A Fast Enhancement/Thresholding Based Blood Vessel Segmentation for Retinal Image Using Contrast Limited Adaptive Histogram Equalization

Delivered by Publishing Technology to BENADICT RAJA JIP 1827315750 On Fri 05 Sep 2014 044059

Copyright American Scientific Publishers

Copyright copy 2014 American Scientific PublishersAll rights reservedPrinted in the United States of America

R E S E A R CH AR T I C L E

Journal of Medical Imaging andHealth InformaticsVol 4 567ndash575 2014

A Fast EnhancementThresholding Based BloodVessel Segmentation for Retinal Image Using

Contrast Limited Adaptive Histogram EqualizationC G Ravichandran1 and J Benadict Raja2lowast

1Excel Engineering College Komarapalayam 637303 Tamil Nadu India2Department of Computer Science and Engineering PSNA College of Engineering and Technology

Dindigul 624622 Tamil Nadu India

Automatic detection of blood vessel in retinal fundus image is an important task in the computer aided diagno-sis of ophthalmology This paper presents a fully automatic enhancementthresholding based vessel extractionmethod The input image is enhanced by histogram matching and Contrast Limited Adaptive Histogram Equal-ization (CLAHE) techniques Following CLAHE Wiener filtering is carried in order to remove the backgroundnoise A local entropy based thresholding technique is then used to extract blood vessel from the 2 dimen-sional Gabor filter response of CLAHErsquod image The performance of the proposed method was evaluated ontwo publicly available DRIVE and STARE databases and compared with the methods reported recently Theproposed method extracts blood vessels in a DRIVE image within 147 seconds(s) with (accuracy sensitivityspecificity)= 957472599799 and (952676939672) for the DRIVE and STARE databasesrespectively The average predictive value on both the databases were also higher (7356) compared to therecent methods

Keywords Retinal Image Segmentation Vessel Detection CLAHE Gabor Filter Entropy Thresholding

1 INTRODUCTIONBlood vessel extraction in retinal fundus (RF) image is impor-tant for an early diagnosis of various ophthalmologic and car-diovascular diseases such as diabetes hypertension chorodialneovascularization and arteriosclerosis16 The morphological fea-tures of retinal blood vessels like length diameter tortuosityand branching angle have diagnostic relevance and also used tomonitor the progression of diseases21 Vessel extraction is alsouseful for registration of retinal images of the same patient gath-ered from different sources11 Sometimes the exclusion of bloodvessel is useful for automatic detection of pathological elementslike microaneurysms or exudates10 Finally the vessel extractionis useful in determining the position of other features of retinasuch as fovea and optic disk2

The manual vessel segmentation can be time consumingtedious when the vascular network is complex or the imagequality is poor Therefore extensive research efforts have beendevoted to automating this process Computer aided vesselextraction is also difficult or time consuming task due to severalreasons the presence of noise pathological effects low contrastbetween vascular network and background variability in vesselwidth and shape10 Some of the existing methods can acquire

lowastAuthor to whom correspondence should be addressed

retinal blood vessels with high accuracy However these methodsare suffers from lengthy processing time

This paper presents a fast and effective algorithm that is capa-ble of extracting blood vessel almost in real time with muchhigher accuracy This paper brings the following contributions1 We introduce a new paradigm based on histogram techniquesfor retinal image enhancement The effects of bright and darkregions of RF image are suppressed by histogram matching Thenthe region based histogram equalization is used to enhance theimage without affecting the details of small vessels2 For vessel enhancement we apply 2D Gabor filter Thereduced filter size and single scale operation ensures the speedof segmentation3 Our method does not require training therefore it can be useddirectly on the image In addition it relies on adaptive thresh-olding so that no numerical parameter is tuned manually

We report experimental results on two publicly availableDRIVE and STARE databases and the obtained results are com-pared with the methods reported recently

The remainder of this paper is organized as follows Section 2reviews briefly the previous work on retinal vascular segmen-tation algorithm Section 3 details the proposed vessel extrac-tion process Experimental results are reported and compared inSections 4 and 5 concludes the paper

J Med Imaging Health Inf Vol 4 No 4 2014 2156-701820144567009 doi101166jmihi20141289 567

Delivered by Publishing Technology to BENADICT RAJA JIP 1827315750 On Fri 05 Sep 2014 044059

Copyright American Scientific Publishers

R E S E A R CH AR T I C L E J Med Imaging Health Inf 4 567ndash575 2014

2 RELATED WORKThe most recent studies222429 discuss the blood vessel segmen-tation algorithm in detail along different dimensions We havedivided the retinal vessel segmentation algorithms into five maincategories(1) tracking based(2) mathematical morphology(3) matched filter(4) model based thresholding and(5) supervised classification methods

They are briefly discussed belowThe tracking based methods select a set of reliable seed points

on the vessel network and track the vessels starting from theseeds The accuracy of tracking method is mainly based on theselection of seed points Nayebitar and Moghaddam6 localizesthe optic disk to determine the seed points and then trackingis done with particle filters A multi-scale line tracking methodis proposed by Vlachos and Dermatas19 they have used bright-ness selection rule to determine the seed points Yin et al35

have used a statistic sampling scheme to select a number ofseed points During the tracking process vessel edge points aredetected iteratively using local grey level statistics and vesselcontinuity properties

Fraz et al23 extracts vessel centerlines by using first orderderivative of a Gaussian filter Then the shape and orientationmap of blood vessel is obtained by applying a multi directionalmorphological top-hat operator with a linear structuring elementMiri and Mahloojitar27 have used curvelet transform for vesselenhancement then morphological operator using multi-structureelement is applied to enhance the edge ridges Further the falseedges are removed by morphological operator by reconstruction

In matched filter (MF) based method a 2 dimensional (2D)matched filter kernel is constructed and convolved with the reti-nal image in different orientation For each pixel the highestresponse of the filter is retained The MF response is foundeffective when used in conjunction with additional processingtechniques25 Chaudhuri et al31 proposed a 2D linear kernel witha Gaussian profile for vessel enhancement The kernel is rotatedin 12 different directions and convolved with the image Fathi andNaghsh-Nilchi1 enhances the blood vessel using complex contin-uous wavelet transform (CCWT) with different scales Then anadaptive histogram-based thresholding is used for vessel extrac-tion Zhang et al5 have used a pair of filter the zero-meanGaussian filter (MF) for vessel enhancement and the first-orderderivative of the Gaussian (FDOG) for adjusting the thresholdA high speed vessel extraction method using phase congruency isproposed by Amin and Yan20 Initially phase congruency of theimage is generated then banks of 2D log Gabor filter are usedfor measuring the phase congruency Ramlugun et al12 appliesa double sided thresholding scheme for extracting vessel fromthe 2D Gabor responses Li et al30 multiplies the MF responseat three scales then applies a simple thresholding for vesselextraction Matched filter with multi-wavelet kernel (MFMK) isused by Wang et al34 for vessel enhancement and then adaptivethresholding is used for vessel extraction

Martinez-Perez et al21 proposed a region growing based vesselextraction method The vessel features such as vessel width sizeand orientation using gradient magnitude and maximum principlecurvature of the hessian tenosr are extracted in multi-scale Lamet al7 combine different concavity measures such as line shape

concavity (to remove dark lesions) differentiable concavity (tohandle bright lesions) and locally normalized concavity to seg-ment the blood vessel network

In supervised classification methods the rule for vessel extrac-tion is learned by the algorithm on the basis of training setInitially the feature vectors are constructed for each pixel thenvessel classification is done by different classifier based on thefeature vectors Ricci and Perfetti10 extract the feature vector foreach pixel using two orthogonal line detectors along with the greylevel of the target pixel Then the pixels are classified as vessel ornon-vessel pixel using support vector machine (SVM) Lupascuet al8 introduces feature-based AdaBoost classifier (FABC) forvessel segmentation For each pixel a 41-D feature vector isconstructed at multiple scales Marin et al9 constructed a 7-Dfeature vector composed of gray-level and moment invariant-based features Then a neural network based classifier is usedfor vessel segmentation Akram et al26 proposed a multivariatem-mediods based classifier for vessel segmentation The inputimage is enhanced by 2D Gabor filter then necessary featuresare extracted for vessel segmentation Soares et al17 have used aGaussian mixer model classifier for vessel extraction The featurevector for each pixel is constructed from the pixel intensity andmulti-scale Gabor wavelet response A ridge based vessel seg-mentation is proposed by Staal et al14 The extracted ridge pixelswere grouped into patches Then the features were extracted foreach pixel on the patches Finally the k-Nearest Neighbor algo-rithm is applied for vessel extraction

3 PROPOSED VESSELSEGMENTATION METHOD

The blood vessel in the RF image is darker than the backgroundthe vessel thickness and illumination are reduced radically out-ward from the optic disk Also the whole image has some whiteGaussian noises therefore necessary preprocessing is requiredto enhance the quality of the image After the preprocessingblood vessels are enhanced using 2 dimensional Gabor filteringthen vessels are extracted using local entropy based thresholdingtechnique The flowchart of the proposed method is depicted inFigure 1

31 PreprocessingThe green channel of RF image is more suitable for vessel extrac-tion since the contrast between vessel and background is higherin green channel image therefore the grayscale image of greenchannel is used for the remaining part of this work In order toreduce the effect of exudates and optic disk in the vessel seg-mentation the histogram of gray scale image of green channelis matched to gray scale image of input RGB image Then thehistogram equalization (HE) is applied for enhancing the imagecontrast and to reduce the effect of non-uniform illuminationThe RF images has different brightness regions such as optic diskand macular region therefore the global based contrast enhance-ment techniques do not produce efficient results15 CLAHE is aregion based enhancement technique where the image is dividedinto N timesN tiles and HE is applied locally to each region Theclip limit ensures that the local contrast is limited and does notincrease to the maximum However result of HE increases back-ground in-homogeneity a region based Wiener filtering (3times3) isapplied for reducing impulse noise and smoothen the background

568

Delivered by Publishing Technology to BENADICT RAJA JIP 1827315750 On Fri 05 Sep 2014 044059

Copyright American Scientific Publishers

R E S E A R CH AR T I C L EJ Med Imaging Health Inf 4 567ndash575 2014

Fig 1 Flowchart of the proposed method

in-homogeneity Based on the experiment experience the tile sizefor CLAHE is set to 4times 4 and clip limit is set to 0012 Theobtained results of preprocessing stage are shown in Figure 2

32 Vessel EnhancementAfter the preprocessing phase a 2D Gabor matched filter isapplied to emphasize the blood vessel in the preprocessed imageThe blood vessels are dark piecewise connected locally linearand have Gaussian shape cross-section profile31 therefore thevessels can be mathematically correlated to a Gabor responseThe multiscale Gabor filter kernel is expressed as

g sx y= expminus

[x2p

sx2+ y2p

sy2

]cos2fxp (1)

where

xp = x cos+y sin yp =minusx sin +y cos (2)

mdashfilter direction smdashscale mdashstandard deviation of Gaussianfmdashthe frequency of cosine wave

In order to produce a single peak response the Gabor kernel isrotated in 12 different directions at 15 degree each and convolved

to the image For each pixel the maximum response of the filter isretained The key parameters in our experiments are as followsx = 9 y = 7 f = 01 and s = 085 these parameters werechosen based on our experiment experience The vessel enhancedimage is shown in Figure 4(a)

33 Vessel ExtractionOnce the vessels are enhanced by Gabor filtering a local entropybased thresholding scheme described by Pal and Pal28 is appliedto extract the blood vessel The procedure for threshold calcula-tion is as follows The Co-occurrence matrix of an image I is anLtimesL square matrix represented by M = tij LtimesL whose elementsare specified by the numbers of transition between all pairs ofgray level in G= 012 Lminus1 Thus tij is defined as

tij =Lsum

x=1

Lsumy=1

(3)

where

= 1 if

⎧⎪⎨⎪⎩Ix y= i and Ix y+1= j

or

Ix y= i and Ix+1 y= j

= 0 otherwise

Let t be a value used to threshold an image which partitioningthe gray level G into two regions G0 = 012 t13 and G1 =t t+ 1 t+ 2 Lminus 113 Also it partitions the matrix M intofour quadrants (see Fig 3) The gray level above t is assumedto be the foreground pixel and the others are background pixelAmong the four quadrants BB and FF are local quadrants thatare used for threshold calculation The probabilities of the twoquadrants are obtained by the following formulae

PBBij = Pij

PBB(4)

P FFij = Pij

PFF(5)

Where

Pij =tijsum

i

sumj tij

(6)

PBB =tsum

i=0

tsumj=0

Pij (7)

PFF =Lminus1sumi=t+1

Lminus1sumj=t+1

Pij (8)

Then the second-order entropy of the quadrants is computedfrom

H2BBt=minus1

2

tsumi=0

tsumj=0

PBBij log2 P

BBij (9)

H2FF t=minus1

2

Lminus1sumi=t+1

Lminus1sumj=t+1

P FFij log2 P

FFij (10)

Using the Eqs (9) and (10) local entropy is defined as

H2T t= H2

BBt+H2FF t (11)

569

Delivered by Publishing Technology to BENADICT RAJA JIP 1827315750 On Fri 05 Sep 2014 044059

Copyright American Scientific Publishers

R E S E A R CH AR T I C L E J Med Imaging Health Inf 4 567ndash575 2014

Fig 2 Obtained results of preprocessing phase (a) Input RF image (01_testtif of DRIVE) (b) Grayscale of RGB image (c) Grayscale of green channelimage (d) Histogram matched image (image (c) is matched to image (b)) (e) CLAHErsquod image (f) Wiener filtered CLAHErsquod image

The optimal local threshold tLE is the one that maximizes localentropy

tLE = argmaxH2T t13 (12)

This value tLE is used to segment the image The vessel extractedimage is shown in Figure 4(b)

34 Post ProcessingAs seen in Figure 4(b) there are some isolated pixels in thevessel extracted image that should be removed for a clean andcomplete result In order to remove the isolated pixels we uselength filtering This technique identifies individual objects byusing the eight-connected neighborhood and label propagationOnce the algorithm is completed the objects are classified asvessel objects and non-vessel objects based on the total count of

BB BF

FB FF

t

t

0 L-1

L-1

whereBBmdashBackground to BackgroundFFmdashForeground to ForegroundBFmdashBackground to Foreground FBmdashForeground to Background

Fig 3 Quadrants of co-occurrence matrix

the pixels in the objects In this work the objects having lesserthan 250 pixels are considered as non-vessel objects that areremoved from the resulting image (see Fig 4(c))

4 RESULTS AND DISCUSSION41 MaterialThe performance of the proposed method is evaluated on pub-licly available DRIVE (Digital Retinal Images for Vessel Extrac-tion) and STARE (STructural Analysis of REtina) databases TheDRIVE database contains 20 training and 20 test retinal images

(a)

(c)

(b)

Fig 4 Illustration of vessel extraction process (a) Vessel enhanced image(Gabor filter response of preprocessed image) (b) Vessel extracted image(c) The response of length filtering

570

Delivered by Publishing Technology to BENADICT RAJA JIP 1827315750 On Fri 05 Sep 2014 044059

Copyright American Scientific Publishers

R E S E A R CH AR T I C L EJ Med Imaging Health Inf 4 567ndash575 2014

Table I Obtained results of all images on the both DRIVE and STARE databases

DRIVE STARE

Execution Accuracy Execution AccuracyTest image time in s Predictive Sensitivity Specificity y in Test image time in s Predictive Sensitivity Specificity y in

1 216 07507 08058 09738 9588 1 288 06214 07399 09609 94332 139 08081 07794 09789 9585 2 188 05707 06423 09655 94403 145 08506 06565 09872 9543 3 184 03639 08433 09062 90244 141 08353 07046 09859 9600 4 178 08992 04160 09963 95335 156 08554 06882 09880 9599 5 189 05965 08328 09440 93406 142 08833 06123 09913 9544 6 188 06424 08702 09637 95727 142 07615 07268 09771 9542 7 180 06108 08961 09503 94598 144 07333 06611 09774 9502 8 181 06227 08997 09560 95189 142 08320 06610 09882 9617 9 184 06678 08632 09633 955510 147 07846 07092 09825 9600 10 188 05690 08486 09438 936111 139 07242 07327 09726 9511 11 181 06332 08502 09622 954212 144 07602 07266 09783 9566 12 183 07351 08821 09734 966313 142 08115 06773 09830 9531 13 178 07396 08106 09721 957714 142 07028 07785 09710 9555 14 183 07706 08104 09760 961015 145 05952 08021 09579 9468 15 181 07415 07743 09745 957216 144 07960 07455 09810 9598 16 180 08493 06517 09869 952617 141 08105 06741 09855 9592 17 184 07549 08414 09731 961418 144 07457 08055 09764 9628 18 184 08615 07007 09940 979219 138 07994 08033 09818 9670 19 184 07923 06101 09928 976320 147 07505 07672 09798 9641 20 189 07903 06021 09886 9628Averge 147 07795 07259 09799 9574 189 06916 07693 09672 9526

obtained from a diabetic retinopathy (DR) screening program inNetherlands Among the 40 images 7 are pathological imageshaving signs of early diabetic retinopathy remaining images donot show any sign for DR The images are acquired using CanonCR5 non-mydriatic 3CCD camera with a 45 field of view Thesize of each image is 768times 584 pixels with 8 bits per colorplane The images are manually segmented by well trained threeobservers The performance of our algorithm is measured on the20 test set images with the segmentations of the first observer asground truth the result of the second observer is used as benchmark for evaluation

The STARE database collected by Hoover et al2 consist of10 pathological and 10 non-pathological images The images are

Table II Performance results compared to recent methods on the both DRIVE and STARE databases

DRIVE STARE DRIVE+STARE

Type of Accuracy Accuracy Accuracymethod Method Sensitivity Specificity in Sensitivity Specificity in in

Second observer 07761 09725 9473 08949 09390 9354 9414Supervised Marin et al9 07068 09801 9452 06944 09819 9526 9489Supervised Ricci and Perfetti10 ndash ndash 9595 ndash ndash 9646 9621Supervised Staal et al14 07192 09773 9441 06970 09810 9516 9479Supervised Soares et al17 07283 09788 9466 07165 09748 9480 9473Supervised You et al33 07410 09751 9434 07260 09756 9497 9466Unsupervised Fathi and Naghsh-Nilchi1 07768 09759 9581 08061 09717 9591 9586Unsupervised Condurache and Mertins4 09086 09580 9509 08907 09654 9574 9542Unsupervised Mendonca and Campilho3 07344 09764 9452 06996 09730 9440 9446Unsupervised Zhang et al5 07120 09724 9382 07177 09753 9484 9433Unsupervised Lam et al7 ndash ndash 9472 ndash ndash 9567 9520Unsupervised Ramlugun et al12 06414 09767 9311 ndash ndash ndash ndashUnsupervised Amin and Yan20 ndash ndash 9191 07261 09807 9199 9195Unsupervised Martinez-Perez et al21 07246 09655 9344 07506 09569 9410 9377Unsupervised Fraz et al23 07152 09759 9430 07311 09680 9442 9436Unsupervised Akram et al26 ndash ndash 9460 ndash ndash 9500 9480Unsupervised Chaudhuri et al31 06168 09741 9284 06134 09755 9384 9334Unsupervised Nguyen et al32 ndash ndash 9407 ndash ndash 9324 9366Unsupervised Wang et al34 ndash ndash 9461 ndash ndash 9501 9481Unsupervised Proposed method 07259 09799 9574 07693 09672 9526 9550

captured with TopCon TRV-50 fundus camera at 35 field ofview The size of each image is 700times605 pixels with 8 bits percolor plane The images are manually segmented by two differentobservers the segmentation of the first observer is used as groundtruth for our evaluation

42 Performance MeasuresSensitivity specificity predictive accuracy and execution timeare the quantitative measures used for evaluating this method themeasures are defined as follows

Sensitivity= TPsTPs+FNs

(13)

571

Delivered by Publishing Technology to BENADICT RAJA JIP 1827315750 On Fri 05 Sep 2014 044059

Copyright American Scientific Publishers

R E S E A R CH AR T I C L E J Med Imaging Health Inf 4 567ndash575 2014

Fig 5 The obtained results on image 9 of the DRIVE database (a) Original image (b) Ground truth image (c) Niemeijer et al18 (d) Martinez-Perez et al21

(e) Staal et al14 (f) Zana and Klein11 (g) Fathi and Naghsh-Nilchi1 (h) Proposed method

Specificity= TNsTNs+FPs

(14)

Predictive= TPsTPs+FPs

(15)

Accuracy= TPs+TNsTPs+TNs+FPs+FNs

(16)

whereTPsmdashthe total number of vessel points correctly detected as

vessel pointsTNsmdashthe total number of non-vessel points correctly detected

as non-vessel pointsFPsmdashthe total number of non-vessel points detected as vessel

points andFNsmdashthe total number of vessel points not detected as vessel

points

The execution time includes mask generation and Gabor kernelconstruction time

In addition the performance of the method is also measuredwith receiver operating characteristic curve (ROC) A ROC curveis a plot of true positive fractions (sensitivity) versus false posi-tive fractions (1-specificity) The closer the curve approaches thetop left corner the better the performance of the method

43 Results on DRIVE and STARE DatabasesAs pointed out in the previous section the experiment was per-formed for the 20 test images of DRIVE and 20 test images ofSTARE databases The method was implemented on 306 GHzcore i3 processor with 317 GB RAM using MATLAB 76 Theobtained results for each image of these databases are presentedin Table I The proposed method shows average accuracy sen-sitivity and specificity respectively higher 955 7476 and9736 on the both databases The average predictive value of itis also higher than 7356

44 DiscussionsTo emphasis the capability of the proposed method the quali-tative and quantitative results of the method are compared withseveral state-of-the-art methods The quantitative results are com-pared in terms of sensitivity specificity accuracy and execu-tion time For this purpose the methods proposed by Fathi andNaghsh-Nilchi1 Hoover et al2 Mendonca and Campilho3 Con-durache and Mertins4 Zhang et al5 Lam et al7 Marin et al9

Ricci and Perfetti10 Zana and Klein11 Ramlugun et al12 Staalet al14 Soares et al17 Niemeijer et al18 Martinez-Perez et al21

Amin and Yan20 Fraz et al23 Akram et al26 Chaudhuri et al31

You et al33 Nguyen et al32 Wang et al34 are used The resultsof these methods are obtained from their original papers

The performance of the proposed method on DRIVE databaseis compared in left half of Table II The accuracy of the pro-posed method on DRIVE database is much higher (9574) thanother methods also the sensitivity and specificity values are com-paratively better than other methods For better comparison thequalitative result on the image 9 of the DRIVE database of theproposed method and some state-of-the-art methods are shownin Figure 5

Since Ramlugun et al12 also proposed a wavelet based methodfor vessel segmentation in retinal images The comparison

Table III Qualitative results compared to other CLAHE based method

572

Delivered by Publishing Technology to BENADICT RAJA JIP 1827315750 On Fri 05 Sep 2014 044059

Copyright American Scientific Publishers

R E S E A R CH AR T I C L EJ Med Imaging Health Inf 4 567ndash575 2014

Table IV Execution time of different methods on the DRIVE database

Method Configuration of computer Actual execution time Execution time normalized to 306 GHz (s)

Nguyen et al32 Intel duo core 24 GHz and 2 GB RAM 25 s 108Proposed method Intel core i3 306 GHz and 317 GB RAM 147 s 147Ramlugun et al12 166 GHz 20 s 526Amin and Yan20 266 GHz 10 s 556Marin et al9 Intel Core2Duo 213 GHz 15 min 3147Soares et al17 AMD Athlon 22 GHz 3 min 10 s 6985Fathi and Naghsh-Nilchi1 Pentium IV PC 32 GHz 1 min 8333Staal et al14 Intel pentium 100 GHz 15 min 17578Lam et al7 Dua CPU 183 GHz 13 min 22543Mendonca and Campilho3 Pentium-IV PC 32 GHz 3 min 25000

Fig 6 The obtained results on image 6 of the STARE database (a) Original image (b) Ground truth image (c) Hoover et al2 (d) You et al33 (d) Conduracheand Mertins4 (f) Fraz et al23 (g) Fathi and Naghsh-Nilchi1 (h) Proposed method

between these two methods shows the ability of the proposedmethod (see Table III) They used CLAHE based enhancementwith different clip limits and two separate processes for extract-ing small and large vessels And also the threshold value forvessel segmentation is set manually for each image Therefore

Fig 7 The ROC curve obtained by the proposed method (a) on the DRIVE database (b) on the STARE database

the execution time of their method is comparatively high (seeTable IV) and the sensitivity is about 64 and 85 lower thanour method They have obtained 931 average accuracy onDRIVE database which is 264 lower than this method andspecificity of their method is also less than this work

573

Delivered by Publishing Technology to BENADICT RAJA JIP 1827315750 On Fri 05 Sep 2014 044059

Copyright American Scientific Publishers

R E S E A R CH AR T I C L E J Med Imaging Health Inf 4 567ndash575 2014

Fig 8 The execution time and obtained accuracy of different method on the DRIVE database

The proposed method also was compared on STARE databasewith recent methods The comparison is presented in theright half of Table II The STARE consists of 10 patholog-ical and 10 non-pathological images The proposed methodprovides much higher accuracy (9629) on non-pathologicalimages (image 11ndash20) while it is slightly lowered (9424)with pathological images (image 1ndash10) For better comparisonthe obtained result of the proposed method on a pathologi-cal image (image 6) of the STARE database is also shown inFigure 6

The ROC curve is also used as a measure in evaluating theperformance of the proposed method The plot of true posi-tive rate (sensitivity) versus false positive rates (1-specificity) ofeach image on the both DRIVE and STARE databases alongwith ROC curve is shown in Figure 7 The proposed methodyields better curves on both databases The curve on the DRIVEdatabase is just over the plot of 2nd observer The sensitivity ofthe proposed method is slightly lower than the 2nd observer onSTARE database therefore the curve on STARE is slightly underthe plot of 2nd observer

From the obtained results the average accuracy and sensitiv-ity of the proposed method is much higher than other recentlypublished methods except Ricci and Perfetti10 method and Fathiand Naghsh-Nilchi1 method

The performance of the Ricci and Perfetti10 method is slightlyhigher than this work The performance of their method heav-ily depends on training samples therefore it is necessary toretrain their classifier on a new database And also their accuracystrongly decreases (9595 to 9266 on DRIVE and 9646 to9452 on STARE database) when training and test samples areindependent The accuracy of Fathi and Naghsh-Nilchi1 is alsoslightly (03) higher than this work but the computation costof their method is much higher The proposed method is about56 times faster than their method (see Fig 8)

Finally the execution time of the proposed method is com-pared with recently published methods (see Table IV) Sincethe methods are implemented on different configuration of PCsthe execution time of those methods should be normalized forcomparison Amin and Yan20 assumed that a 100 increase inmachine speed will increase execution-speed by 200 based

on this assumption the actual processing time of various meth-ods are normalized as follows The proposed method is imple-mented on 306 GHz machine which is 306minus1times100= 206increase in machine speed and 306minus1times200= 412 increasein execution-speed than the machine (10 GHz) that Staal et al14

used Their method took 15 min= 900 s on 10 GHz machine andbased on the assumption it takes 9001+ 412100 = 17578s on a 306 GHz machine Similarly the actual execution timeof other methods is normalized and given in last column ofTable IV

The recently published Nguyen et al33 method is slightly fasterthan this work but the accuracy of their method is comparativelylow (see Fig 8)

5 CONCLUSIONIn this paper a simple fast enhancementthresholding based ves-sel segmentation algorithm is proposed The input RF image isenhanced by histogram based techniques followed by 2D Gaborfiltering The histogram of gray scale of the green channel ismatched to gray scale of input RGB This matching highly sup-presses the effect of bright regions of the fundus image Furtherthe image is enhanced by the region based contrast enhancementtechnique Then the blood vessel network is enhanced by thefine tuned 2D Gabor filtering and extracted by entropy basedthresholding

The proposed method detects blood vessels in retinal imagealmost in real time however most of the existing methods arecostly in terms of execution time Compared to other vesselextraction methods this method is fully automatic and extractsthe vessel network of one DRIVE image within 147 s withoutoptimization of its Matlab code Therefore this method can beeasily scaled to deal with high resolution retinal images with anacceptable time The obtained average sensitivity and specificityof the vessel extraction on both the DRIVE and STARE databasesare 7476 and 9735 respectively Also the average accuracyand predictive value of it are 955 and 7356 respectivelyThe demonstrated performance effectiveness with its simplic-ity makes this method a suitable tool for real time blood vesselanalysis

574

Delivered by Publishing Technology to BENADICT RAJA JIP 1827315750 On Fri 05 Sep 2014 044059

Copyright American Scientific Publishers

R E S E A R CH AR T I C L EJ Med Imaging Health Inf 4 567ndash575 2014

References and Notes1 A Fathi and A R Naghsh-Nilchi Automatic wavelet-based retinal blood ves-

sels segmentation and vessel diameter estimation Biomedical Signal Pro-cessing and Control 8 71 (2013)

2 A Hoover V Kouznetsova and M Goldbaum Locating blood vessels in reti-nal images by piecewise threshold probing of a matched filter response IEEETransactions on Medical Imaging 19 203 (2000)

3 A M Mendonca and A Campilho Segmentation of retinal blood vesselsby combining the detection of centerlines and morphological reconstructionIEEE Transactions on Medical Imaging 25 1200 (2006)

4 A P Condurache and A Mertins Segmentation of retinal vessels with hys-teresis binary-classification paradigm Computerized Medical Imaging andGraphics 36 325 (2012)

5 B Zhang L Zhang L Zhang and F Karray Retinal vessel extraction bymatched filter with first-order derivative of Gaussian Computers in Biologyand Medicine 40 438 (2010)

6 B Nayebitar and H A Moghaddam A novel method for retinal vessel trackingusing particle filters Computers in Biology and Medicine 43 541 (2013)

7 B S Y Lam Y Gao and A W C Liew General retinal vessel segmenta-tion using regularization-based multiconcavity modeling IEEE Transactionson Medical Imaging 29 1369 (2010)

8 C A Lupascu D Tegolo and E Trucco FABC Retinal vessel segmentationusing AdaBoost IEEE Transactions on Information Technology in Biomedicine14 1267 (2010)

9 D Marin A Aquino M E Gegundez-Arias and J M Bravo A new super-vised method for blood vessel segmentation in retinal images by using gray-level and moment invariants-based features IEEE Transactions on MedicalImaging 30 146 (2011)

10 E Ricci and R Perfetti Retinal blood vessel segmentation using line operatorand support vector classification IEEE Transactions on Medical Imaging 261357 (2007)

11 F Zana and J C Klein A multimodal registration algorithm of eye fundusimages using vessels detection and Hough transform IEEE Transactions onMedical Imaging 18 419 (1999)

12 G S Ramlugun V K Nagarajan and C Chakraborty Small retinal vesselsextraction towards proliferative diabetic retinopathy screening Expert Sys-tems with Applications 39 1141 (2012)

13 httpwww isi uunlResearchDatabasesDRIVEresultsphp14 J Staal M D Abramoff M Niemeijer M A Viergever and B V Ginneken

Ridge based vessel segmentation in color images of the retina IEEE Trans-actions on Medical Imaging 23 501 (2004)

15 J A Stark Adaptive image contrast enhancement using generalization of his-togram equalization IEEE Transactions on Image Processing 9 889 (2000)

16 J J Kanski Clinical Ophthalmology 6th edn Elsevier Health SciencesLondon UK (2007)

17 J V B Soares J J G Leandro R M Cesar H F Jelinek and M J GreeRetinal vessel segmentation using 2-D Gabor wavelet and supervised classi-fication IEEE Transactions on Medical Imaging 25 1214 (2006)

18 M Niemeijer J J Staal V Ginneken M Loog and M D Abramoff Compar-ative study of retinal vessel segmentation methods on a new publicly availabledatabase SPIE Medical Imaging 5370 648 (2004)

19 M Vlachos and E Dermatas Multi-scale retinal vessel segmentationusing line tracking Computerized Medical Imaging and Graphics 34 213(2010)

20 M A Amin and H Yan High speed detection of retinal blood vessels in fundusimage using phase congruency Soft ComputingmdashA Fusion of FoundationsMethodologies and Applications 15 1217 (2011)

21 M E Martinez-Perez A D Hughes S A Thom A A Bharath and K HParker Segmentation of blood vessels from red-free and fluorescein retinalimages Medical Image Analysis 11 47 (2007)

22 M M Fraz P Remagnio A Hoppe B Uyyanonvara and A R RudnickaBlood vessel segmentation methodologies in retinal imagesmdashA survey Com-puter Methods and Programs in Biomedicine 108 407 (2012)

23 M M Fraz S A Barman P Remagnino A Hoppe A Basit B UyyanonvaraA R Rudnicka and C G Owen An approach to localize the retinal bloodvessels using bit planes and centerline detection Computer Methods andPrograms in Biomedicine 108 600 (2012)

24 M R K Mookiah U Rajendra Acharya C K Chua C M Lim E Y K Ngand A Laude Computer-aided diagnosis of diabetic retinopathy A reviewComputers in Biology and Medicine 43 2136 (2013)

25 M R K Mookiah U R Acharya R J Martis C K Chua L C Min E Y KNg and A Laude Evolutionary algorithm based classifier parameter tuning forautomatic diabetic retinopathy grading A hybrid feature extraction approachKnowledge Based System 39 9 (2013)

26 M U Akram S Khalid A Tariq and M Y Javed Detection of neovascular-ization in retinal images using multivariate m-Mediods based classifier Com-puterized Medical Imaging and Graphics 37 346 (2013)

27 M S Miri and A Mahloojitar Retinal image analysis using curvelet transformand multistructure elements morphology by reconstruction IEEE Transactionson Biomedical Engineering 58 1183 (2011)

28 N R Pal and S K Pal Entropic thresholding Signal Processing 16 97(1989)

29 O Faust R Acharya U E V K Ng K-H Ng and J S Suri Algorithms forthe automated detection of diabetic retinopathy using digital fundus imagesA review J Med Syst 36 145 (2012)

30 Q Li J You and D Zhang Vessel segmentation and width estimation inretinal images using multiscale production of matched filter responses ExpertSystems with Applications 39 7600 (2012)

31 S Chaudhuri S Chatterjee N Katz M Nelson and M Goldbaum Detectionof blood vessels in retinal images using two-dimensional matched filters IEEETransactions on Medical Imaging 8 263 (1989)

32 U T V Nguyen A Bhuiyan L A F Park and K Ramamohanarao An effec-tive retinal blood vessel segmentation method using multi-scale line detectionPattern Recognition 46 703 (2013)

33 X You Q Peng Y Yuan Y-M Cheung and J Lei Segmentation of reti-nal blood vessels using the radial projection and semi-supervised approachPattern Recognition 44 2314 (2011)

34 Y Wang G Ji P Lin and E Trucco Retinal vessel segmentation using mul-tiwavelet kernels and multiscale hierarchical decomposition Pattern Recogni-tion 46 2117 (2013)

35 Y Yin M Adel and S Bourennane Retinal vessel segmentation using aprobabilistic tracking method Pattern Recognition 45 1235 (2012)

Received 27 December 2013 Accepted 10 February 2014

575