus624 +3 minus3 +45 +76 +58 minus65 minus5 minus28 +4 +14Accuracy increment () minus108 8977 1280 2288 1052 078 1305 369 minus1603 23Overall accuracy increment () 16 116

Mobile Information Systems 13

validation and test accuracies are even lower than thedefault one From this result we believe that overttingarises when optimizing parameters in SVM

Table 5 shows the confusion matrix of the classicationresults from the two DCNN networks (before performingSVM classication) Network 1 is the architecture shown on

Yes

No

Reject

Yes

No

No

Yes

Yes

Start

Initial students terminal criterion

i = 1

Calculate mean of each design variables

Identify best teachersolution

Based on best teachersolution to modifyXnewi = Xoldi + differencei = Xoldi +

ri (Xteacher ndash (TF)mean)

Randomly select any solution Xj (Xj ne Xi)

i = i + 1

Final value of solution and end

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is Xi better than Xjf (Xi) gt f (Xj)

Is new solution better than existingf (Xnewi) gt f (Xoldi)

Is termination criteria satisfiedf (Xnewi) gt f (Xoldi)

No

Yes

No

Xnewi = Xoldi + ri (Xi ndash Xj)

Xi = Xnewi

Xnewi = Xoldi + ri (Xj ndash Xi)

i le N

Xi = Xnewi

Figure 8 TLBO owchart

10 Mobile Information Systems

the left side of Figure 9 and network 2 is the one on the rightside From Table 5 it shows that the lesion classications of 0and 2 are better than the other categories For lesion 1 mostof the prediction results are incorrect Also for lesions 3 and4 the majority of the results are misclassications that areclassied into lesion 2

Table 6 shows the confusion matrix of the classicationresults using the full procedure of the proposed method(using SVM with TLBO) Table 7 displays the dierencebetween Tables 5 and 6 which serves as a performancecomparison between the two methods (using DCNN onlyand DCNN + SVM + TLBO) From the table every classexcept class 0 and overall accuracy is increased in network 1For network 2 each accuracy except class 3 is increasede decline in accuracy of class 3 is mainly caused by

misclassication of class 2 Table 8 shows the confusionmatrix of the classication results using the baseline methodas reported in [39] for comparison purposes

62 Deep Learning vs Traditional Classication MethodsMany traditional classication methods try to solve theproblem of DR detection by (1) using image processing tocapture symptoms in fundus images and then (2) building aclassier to make decisions based on the detected symptoms(1) e shortcoming of image processing methods is that themanifestations of the symptoms are random across dierentimages therefore it is extremely time-consuming and re-quires intense eorts to label the locations of the symptomsAbiding by the new philosophy that comes with the emer-gence of the deep learning technology our proposed methodis trying to learn how to make decisions directly from theimage data itself Dierent than the former approaches ourimages only need to be labeled with lesion number instead oflabeling symptom locations Consequently it saves consid-erable time during the database preprocessing stage On top ofthe classication results by the two DCNN networks we useSVM optimized by TLBO to generate an improved outcomeand we achieve 8617 accuracy Our result is better than therst-place winner in the Kaggle competition It shows that ourresearch result is the state-of-the-art

63 Limitation In our current datasets the number ofimages of lesions 3 and 4 is not suiexclcient to train a networkwhich is a limitation of the proposed methoderefore oneof our future works is to develop deeper collaborative re-lations with hospitals and clinics to acquire more data oflesions 3 and 4 With more data we believe the classicationaccuracy will be further increased In addition from ourresult we found that it is hard to dierentiate the imagesbetween lesions 0 and 1 erefore when we collect newdata it is desirable to collect more images belonging tolesions 0 and 1 Also we can attempt to use a dierentnetwork architecture for this problem

7 Conclusion

It is feasible to train a deep learning model for automaticdiagnosis of DR as long as we have enough data for sta-tistical model training Furthermore the database prepa-ration stage only needs a categorical label for each trainingimage It does not require detailed annotation for retinalvessel tracking in every image Hence it is time-eiexclcientcompared to the traditional machine learning-based methodfor automatic diagnosis of DR e nal accuracy canachieve 8617 and 9105 for ve-class and binary classclassications respectively

e sensitivity and specicity of binary classication are08930 and 09089 respectively which is a satisfactory resultFurthermore we developed an automatic inspection appthat can be used in both personal examination and remotemedical care With more image data collected we expect theaccuracy can be even more enhanced further improving oursystem

Table 3 Detailed information of each dataset

Dataset of imagesDR lesions

0 1 2 3 4Train 34124 73 6 15 3 3Validation 1000 70 8 15 4 3Test 53572 74 7 15 2 2

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

3 times 3 conv 320

3 times 3 conv 352

3 times 3 conv 384

2 times 2 conv 416

1 times 1 conv 448

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

Softmax

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

2 times 2 conv 320

1 times 1 conv 356

Softmax

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

Figure 9 Architecture of the two DCNN networks that we used

Mobile Information Systems 11

Table 4 Accuracy comparison when using different classifiers and different parameter optimization methods on each dataset

DatasetMethod

SVM RF [39]Without optimization Default parameter searching TLBO

Five-fold cross validation (using training data) 88744 965724 mdash mdashValidation data 854 796 mdash 847Test data 861177 810573 8617 8602

Table 5 Confusion matrix of the classification results from the two DCNN networks before performing SVM classification

Determined levelGround-truth level

Network 1 Network 20 1 2 3 4 0 1 2 3 4

0 39031 2693 2361 75 160 38310 2233 1617 51 781 118 438 185 0 1 327 735 290 1 32 339 626 5058 738 392 790 786 5509 620 3453 0 0 170 332 102 4 1 338 486 1854 43 5 85 69 551 100 7 105 56 595Accuracy () 9874 1164 6436 2735 4569 9691 1954 7001 4003 4934Overall accuracy () 8476 8518

(a) (b) (c) (d)

Figure 10 App interfaces

Table 6 Confusion matrix of the classification results using the full procedure of the proposed method (using SVM with TLBO)

Determined levelGround-truth level

SVM optimized with TLBO0 1 2 3 4

0 38611 2233 1513 44 801 312 831 349 1 12 570 696 5706 701 3933 3 0 215 408 1234 35 2 76 60 609Accuracy () 9767 2209 7260 3361 5050Overall accuracy () 8617

12 Mobile Information Systems

Data Availability

All experiment data come from a Kaggle contest calledldquoIdentify signs of diabetic retinopathy in eye imagesrdquo(website httpswwwkagglecomcdiabetic-retinopathy-detection)

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e authors would like to thank Mr Chia-Ming Hu for hishelp in preparing computational instruments and per-forming parts of the experiments in this paper

References

[1] N H Cho J E Shaw S Karuranga et al ldquoIDF Diabetes Atlasglobal estimates of diabetes prevalence for 2017 and pro-jections for 2045rdquo Diabetes Research and Clinical Practicevol 138 pp 271ndash281 2018

[2] American Diabetes Association Eye Complications AmericanDiabetes Association Arlington VA USA 2013 httpwwwdiabetesorgliving-with-diabetescomplicationseye-complicationsreferrerhttpswwwgooglecomtw

[3] International Diabetes Foundation (IDF) Diabetes AtlasInternational Diabetes Federation Brussels Belgium 7thedition 2015

[4] C P Wilkinson F L Ferris R E Klein et al ldquoProposedinternational clinical diabetic retinopathy and diabetic

macular edema disease severity scalesrdquo Ophthalmologyvol 110 no 9 pp 1677ndash1682 2003

[5] A Tufail C Rudisill C Egan et al ldquoAutomated diabeticretinopathy image assessment software diagnostic accuracyand cost-effectiveness compared with human gradersrdquoOphthalmology vol 124 no 3 pp 343ndash351 2017

[6] A Tufail V V Kapetanakis S Salas-Vega et al ldquoAn ob-servational study to assess if automated diabetic retinopathyimage assessment software can replace one or more steps ofmanual imaging grading and to determine their cost-effec-tivenessrdquo Health Technology Assessment vol 20 no 92pp 1ndash72 2016

[7] I Liu and Y Sun ldquoRecursive tracking of vascular networks inangiograms based on the detection-deletion schemerdquo IEEETransactions on Medical Imaging vol 12 no 2 pp 334ndash3411993

[8] A Can H Shen J N Turner H L Tanenbaum andB Roysam ldquoRapid automated tracing and feature extractionfrom retinal fundus images using direct exploratory algo-rithmsrdquo IEEE Transactions on Information Technology inBiomedicine vol 3 no 2 pp 125ndash138 1999

[9] M Vlachos and E Dermatas ldquoMulti-scale retinal vesselsegmentation using line trackingrdquo Computerized MedicalImaging and Graphics vol 34 no 3 pp 213ndash227 2010

[10] Y Yin M Adel and S Bourennane ldquoAn automatic trackingmethod for retinal vascular tree extractionrdquo in Proceedings ofthe IEEE International Conference on Acoustics Speech andSignal Processing IEEE Kyoto Japan March 2012

[11] S Chaudhuri S Chatterjee N Katz M Nelson andM Goldbaum ldquoDetection of blood vessels in retinal imagesusing two-dimensional matched filtersrdquo IEEE Transactions onMedical Imaging vol 8 no 3 pp 263ndash269 1989

Table 8 Confusion matrix of the classification using the baseline method [39] and its overall accuracy

DeterminedOriginal

Random forest0 1 2 3 4

0 38228 2001 1331 38 641 546 1026 434 3 22 710 732 5726 635 3583 2 1 279 465 1464 45 2 89 73 636Accuracy () 9670 2727 7286 3830 5274Overall accuracy () 8601

Table 7 Differences between Tables 5 and 6

Determined lesionGround-truth lesion

SVM-network 1 SVM-network 20 1 2 3 4 0 1 2 3 4

0 +420 minus460 minus848 minus31 minus80 +310 0 minus104 minus43 +21 +194 +393 +164 +1 0 minus15 +96 +59 0 minus22 +231 +70 +648 minus37 +1 minus220 minus90 +197 +81 +483 +3 0 +45 +76 +21 +1 minus1 minus123 minus78 minus624 +3 minus3 +45 +76 +58 minus65 minus5 minus28 +4 +14Accuracy increment () minus108 8977 1280 2288 1052 078 1305 369 minus1603 23Overall accuracy increment () 16 116

Mobile Information Systems 13

the left side of Figure 9 and network 2 is the one on the rightside From Table 5 it shows that the lesion classications of 0and 2 are better than the other categories For lesion 1 mostof the prediction results are incorrect Also for lesions 3 and4 the majority of the results are misclassications that areclassied into lesion 2

Table 6 shows the confusion matrix of the classicationresults using the full procedure of the proposed method(using SVM with TLBO) Table 7 displays the dierencebetween Tables 5 and 6 which serves as a performancecomparison between the two methods (using DCNN onlyand DCNN + SVM + TLBO) From the table every classexcept class 0 and overall accuracy is increased in network 1For network 2 each accuracy except class 3 is increasede decline in accuracy of class 3 is mainly caused by

misclassication of class 2 Table 8 shows the confusionmatrix of the classication results using the baseline methodas reported in [39] for comparison purposes

62 Deep Learning vs Traditional Classication MethodsMany traditional classication methods try to solve theproblem of DR detection by (1) using image processing tocapture symptoms in fundus images and then (2) building aclassier to make decisions based on the detected symptoms(1) e shortcoming of image processing methods is that themanifestations of the symptoms are random across dierentimages therefore it is extremely time-consuming and re-quires intense eorts to label the locations of the symptomsAbiding by the new philosophy that comes with the emer-gence of the deep learning technology our proposed methodis trying to learn how to make decisions directly from theimage data itself Dierent than the former approaches ourimages only need to be labeled with lesion number instead oflabeling symptom locations Consequently it saves consid-erable time during the database preprocessing stage On top ofthe classication results by the two DCNN networks we useSVM optimized by TLBO to generate an improved outcomeand we achieve 8617 accuracy Our result is better than therst-place winner in the Kaggle competition It shows that ourresearch result is the state-of-the-art

63 Limitation In our current datasets the number ofimages of lesions 3 and 4 is not suiexclcient to train a networkwhich is a limitation of the proposed methoderefore oneof our future works is to develop deeper collaborative re-lations with hospitals and clinics to acquire more data oflesions 3 and 4 With more data we believe the classicationaccuracy will be further increased In addition from ourresult we found that it is hard to dierentiate the imagesbetween lesions 0 and 1 erefore when we collect newdata it is desirable to collect more images belonging tolesions 0 and 1 Also we can attempt to use a dierentnetwork architecture for this problem

7 Conclusion

It is feasible to train a deep learning model for automaticdiagnosis of DR as long as we have enough data for sta-tistical model training Furthermore the database prepa-ration stage only needs a categorical label for each trainingimage It does not require detailed annotation for retinalvessel tracking in every image Hence it is time-eiexclcientcompared to the traditional machine learning-based methodfor automatic diagnosis of DR e nal accuracy canachieve 8617 and 9105 for ve-class and binary classclassications respectively

e sensitivity and specicity of binary classication are08930 and 09089 respectively which is a satisfactory resultFurthermore we developed an automatic inspection appthat can be used in both personal examination and remotemedical care With more image data collected we expect theaccuracy can be even more enhanced further improving oursystem

Table 3 Detailed information of each dataset

Dataset of imagesDR lesions

0 1 2 3 4Train 34124 73 6 15 3 3Validation 1000 70 8 15 4 3Test 53572 74 7 15 2 2

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

3 times 3 conv 320

3 times 3 conv 352

3 times 3 conv 384

2 times 2 conv 416

1 times 1 conv 448

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

15 FMP

Softmax

5 times 5 conv 32

3 times 3 conv 64

3 times 3 conv 96

3 times 3 conv 128

3 times 3 conv 160

3 times 3 conv 192

3 times 3 conv 224

3 times 3 conv 256

3 times 3 conv 288

2 times 2 conv 320

1 times 1 conv 356

Softmax

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

18 FMP

Figure 9 Architecture of the two DCNN networks that we used

Mobile Information Systems 11

Table 4 Accuracy comparison when using different classifiers and different parameter optimization methods on each dataset

DatasetMethod

SVM RF [39]Without optimization Default parameter searching TLBO

Five-fold cross validation (using training data) 88744 965724 mdash mdashValidation data 854 796 mdash 847Test data 861177 810573 8617 8602

Table 5 Confusion matrix of the classification results from the two DCNN networks before performing SVM classification

Determined levelGround-truth level

Network 1 Network 20 1 2 3 4 0 1 2 3 4

0 39031 2693 2361 75 160 38310 2233 1617 51 781 118 438 185 0 1 327 735 290 1 32 339 626 5058 738 392 790 786 5509 620 3453 0 0 170 332 102 4 1 338 486 1854 43 5 85 69 551 100 7 105 56 595Accuracy () 9874 1164 6436 2735 4569 9691 1954 7001 4003 4934Overall accuracy () 8476 8518

(a) (b) (c) (d)

Figure 10 App interfaces

Table 6 Confusion matrix of the classification results using the full procedure of the proposed method (using SVM with TLBO)

Determined levelGround-truth level

SVM optimized with TLBO0 1 2 3 4

0 38611 2233 1513 44 801 312 831 349 1 12 570 696 5706 701 3933 3 0 215 408 1234 35 2 76 60 609Accuracy () 9767 2209 7260 3361 5050Overall accuracy () 8617

12 Mobile Information Systems

Data Availability

All experiment data come from a Kaggle contest calledldquoIdentify signs of diabetic retinopathy in eye imagesrdquo(website httpswwwkagglecomcdiabetic-retinopathy-detection)

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e authors would like to thank Mr Chia-Ming Hu for hishelp in preparing computational instruments and per-forming parts of the experiments in this paper

References

[1] N H Cho J E Shaw S Karuranga et al ldquoIDF Diabetes Atlasglobal estimates of diabetes prevalence for 2017 and pro-jections for 2045rdquo Diabetes Research and Clinical Practicevol 138 pp 271ndash281 2018

[2] American Diabetes Association Eye Complications AmericanDiabetes Association Arlington VA USA 2013 httpwwwdiabetesorgliving-with-diabetescomplicationseye-complicationsreferrerhttpswwwgooglecomtw

[3] International Diabetes Foundation (IDF) Diabetes AtlasInternational Diabetes Federation Brussels Belgium 7thedition 2015

[4] C P Wilkinson F L Ferris R E Klein et al ldquoProposedinternational clinical diabetic retinopathy and diabetic

macular edema disease severity scalesrdquo Ophthalmologyvol 110 no 9 pp 1677ndash1682 2003

[5] A Tufail C Rudisill C Egan et al ldquoAutomated diabeticretinopathy image assessment software diagnostic accuracyand cost-effectiveness compared with human gradersrdquoOphthalmology vol 124 no 3 pp 343ndash351 2017

[6] A Tufail V V Kapetanakis S Salas-Vega et al ldquoAn ob-servational study to assess if automated diabetic retinopathyimage assessment software can replace one or more steps ofmanual imaging grading and to determine their cost-effec-tivenessrdquo Health Technology Assessment vol 20 no 92pp 1ndash72 2016

[7] I Liu and Y Sun ldquoRecursive tracking of vascular networks inangiograms based on the detection-deletion schemerdquo IEEETransactions on Medical Imaging vol 12 no 2 pp 334ndash3411993

[8] A Can H Shen J N Turner H L Tanenbaum andB Roysam ldquoRapid automated tracing and feature extractionfrom retinal fundus images using direct exploratory algo-rithmsrdquo IEEE Transactions on Information Technology inBiomedicine vol 3 no 2 pp 125ndash138 1999

[9] M Vlachos and E Dermatas ldquoMulti-scale retinal vesselsegmentation using line trackingrdquo Computerized MedicalImaging and Graphics vol 34 no 3 pp 213ndash227 2010

[10] Y Yin M Adel and S Bourennane ldquoAn automatic trackingmethod for retinal vascular tree extractionrdquo in Proceedings ofthe IEEE International Conference on Acoustics Speech andSignal Processing IEEE Kyoto Japan March 2012

[11] S Chaudhuri S Chatterjee N Katz M Nelson andM Goldbaum ldquoDetection of blood vessels in retinal imagesusing two-dimensional matched filtersrdquo IEEE Transactions onMedical Imaging vol 8 no 3 pp 263ndash269 1989

Table 8 Confusion matrix of the classification using the baseline method [39] and its overall accuracy

DeterminedOriginal

Random forest0 1 2 3 4

0 38228 2001 1331 38 641 546 1026 434 3 22 710 732 5726 635 3583 2 1 279 465 1464 45 2 89 73 636Accuracy () 9670 2727 7286 3830 5274Overall accuracy () 8601

Table 7 Differences between Tables 5 and 6

Determined lesionGround-truth lesion

SVM-network 1 SVM-network 20 1 2 3 4 0 1 2 3 4

0 +420 minus460 minus848 minus31 minus80 +310 0 minus104 minus43 +21 +194 +393 +164 +1 0 minus15 +96 +59 0 minus22 +231 +70 +648 minus37 +1 minus220 minus90 +197 +81 +483 +3 0 +45 +76 +21 +1 minus1 minus123 minus78 minus624 +3 minus3 +45 +76 +58 minus65 minus5 minus28 +4 +14Accuracy increment () minus108 8977 1280 2288 1052 078 1305 369 minus1603 23Overall accuracy increment () 16 116

Mobile Information Systems 13

Table 4 Accuracy comparison when using different classifiers and different parameter optimization methods on each dataset

DatasetMethod

SVM RF [39]Without optimization Default parameter searching TLBO

Five-fold cross validation (using training data) 88744 965724 mdash mdashValidation data 854 796 mdash 847Test data 861177 810573 8617 8602

Table 5 Confusion matrix of the classification results from the two DCNN networks before performing SVM classification

Determined levelGround-truth level

Network 1 Network 20 1 2 3 4 0 1 2 3 4

0 39031 2693 2361 75 160 38310 2233 1617 51 781 118 438 185 0 1 327 735 290 1 32 339 626 5058 738 392 790 786 5509 620 3453 0 0 170 332 102 4 1 338 486 1854 43 5 85 69 551 100 7 105 56 595Accuracy () 9874 1164 6436 2735 4569 9691 1954 7001 4003 4934Overall accuracy () 8476 8518

(a) (b) (c) (d)

Figure 10 App interfaces

Table 6 Confusion matrix of the classification results using the full procedure of the proposed method (using SVM with TLBO)

Determined levelGround-truth level

SVM optimized with TLBO0 1 2 3 4

0 38611 2233 1513 44 801 312 831 349 1 12 570 696 5706 701 3933 3 0 215 408 1234 35 2 76 60 609Accuracy () 9767 2209 7260 3361 5050Overall accuracy () 8617

12 Mobile Information Systems

Data Availability

All experiment data come from a Kaggle contest calledldquoIdentify signs of diabetic retinopathy in eye imagesrdquo(website httpswwwkagglecomcdiabetic-retinopathy-detection)

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e authors would like to thank Mr Chia-Ming Hu for hishelp in preparing computational instruments and per-forming parts of the experiments in this paper

References

[1] N H Cho J E Shaw S Karuranga et al ldquoIDF Diabetes Atlasglobal estimates of diabetes prevalence for 2017 and pro-jections for 2045rdquo Diabetes Research and Clinical Practicevol 138 pp 271ndash281 2018

[2] American Diabetes Association Eye Complications AmericanDiabetes Association Arlington VA USA 2013 httpwwwdiabetesorgliving-with-diabetescomplicationseye-complicationsreferrerhttpswwwgooglecomtw

[3] International Diabetes Foundation (IDF) Diabetes AtlasInternational Diabetes Federation Brussels Belgium 7thedition 2015

[4] C P Wilkinson F L Ferris R E Klein et al ldquoProposedinternational clinical diabetic retinopathy and diabetic

macular edema disease severity scalesrdquo Ophthalmologyvol 110 no 9 pp 1677ndash1682 2003

[5] A Tufail C Rudisill C Egan et al ldquoAutomated diabeticretinopathy image assessment software diagnostic accuracyand cost-effectiveness compared with human gradersrdquoOphthalmology vol 124 no 3 pp 343ndash351 2017

[6] A Tufail V V Kapetanakis S Salas-Vega et al ldquoAn ob-servational study to assess if automated diabetic retinopathyimage assessment software can replace one or more steps ofmanual imaging grading and to determine their cost-effec-tivenessrdquo Health Technology Assessment vol 20 no 92pp 1ndash72 2016

[7] I Liu and Y Sun ldquoRecursive tracking of vascular networks inangiograms based on the detection-deletion schemerdquo IEEETransactions on Medical Imaging vol 12 no 2 pp 334ndash3411993

[8] A Can H Shen J N Turner H L Tanenbaum andB Roysam ldquoRapid automated tracing and feature extractionfrom retinal fundus images using direct exploratory algo-rithmsrdquo IEEE Transactions on Information Technology inBiomedicine vol 3 no 2 pp 125ndash138 1999

[9] M Vlachos and E Dermatas ldquoMulti-scale retinal vesselsegmentation using line trackingrdquo Computerized MedicalImaging and Graphics vol 34 no 3 pp 213ndash227 2010

[10] Y Yin M Adel and S Bourennane ldquoAn automatic trackingmethod for retinal vascular tree extractionrdquo in Proceedings ofthe IEEE International Conference on Acoustics Speech andSignal Processing IEEE Kyoto Japan March 2012

[11] S Chaudhuri S Chatterjee N Katz M Nelson andM Goldbaum ldquoDetection of blood vessels in retinal imagesusing two-dimensional matched filtersrdquo IEEE Transactions onMedical Imaging vol 8 no 3 pp 263ndash269 1989

Table 8 Confusion matrix of the classification using the baseline method [39] and its overall accuracy

DeterminedOriginal

Random forest0 1 2 3 4

0 38228 2001 1331 38 641 546 1026 434 3 22 710 732 5726 635 3583 2 1 279 465 1464 45 2 89 73 636Accuracy () 9670 2727 7286 3830 5274Overall accuracy () 8601

Table 7 Differences between Tables 5 and 6

Determined lesionGround-truth lesion

SVM-network 1 SVM-network 20 1 2 3 4 0 1 2 3 4

0 +420 minus460 minus848 minus31 minus80 +310 0 minus104 minus43 +21 +194 +393 +164 +1 0 minus15 +96 +59 0 minus22 +231 +70 +648 minus37 +1 minus220 minus90 +197 +81 +483 +3 0 +45 +76 +21 +1 minus1 minus123 minus78 minus624 +3 minus3 +45 +76 +58 minus65 minus5 minus28 +4 +14Accuracy increment () minus108 8977 1280 2288 1052 078 1305 369 minus1603 23Overall accuracy increment () 16 116

Mobile Information Systems 13

Data Availability

All experiment data come from a Kaggle contest calledldquoIdentify signs of diabetic retinopathy in eye imagesrdquo(website httpswwwkagglecomcdiabetic-retinopathy-detection)

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

e authors would like to thank Mr Chia-Ming Hu for hishelp in preparing computational instruments and per-forming parts of the experiments in this paper

References

[1] N H Cho J E Shaw S Karuranga et al ldquoIDF Diabetes Atlasglobal estimates of diabetes prevalence for 2017 and pro-jections for 2045rdquo Diabetes Research and Clinical Practicevol 138 pp 271ndash281 2018

[2] American Diabetes Association Eye Complications AmericanDiabetes Association Arlington VA USA 2013 httpwwwdiabetesorgliving-with-diabetescomplicationseye-complicationsreferrerhttpswwwgooglecomtw

[3] International Diabetes Foundation (IDF) Diabetes AtlasInternational Diabetes Federation Brussels Belgium 7thedition 2015

[4] C P Wilkinson F L Ferris R E Klein et al ldquoProposedinternational clinical diabetic retinopathy and diabetic

macular edema disease severity scalesrdquo Ophthalmologyvol 110 no 9 pp 1677ndash1682 2003

[5] A Tufail C Rudisill C Egan et al ldquoAutomated diabeticretinopathy image assessment software diagnostic accuracyand cost-effectiveness compared with human gradersrdquoOphthalmology vol 124 no 3 pp 343ndash351 2017

[6] A Tufail V V Kapetanakis S Salas-Vega et al ldquoAn ob-servational study to assess if automated diabetic retinopathyimage assessment software can replace one or more steps ofmanual imaging grading and to determine their cost-effec-tivenessrdquo Health Technology Assessment vol 20 no 92pp 1ndash72 2016

[7] I Liu and Y Sun ldquoRecursive tracking of vascular networks inangiograms based on the detection-deletion schemerdquo IEEETransactions on Medical Imaging vol 12 no 2 pp 334ndash3411993

[8] A Can H Shen J N Turner H L Tanenbaum andB Roysam ldquoRapid automated tracing and feature extractionfrom retinal fundus images using direct exploratory algo-rithmsrdquo IEEE Transactions on Information Technology inBiomedicine vol 3 no 2 pp 125ndash138 1999

[9] M Vlachos and E Dermatas ldquoMulti-scale retinal vesselsegmentation using line trackingrdquo Computerized MedicalImaging and Graphics vol 34 no 3 pp 213ndash227 2010

[10] Y Yin M Adel and S Bourennane ldquoAn automatic trackingmethod for retinal vascular tree extractionrdquo in Proceedings ofthe IEEE International Conference on Acoustics Speech andSignal Processing IEEE Kyoto Japan March 2012

[11] S Chaudhuri S Chatterjee N Katz M Nelson andM Goldbaum ldquoDetection of blood vessels in retinal imagesusing two-dimensional matched filtersrdquo IEEE Transactions onMedical Imaging vol 8 no 3 pp 263ndash269 1989

Table 8 Confusion matrix of the classification using the baseline method [39] and its overall accuracy

DeterminedOriginal

Random forest0 1 2 3 4

0 38228 2001 1331 38 641 546 1026 434 3 22 710 732 5726 635 3583 2 1 279 465 1464 45 2 89 73 636Accuracy () 9670 2727 7286 3830 5274Overall accuracy () 8601

Table 7 Differences between Tables 5 and 6

Determined lesionGround-truth lesion

SVM-network 1 SVM-network 20 1 2 3 4 0 1 2 3 4

0 +420 minus460 minus848 minus31 minus80 +310 0 minus104 minus43 +21 +194 +393 +164 +1 0 minus15 +96 +59 0 minus22 +231 +70 +648 minus37 +1 minus220 minus90 +197 +81 +483 +3 0 +45 +76 +21 +1 minus1 minus123 minus78 minus624 +3 minus3 +45 +76 +58 minus65 minus5 minus28 +4 +14Accuracy increment () minus108 8977 1280 2288 1052 078 1305 369 minus1603 23Overall accuracy increment () 16 116

Mobile Information Systems 13

[12] A Hoover V Kouznetsova and M Goldbaum ldquoLocatingblood vessels in retinal images by piecewise threshold probingof a matched filter responserdquo IEEE Transactions on MedicalImaging vol 19 no 3 pp 203ndash210 2000

[13] X Jiang and D Mojon ldquoAdaptive local thresholding byverification-based multithreshold probing with application tovessel detection in retinal imagesrdquo IEEE Transactions onPattern Analysis and Machine Intelligence vol 25 no 1pp 131ndash137 2003

[14] L Zhang Q Li J You and D Zhang ldquoA modified matchedfilter with double-sided thresholding for screening pro-liferative diabetic retinopathyrdquo IEEE Transactions on In-formation Technology in Biomedicine vol 13 no 4pp 528ndash534 2009

[15] Q Li J You and D Zhang ldquoVessel segmentation and widthestimation in retinal images using multiscale production ofmatched filter responsesrdquo Expert Systems with Applicationsvol 39 no 9 pp 7600ndash7610 2012

[16] F Zana and J C Klein ldquoSegmentation of vessel-like patternsusing mathematical morphology and curvature evaluationrdquoIEEE Transactions on Image Processing vol 10 no 7pp 1010ndash1019 2001

[17] G Ayala T Leon and V Zapater ldquoDifferent averages of afuzzy set with an application to vessel segmentationrdquo IEEETransactions on Fuzzy Systems vol 13 no 3 pp 384ndash3932005

[18] M S Miri and A Mahloojifar ldquoRetinal image analysis usingcurvelet transform and multistructure elements morphologyby reconstructionrdquo IEEE Transactions on Biomedical Engi-neering vol 58 no 5 pp 1183ndash1192 2011

[19] D Karthika and A Marimuthu ldquoRetinal image analysisusing contourlet transform and multistructure elementsmorphology by reconstructionrdquo in Proceedings of the WorldCongress on Computing and Communication TechnologiesIEEE Tiruchirappalli India 2014

[20] M Kass A Witkin and D Terzopoulos ldquoSnake activecontour modelsrdquo International Journal of Computer Visionvol 1 no 4 pp 321ndash331 1988

[21] L Espona M J Carreira M Ortega et al ldquoA snake for retinalvessel segmentationrdquo in Proceedings of the 3rd Iberian Con-ference on Pattern Recognition and Image Analysis Springer-erlag Berlin and Heidelberg GmbH amp Co KG Girona SpainJune 2007

[22] B Al-Diri and A Hunter ldquoA ribbon of twins for extractingvessel boundariesrdquo in Proceedings of the 3rd EuropeanMedicaland Biological Engineering Conference Prague Czech Re-public November 2005

[23] Y Zhang W Hsu and M L Lee ldquoDetection of retinal bloodvessels based on nonlinear projectionsrdquo Journal of SignalProcessing Systems vol 55 no 1ndash3 pp 103ndash112 2009

[24] Y Q Zhao X H Wang X F Wang and F Y Shih ldquoRetinalvessels segmentation based on level set and region growingrdquoPattern Recognition vol 47 no 7 pp 2437ndash2446 2014

[25] E Ricci and R Perfetti ldquoRetinal blood vessel segmentationusing line operators and support vector classificationrdquo IEEETransactions on Medical Imaging vol 26 no 10 pp 1357ndash1365 2007

[26] D Marin A Aquino M E Gegundez-Arias and J M BravoldquoA new supervised method for blood vessel segmentation inretinal images by using gray-level and moment invariants-based featuresrdquo IEEE Transactions on Medical Imagingvol 30 no 1 pp 146ndash158 2011

[27] V Shanmugam and R S D W Banu ldquoRetinal blood vesselsegmentation using an extreme learning machine approachrdquo

in Proceedings of the 2013 IEEE Point-of-Care HealthcareTechnologies IEEE Bangalore India January 2013

[28] S Wang Y Yin G Cao B Wei Y Zheng and G YangldquoHierarchical retinal blood vessel segmentation based onfeature and ensemble learningrdquo Neurocomputing vol 149pp 708ndash717 2015

[29] A Salazar-Gonzalez D Kaba Y Li and X Liu ldquoSegmen-tation of the blood vessels and optic disk in retinal imagesrdquoIEEE Journal of Biomedical and Health Informatics vol 18no 6 pp 1874ndash1886 2014

[30] R M Cesar Jr and H F Jelinek ldquoSegmentation of retinalfundus vasculature in nonmydriatic camera images usingwaveletsrdquo in Angiography and Plaque Imaging AdvancedSegmentation Techniques J S Suri and S Laxminarayan Edspp 193ndash224 CRC Press Boca Raton FL USA 2003

[31] J J G Leandro J V B Soares R M Cesar et al ldquoBloodvessels segmentation in nonmydriatic images using waveletsand statistical classifiersrdquo in Proceedings of the XVI BrazilianSymposium on Computer Graphics and Image ProcessingIEEE Satildeo Carlos Brazil October 2003

[32] Y A Tolias and S M Panas ldquoA fuzzy vessel tracking algo-rithm for retinal images based on fuzzy clusteringrdquo IEEETransactions on Medical Imaging vol 17 no 2 pp 263ndash2731998

[33] S Xie and H Nie ldquoRetinal vascular image segmentation usinggenetic algorithm plus FCM clusteringrdquo in Proceedings of the2013 Bird International Conference on Intelligent SystemDesign and Engineering Applications IEEE Hong KongChina January 2013

[34] B M Ege O K Hejlesen O V Larsen et al ldquoScreening fordiabetic retinopathy using computer based image analysis andstatistical classificationrdquo Computer Methods and Programs inBiomedicine vol 62 no 3 pp 165ndash175 2000

[35] N Silberman K Ahrlich R Fergus et al ldquoCase for automateddetection of diabetic retinopathyrdquo in Proceedings of the AAAISpring Symposium Artificial Intelligence for DevelopmentStanford CA USA March 2010

[36] A G Karegowda A Nasiha M A Jayaram andA S Manjunath ldquoExudates detection in retinal images usingback propagation neural networkrdquo International Journal ofComputer Applications vol 25 no 3 pp 25ndash31 2011

[37] S Kavitha and K Duraiswamy ldquoAutomatic detection of hardand soft exudates in fundus images using color histogramthresholdingrdquo European Journal of Scientific Research vol 48pp 493ndash504 2011

[38] J de la Calleja L Tecuapetla M A Medina et al ldquoLBP andmachine learning for diabetic retinopathy detectionrdquo inProceedings of the 2014 International Conference on In-telligent Data Engineering and Automated LearningSpringer Salamanca Spain September 2014

[39] B Graham ldquoFractional max-poolingrdquo 2015 httpsarxivorgpdf14126071v4pdf

[40] C-C Chang and C-J Lin ldquoLIBSVM a library for supportvector machinesrdquo ACM Transactions on Intelligent Systemsand Technology vol 2 no 3 p 27 2011

[41] R V Rao V J Savsani and D P Vakharia ldquoTeachingndashlearning-based optimization a novel method for constrainedmechanical design optimization problemsrdquo Computer-AidedDesign vol 43 no 3 pp 303ndash315 2011

[42] Kaggle contests Identify Signs of Diabetic Retinopathy in EyeImages Kaggle San Francisco CA USA 2015 httpswwwkagglecomcdiabetic-retinopathy-detection

14 Mobile Information Systems

Computer Games Technology

International Journal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Journal ofEngineeringVolume 2018

Advances in

FuzzySystems

Hindawiwwwhindawicom

Volume 2018

International Journal of

ReconfigurableComputing

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

thinspArtificial Intelligence

Hindawiwwwhindawicom Volumethinsp2018

Hindawiwwwhindawicom Volume 2018

Civil EngineeringAdvances in

Hindawiwwwhindawicom Volume 2018

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawiwwwhindawicom Volume 2018

Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

RoboticsJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Computational Intelligence and Neuroscience

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018

Human-ComputerInteraction

Advances in

Hindawiwwwhindawicom Volume 2018

Scientic Programming

Submit your manuscripts atwwwhindawicom