automated detection of red lesions using superpixel...

Research ArticleAutomated Detection of Red Lesions Using SuperpixelMultichannel Multifeature

Wei Zhou12 ChengdongWu12 Dali Chen1 ZhenzhuWang1 Yugen Yi3 andWenyou Du1

1College of Information Science and Engineering Northeastern University Shenyang Liaoning 110004 China2Faculty of Robot Science and Engineering Northeastern University Shenyang Liaoning 110004 China3School of Software Jiangxi Normal University Nanchang Jiangxi 330022 China

Correspondence should be addressed to Yugen Yi yiyg510gmailcom

Received 19 November 2016 Accepted 27 March 2017 Published 23 April 2017

Academic Editor Alessandro Arsie

Copyright copy 2017 Wei Zhou et alThis is an open access article distributed under theCreativeCommonsAttribution License whichpermits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Red lesions can be regarded as one of the earliest lesions in diabetic retinopathy (DR) and automatic detection of red lesions playsa critical role in diabetic retinopathy diagnosis In this paper a novel superpixel Multichannel Multifeature (MCMF) classificationapproach is proposed for red lesion detection In this paper firstly a new candidate extraction method based on superpixel isproposed Then these candidates are characterized by multichannel features as well as the contextual feature Next FDA classifieris introduced to classify the red lesions among the candidates Finally a postprocessing technique based on multiscale bloodvessels detection is modified for removing nonlesions appearing as red Experiments on publicly available DiaretDB1 databaseare conducted to verify the effectiveness of our proposed method

1 Introduction

Diabetic retinopathy (DR) is one of the most serious per-formances of diabetic and the majority of people sufferingfrom diabetes mellitus for more than ten years will eventuallydevelop DR [1] Moreover with the development of thedisease it will cause vision loss [2] Regular follow-up hasbeen shown to help patients delay the progression of blind-ness and visual loss Digital color fundus photography ownsits low-cost and patient friendliness which is a prerequisitefor large scale screening [3] However in DR screeningprogram the limited number of specialists cannot keep upwith the rapid increasing number of DR patients Under thiscircumstance developing an automatic detection techniquebased on fundus images becomes extremely essential andurgent [4]

There are several different components in retinal images(see Figure 1) such as blood vessels fovea macula andoptic disc In clinical ophthalmologists classify DR into twoprimarily phases namely nonproliferative DR (NPDR) andproliferative DR (PDR) [5] NPDR can be regarded as theinitial phase of DR During this phase the blood vesselsbecome thin and leak fluid onto the retina [5] Several lesions

such as red lesions (microaneurysms and hemorrhages seeFigure 1) yellowish or bright spots (hard and soft exudatessee Figure 1) and interretinal microvascular abnormalities(IRMA) [6] can be produced during the phase of NPDRThesecond phase of DR is the PDR because the blood vesselscannot obtain enough oxygen causing the blood vessels togrow in different regions of the retina images for maintainingthe adequate oxygen Naturally these blood vessels are proneto leakage which may lead to vision lost In this paperwe mainly focus on the detection of red lesions containingmicroaneurysms and hemorrhages which are the earliestlesions and more complicated compared to other kinds oflesions detection in the DR

Numerous approaches have been proposed for red lesiondetection Among them the earliest paper based on MAdetection was proposed by Baudoin et al [7] which utilizedmathematical morphology approach to detect the microa-neurysms in fluorescein angiography images of the fundusAfter that two variants of morphological top-hat trans-formation methods for detecting MAs within fluoresceinangiograms were developed by Spencer et al [8] and Frameet al [9] Although the fluorescein angiograms can improvethe contrast between the fundus and their background the

HindawiComputational and Mathematical Methods in MedicineVolume 2017 Article ID 9854825 13 pageshttpsdoiorg10115520179854825

2 Computational and Mathematical Methods in Medicine

A optic discB foveaC maculaD blood vessels

E hemorrhageF microaneurysmG hard exudates

E

D

D

A BC

DF

G

Figure 1 A retinal image with different types of lesions and mainanatomical features

usage of intravenous contrast agents is not applicable foreveryone especially the pregnant woman [10] Thereforefluorescein angiograms cannot be widely used in public DRscreening programs and the adaption of digital color fundusphotographs is a better choice for screening purpose

A number of algorithms based on filtering have beenproposed for red lesion detection in digital color fundusphotographs Quellec et al [11] adopted a template matchingmethod based on subbands of wavelet transformed imagesfor microaneurysm detection which can solve the unevenillumination or high-frequency noise problems BesidesZhang et al [12] employed Multiscale Gaussian CorrelationCoefficients (MSCF) by computing the correlation coefficientbetween the grayscale distribution of microaneurysms andthe Gaussian function to detect MAs However MAs andhemorrhages have large variations in their appearance forexample shape and size Therefore how to design a suitabletemplate to match them becomes a vital problem

Apart from the above-described detection approachesseveral mathematical morphology based approaches havebeen proposed for the detection of red lesions Junior andWelfer [13] designed a five-stage detection approach usingmathematical morphology for detecting microaneurysmsand hemorrhages obtaining a sensitivity of 8769 and aspecificity of 9244 Ravishankar et al [14] put forwardan automatic feature extraction method for multiple lesionsdetection In their method geometrical relationships ofdifferent features and lesions can be used along with sim-ple morphological operations Jaafar et al [15] suggested acombination of mathematical morphology and rule-basedclassification approach for the detection of red lesions whichachieved a sensitivity of 8620 and a specificity of 9880However the above-mentioned mathematical morphologybased approaches heavily depend on the choosing of struc-turing elements and it may weaken the performance ofalgorithms when changing their size and shape To addressthis limitation some pixel classification based methods have

been proposed for the detection of red lesions in color fun-dus photographs [16ndash20] Niemeijer et al [16] incorporatedmorphological top-hat transform with a 119896NN classifier intoa unified framework for automatic detection of red lesionsAfter that Sanchez et al [17] employed the Gaussian mixturemodel and logistic regression classification for the detectionof red lesions Furthermore Zhang et al [18] designed amicroaneurysm detection approach which integrates thedictionary learning (DL) with Sparse Representation Clas-sification (SRC) The microaneurysms can be identified bycalculated reconstruction error Considering the importanceof multifeature Zhou et al [19] presented a novel approachwhich combines the multiple features with dictionary learn-ing for microaneurysm detection The total reconstructionerror of multifeature is used for microaneurysm classifica-tion Besides a microaneurysm detection approach usingSparse PCA and 1198792 statistic is developed in [20] Sincetheir approach just needs to learn one class data the classimbalance problem can be addressed

All the above-mentioned detection approaches regardimage pixels as the basic unit to distinguish red lesions fromnonred lesions However image pixels are a consequenceof the discrete representation of images but not naturalentities Compared to the pixel-based image representationsuperpixel-based image representation is more consistentwith human visual cognition and contains less redundancy[21ndash23] The reason lies in the fact that superpixel groupspixels into perceptually meaningful atomic regions with sim-ilar color and texture and has the ability to adhere to imageboundaries Therefore it can not only provide a convenientmanner to compute image features but also greatly reduce thecomplexity of subsequent image processing tasks Howeverhow to detect the true red lesions from a series of superpixelsegmentation results is still a problem

To overcome the above-mentioned issues in this paperwe propose a novel red lesion detection approach based onsuperpixel Multichannel Multifeature (MCMF) Two maincontributions are as follows on one hand a novel candidateextraction scheme based on superpixel segmentation is givenwhich is more consistent with human visual cognition andcontains less redundancy improving efficiency and accuracyof subsequent image processing tasks On the other handextensive features extracted from multiple-channel imageshave been proposed and introduced to our feature extractionprocess which can improve the performance of red lesiondetection

Our proposed approach consists of the following fivephases first of all preprocessing is used to make redlesions more visible Next candidates can be extracted byapplying superpixel segmentation in digital color fundusphotographs And then these candidates are characterizedusing not only intensity features in multichannel images(here ldquomultichannelrdquo means that a series of images canbe produced by different operations based on the originalimage) but also the contextual feature Besides Fisher Dis-criminant Analysis (FDA) [24] is used to perform the classi-fication of candidates Finally a postprocessing technique isapplied to distinguish red lesions from nonlesions appearingas red such as fovea and blood vessels Experiments on

Computational and Mathematical Methods in Medicine 3

Input image Preprocessing Candidateextraction

FeatureextractionClassificationPostprocessing

Output

Figure 2 A system description of the proposed approach

DiaretDB1 database [25] demonstrate the effectiveness of ourproposed method

The remainder of the paper is organized as followsSection 2 describes the proposedMCMF red lesion detectionalgorithmThe proposed approach is verified in experimentsand the corresponding experimental results are reported inSection 3 We end with the conclusion in Section 4

2 The Proposed Method

In this section we mainly introduce the proposed MCMFmethod for the detection of red lesions Figure 2 depictsa flowchart for our proposed approach which consists ofthe following five stages preprocessing candidate extractionfeature extraction classification and postprocessing Eachstage will be described in detail in the following subsections

21 Preprocessing The large luminosity poor contrast andnoise always occur in retinal fundus images [11] whichaffect seriously the diagnostic process of DR and automaticdetection of lesions especially for red lesions In orderto address these problems and make a suitable image forred lesion detection contrast limited adaptive histogramequalization (CLAHE) [26] method is applied to make thehidden features more visible Besides Gaussian smoothingfilter with a width of 5 and a standard deviation of 1 is alsoincorporated for reducing the effect of noise further Twoinstances for describing an improvement in color saturationand contrast between lesions and background are shown inFigure 3

22 Candidate Extraction Given a preprocessed image 119868firstly SLIC transforms the image 119868 into CIELAB color spaceAssume that there are 119875 pixels in the image and the numberof initialized cluster centers is 119896 the grid interval 119878 is definedas 119878 = radic119875119896 Secondly the distance between pixels and thecluster centers should be calculated in each 2119878 times 2119878 regionHere SLIC combines the color and pixel positions with thecluster centers [119897 119886 119887 119909 119910]119879 to compute the distance where[119897 119886 119887]119879 is three parameters from CIELAB color space and

[119909 119910]119879 is the position of the pixels The distance of the pixel 119894to the 119896th cluster center can be defined as in the following

119863119894119896 = radic1198892119897119886119887 + (119898119878 )2 1198892119909119910

119889119897119886119887 = radic(119897119894 minus 119897119896)2 + (119886119894 minus 119886119896)2 + (119887119894 minus 119887119896)2119889119909119910 = radic(119909119894 minus 119909119896)2 + (119910119894 minus 119910119896)2

(1)

where 119898 is the weight factor and its range varies from 1 to40 [21] and (119897119894 119886119894 119887119894) is the color values of the pixel 119894 at theposition of (119909119894 119910119894) of the image 119868

Region size means the size of superpixel segmentationwhich plays a vital role in SLIC For larger region sizes spatialdistances overweigh color proximity causing the obtainedsuperpixels to not adhere well to image boundaries Forsmaller region sizes the converse is true [21] Consideringthe red lesions always vary in size we will choose a suitablesize as our segmentation standard which will be discussed inour experiments Figure 4 gives the superpixel segmentationresults with different region sizes

23 Feature Extraction As for red lesion detection somespecific properties need to be considered Firstly some redlesions are very near the blood vessels making it hardto distinguish them just in traditional RGB color spaceMoreover the appearance of red lesions is similar to thenormal structures of the retinal such as the blood vessels andthe fovea causing too much false positive (FP) Finally theshape and size of red lesions especially for large hemorrhagesare varied Based on these facts we adopt the strategy ofextracting different features based on multichannel imagesfor each candidate in this paper The chosen multichannelimages are listed as follows

Channel_1-Channel_2Green channel image 119868119866 of the originalimage 119868original and enhanced green channel image 119868green areobtained after preprocessing (see Figures 5(a) and 5(b))

Channel_3Alternating Sequential Filtering (ASF) containinga set of morphological closing 120574 and opening 120599 operationswith different sizes of disc-shaped structuring element119870 (1020 and 40) is used to estimate the background of prepro-cessing image in terms of (2) The image of preprocessingis removed from 119891ASF and we will obtain the result image119868dark_enhanced according to (3) A typical result of this operationis illustrated in Figure 5(c)

119891ASF = 120599(119899119870) (sdot sdot sdot (120574(2119870) (120599(119870) (120574(119870) (119868green)))) sdot sdot sdot) (2)

119868dark_enhanced = 119891ASF minus 119868green (3)

Channel_4 Median filtering with a 25 times 25 pixel kernel to119868green is used for calculating background image 119868bg and then119868bg can be removed from 119868green to obtain the shade correlationimage 119868sc A series of morphological open operations using 12line structures of length 9 pixels with different angles ranging


(a) (b)

(c) (d)

Figure 3 From left to right original RGB color retinal images (a) and (c) enhancement images (b) and (d)

from 15 degrees to 165 degrees with the increase of 15 areapplied to 119868sc for locating blood vessels By combining themaximum pixel value at each pixel position in all 12 imagesthe blood vessels can be obtained These blood vessels arethen subtracted from 119868sc to form 119868lesions containing mainlyred lesions with small size The result of 119868lesions is depicted inFigure 5(d)

Channel_5 Morphological close operation [27] is applied to119868green for eliminating shrinks or thin objects by using a disc-shaped structuring element 119861 with radius of 10 pixels theresult image 119868close is illustrated in Figure 5(e)

Channel_6 119868Hue_enhanced can be obtained by extracting the huechannel image of preprocessing image (see Figure 5(f))

Channel_7 119868119872 is 119872 component image of the preprocessingimage in CMYK color space The dark structures such as redlesions blood vessels and the fovea in 119868119872 are shown as whiteOn the contrary the bright structures such as exudates andthe optic disc appear as black (see Figure 5(g))

For each of the above-described multichannel imagesfour kinds of statistic features including the maximum min-imum mean and median are extracted from each candidateBesides the total average intensity and standard deviation ofthe each preprocessing retinal image are also imported

Since red lesions appear as darker regions with brightersurroundings based on this characteristic we also develop a

novel and effective feature for distinguishing a candidate fromits surroundings and background by mean intensity and thebarycenter distance between the neighbor candidates Herelet av_119866119894 and 119901119894 denote the mean color in original greenchannel and barycenter position of the 119894th candidate (119877119894 119894 =1 2 119873) respectively The proposed contextual feature islisted as follows

119878119894 = sum119895isin119873(119894) ((av_119866119894av_119868green) times (11988911198892))1198731198941198891 =

10038171003817100381710038171003817av_119866119894 minus av_1198661198951003817100381710038171003817100381722 if av_119866119894 le av_119866119895minus 10038171003817100381710038171003817av_119866119894 minus av_1198661198951003817100381710038171003817100381722 if av_119866119894 gt av_119866119895

1198892 = 10038171003817100381710038171003817119901119894 minus 1199011198951003817100381710038171003817100381722 =10038171003817100381710038171003817100381710038171003817100381710038171003817sum119868119898isin119877119894 11986811987511989810038161003816100381610038161198771198941003816100381610038161003816 minus sum119868119898isin119877119895 1198681198751198981003816100381610038161003816100381611987711989510038161003816100381610038161003816

100381710038171003817100381710038171003817100381710038171003817100381710038172

2

119895 isin 119873 (119894)

(4)

where sdot 22 denotes a quadratic term of 1198972-norm 119873(119894)represents the set of neighbors of candidate 119894 and 119873119894 is atotal of pixels of candidate 119894 119868119875119898 is the barycenter positionvector constituted by position vector of pixel 119868119898 1198891 is themean intensity difference and 1198892 is the barycenter positionsdifference between candidate 119894 and its neighbor candidate 119895Here the neighbor is defined as the 7 times of the region sizeempirically


(a) (b)

(c) (d) (e)

Figure 4 Retinal image segmentation using SLIC (a) Original retinal fundus image (b) detail from part of (a) (c)ndash(e) superpixelsegmentation with region sizes 10 30 and 50 pixels respectively

Basically speaking our proposed contextual feature needsto calculate the mean gray value of each candidate av_119866119894with the aim of making the feature more stable and precisefor red lesions Global mean of image av_119868green is importedfor eliminating the influence caused by the varying retinalpigmentations and different image acquisition processes Ifthe value of 1198891 is positive and large current candidate is morelikely to be a true red lesion and the converse is false Besidesthe value of 119878119894 is also affected by the barycenter positionsdistance 1198892 between candidate 119894 and candidate 119895 From (3)it is clear that a larger 119878119894 indicates that the current candidatemore likely belongs to red lesions and vice versa

In summary thirty-one features are computed on eachcandidate and they can be represented as a vector in a 31-dimension feature set 119865 = 1198911 1198912 11989131 Since the differentfeatures 119891119894 always varied in values and ranges we normalizeeach feature for all the candidates to have zero mean and unitvariance by using the following

1198911015840119894 = 119891119894 minus 120583119894120590119894 (5)

where 120583119894 is the mean of the 119894th feature and 120590119894 is its standarddeviation

24 Classification In this section Fisher Discriminant Anal-ysis (FDA) [24] is applied to perform the classification ofthe red lesion candidates The basic idea of FDA is to seeka transformation matrix which maximizes the between-classscatter andminimizes thewithin-class scatter simultaneouslyAssume a labeled candidate dataset matrix 119883 = 1198831 1198832where 1198831 = 1199091 1199092 1199091198991 and 1198832 = 1199091 1199092 1199091198992the matrix 119883119896 (119896 isin 1 2 1 = red 2 = non-red) is from119877119889times119899119896 119889 denotes corresponding feature dimension and 119899119896 isthe total number of samples (119899 = 1198991 + 1198992) in the 119896th class

Let 119878119887 and 119878119908 be the between-class scatter matrix andwithin-class scatter matrix

119878119887 = 2sum119896=1

119899119896 (120583119896 minus 120583) (120583119896 minus 120583)119879

119878119908 = 2sum119896=1

sum119909119894isin119862119896

(119909119894 minus 120583119896) (119909119894 minus 120583119896)119879 (6)


(a) (b)

(c) (d)

(e) (f)

(g)

Figure 5 Multichannel images (a) Original green channel image 119868119866 (b)ndash(g) enhanced green channel image 119868green enhanced low intensitystructure image 119868dark_enhanced 119868lesions close operation image 119868close enhanced hue image 119868Hue_enhanced and119872 component image 119868119872 of CMYKcolor space respectively

where 120583119896 = (1119899119896) sum119909119894isin119862119896 119909119894 is the mean vector of the 119896thclass and 120583 = (1198831 + 1198832)119899 is the mean vector of the wholecandidate dataset

FDA transformationmatrix119882 can be found bymaximiz-ing the following optimization problem

119882 = 119908119879119878119887119908119908119879119878119908119908 (7)

The above optimization problem can be regarded as thegeneralized eigenvalue problem below [28]

119878119887120593 = 120582119878119908120593 (8)

where 120582 is the generalized eigenvalue and the vector 120593 is thecorresponding eigenvector which is one of columns of theFDA transform matrix119882


(a) (b)

(c) (d)

Figure 6 A series of morphological operations are used for blood vessels detection (a) 1198912 reduction of high intensity structures (b) 1198913 sumof morphological openings (c) 1198914 morphological reconstruction by dilation (d) 119891scale

5 regional minimum and close operation (eg we takescale (4) = 5)

Since the projected class means1198821198791205831 and1198821198791205832 are wellseparated [24] we can choose average of the two projectedmeans as a threshold for classification In this case thethreshold parameter 119888 can be found

119888 = 119882119879 (1205831 + 1205832)2 (9)

Given a new sample 119909 it belongs to class 1 if classificationrule119882119879119909 gt 119888 or else it belongs to class 225 Post-Processing After the classification stage some non-lesions appearing as red such as blood vessels and foveaare often present and may be erroneously detected as redlesions In order to avoid this problem a postprocessing stageis incorporated into our approach for removing them andimproving the robustness of the proposed method

251 Multiscale Blood Vessels Detection Since red lesionsand blood vessel have the similar appearance it is hard todistinguish them effectively Besides the red lesions cannotoccur on the blood vessels [12] In order to remove anypossible nonlesions caused by blood vessels a multiscalemorphological blood vessel extraction method (MSM) basedon [13] is modified Here multiscale consists of four differentsizes of disc-shaped structure element such as scale =[2 3 4 5] according to [29] For each scale we conductSteps 1 to 5 (scale_num is the number of disc structureelements and scale_num equals 4) to extract blood vesselsmap and the final blood vesselsmap can be obtained by fusing

the blood vessels segmentation maps under all the scalesMore details are listed as follows

Step 1 Morphological opening 120599 and closing 120574 operationswith structuring element 119870 = scale (119894) (119894 represents the 119894thiteration 119894 = 1 2 3 4 such as scale (1) = 2 and scale (4) = 5)are used to estimate the background of preprocessing image119868CLAHE according to (10) and the result 1198911 can be obtained

1198911 = 120574(119870) (120599(119870) (119868CLAHE)) (10)

Step 2 The high intensity structures can be eliminated bysubtracting the CLAHE result image from1198911 see Figure 6(a)

1198912 = 1198911 minus 119868CLAHE (11)

Step 3 Morphological opening with varying structuringelements is applied to 1198912 for locating blood vessels Here weuse the linear structuring elements 120595 with 12 different anglesranging from 15 degrees to 165 degrees with the increase of15 By accumulating the pixel values at each pixel position inall 12 images we can obtain the image 1198913 according to thefollowing (see Figure 6(b))

1198913 = 12059511198912 + 12059521198912 + sdot sdot sdot 120595121198912 (12)

Step 4 According to (13) a morphological dilation recon-struction is used to detect the blood vessels furthermorewhich is denoted by 119877 (see Figure 6(c))

1198914 = 119877 (1198913) (13)


(e)

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

+ + +

(a1) (b1) (c1) (d1)

Figure 7 Blood vessels extraction results (a)ndash(d) show the different detection results with varying sizes of disc-shaped structuring elements2 3 4 and 5 respectively (e) the combination result of (a)ndash(d) (a1)ndash(d1) are the blood vessels map differences between (e) and (a)ndash(d)respectively

Step 5 A binary vessel structure map 119891scale5 can be obtained

by combining morphological operator of regional minimumRMwith close operation119877close according to the following (seeFigure 6(d))

119891scale5 = 119877close (minusRM (1198914)) (14)

Repeat Steps 1 to 5 until the value of scale reaches themaximum iteration number (scale_num)

At last we will obtain the final blood vessels map 1198915 bycombining each of the blood vessels segmentation maps withthe logical OR operation under all the scales (see Figures7(a)ndash7(d)) (ldquomdashrdquo represents logical OR operation and theentire vessel network is shown in Figure 7(e))

1198915 = 11989115 1003816100381610038161003816100381611989125 sdot sdot sdot 10038161003816100381610038161003816119891scale_num5 (15)

The main difference between our proposed MSM bloodvessels extraction method and the single scale blood vesselsdetection method [13] lies in whether to take the multiscaleinto consideration or not Figures 7(a)ndash7(d) show the blood

vessels maps produced by [13] and Figures 7(a1)ndash7(d1) arethe difference images by subtracting the images in Figures7(a)ndash7(d) from MSM detection result shown in Figure 7(e)Comparing with these results we can learn that our proposedapproach achieves a better robust and complete blood vesselssegmentation result than the single scale blood vessels detec-tion [13] which can do well in reducing the FP in red lesiondetection

252 Elimination of the Fovea The fovea appears as darkregion located in the center of the macula region of theretina Since the fovea region has a relative low intensityprofile and its appearance is commonly much similar withthe background causing it to be hard to distinguish from truered lesions In order to improve the accuracy of the proposedmethod the removal of fovea is indispensable Amethod [16]by considering the spatial relationship between the diameterof the optic disc and the region of the fovea is adopted toremove the fovea The final red lesion map result is shownin Figure 8


(a) (b)

(c) (d)

Figure 8 The final result of our proposed MCMF (a) Original retinal image (b) the corresponding ground truth (c) final red lesions map(d) overlaying red lesions map on original retinal image

Figure 9 depicts more examples for our proposedMCMFapproach for red lesions detection in the retinal imagescontaining red lesions Figure 9(a) shows original colorretinal images Figure 9(b) illustrates their detected red lesionmaps From these results we can learn that our proposedMCMF red lesion detection approach can detect most of redlesions in the retinal images regardless of their varying sizesand shapes

3 Experimental Results and Analysis

31 Database Description In this section we conduct exten-sive experiments to validate and evaluate the effectivenessof our proposed red lesion detection method on publicDiaretDB1 retinal image database [25]

The DiaretDB1 database (Standard Diabetic RetinopathyDatabase Calibration level 1 version 1) [25] is a publicdatabase available on the web The database contains a totalof 89 RGB color fundus images with the fixed 1500 times 1152resolution and 50∘ field of view Among the 89 fundus images5 images are healthy and the remaining 84 retinal imagesare abnormal which are annotated by four different clinicalexperts In our experiment the training set consists of 40retinal images selected randomlyThe testing set is composedof the remaining 49 images

32 Assessment of Classification Performance We take twoevaluation criterions to verify the effectiveness of our pro-posed red lesion detection method such as sensitivity andspecificity These measures are calculated based on the given

red lesion ground truth which can be seen from the follow-ing

sensitivity = TPTP + FN

specificity = TNTN + FP

(16)

where true positive (TP) is the number of red lesions thatare correctly identified false negative (FN) is the number oflesions incorrectly found as nonred lesions false positive (FP)is the number of lesions incorrectly found as red lesions andtrue negative (TN) is the number of nonred lesions that arecorrectly identified These criteria are also used to evaluatethe performance of different methods for the detection of redlesions

In our experiments we employ the Receiver OperatingCharacteristics (ROC) curve to evaluate the effectiveness ofthe proposed red lesion detection method An ROC curve isthe plot of sensitivity on the vertical axis and (1 minus specificity)on the horizontal axis In ROC curve the upper left cornerrepresents perfect classification Besides the area under theROC curve is denoted as the AUC value for measuring anddescribing the algorithm performance A larger AUC valueindicates a better classifier performance

33 Results In this section we will carry out three experi-ments based on DiaretDB1 dataset to verify the effectivenessof our proposed method In our experiments we employtwo different kinds of criteria to evaluate the method per-formance including image-based [25] and pixel-based [30]


(a) (b)

Figure 9 Results of our proposed red lesion detection approach on abnormal retinal images (a)Original retinal images (b) the correspondingred lesions maps


criteria Image-based criterion means to classify an imageeither as ldquonormalrdquo or ldquoabnormalrdquo (ie detecting the absenceor presence of red lesions anywhere in the image) That isto say a retinal image is considered as pathological if itpresents one or more red lesions otherwise it is normal [22]However in pixel-based criterion we adopt the connectedcomponent level validation [28] by counting the numberof pixels detected correctly According to [28] connectedcomponents are considered as true positive (TP) if theytotally or partially overlap with the ground truth (ie ared lesion is considered as TP if the detected connectedcomponent overlapped at least 75 of the area manuallyannotated by the expert but less than 100 and all otherconditions are considered to be false detection) The TN FPand FN are calculated in the same manner Based on theseobservations algorithm validated based on the image-basedcriterion typically achieves a better performance than thepixel-based criterion since it is not necessary to detect allthe red lesions in the images For more details refer to [30]Taking the clinical point of view and screening applicationsinto consideration it is more interesting to evaluate theexperiment results at image-based criterion [31]

There is a parameter in our proposed method regionsize (the size of superpixel) which impacts the performanceof our proposed method How to choose a suitable sizebecomes a critical problem in our experiment Besides for thesame test sample different classifiers may obtain the differentclassification results So the choosing of classifier is equallyimportant

In our first experiment two supervised classifiers FDAand 119896NN (119896 nearest neighborhood) [32] as the underliningclassifiers are selected and the optimal one will be used forour following experiments Here we set 119888 as 001 for FDAclassification according to (9) Three different region sizes(10 30 and 50) as our segmentation standards are used toselect the optimal classifier based on image-based criterionFirstly training samples with three different region sizes areused to train corresponding classifiers containing FDA and119896NN respectively and then we adopt the trained classifiersto classify test samples with various region sizes and obtaincorresponding classification scores The experiment resultsare shown in Figure 10

According to Figure 10 we can learn that FDA classifiercan achieve the better results than 119896NN classifier in all theregion sizes The reasons are listed as follows since the 119896NNclassifier works by comparing the Euclidean distance of atest sample with 119896 labeled training samples which are alsoknown as 119896 nearest neighbors on the whole feature space dueto the original feature space containing some irrelevant orredundant features which causes misclassification for 119896NNDifferent from 119896NN FDA is a well-known linear techniquefor dimensionality reduction and feature extraction whichcan avoid the above problem greatly FDA makes full use ofthe label information to find the optimal projection vectorsby maximizing between-class scatter matrix and minimizingwithin-class scatter matrix simultaneously and achieves abetter classification performance Therefore the FDA willbe chosen as our optimal classifier for the following experi-ments

FDA region size = 10 AUC = 098FDA region size = 30 AUC = 096FDA region size = 50 AUC = 095kNN region size = 10 AUC = 089kNN region size = 30 AUC = 084kNN region size = 50 AUC = 079

01 02 03 04 05 06 07 08 09 101 minus specificity

0010203040506070809

1

Sens

itivi

tyFigure 10 Sensitivity versus 1minusspecificity curves with varied regionsizes by applying two different classifiers

Region size = 10Region size = 30Region size = 50


0010203040506070809

1

Sens

itivi

ty

Figure 11 ROC curves of pixel-based criterion on testing set withvarying region sizes

Besides judging from Figure 10 we can see that when theregion size is set to 10 our proposed algorithm achieves thebest performance both in FDA and in 119896NN Yet we can alsonotice that for each kind of classifiers the different choosingof region sizes is not sensitivity for our classification resultwhich can achieve quite similar ROC curves

In our second experiment we will find the optimal regionsize based on pixel-based criterion [30] The ROC curves areshown in Figure 11

Judging from the above experiment results it is easy tofind that the performances of all the region sizes in the pixel-based criterion are relatively lower than those in the image-based criterion The reason lies in the fact that the range of


Table 1 Average execute time of per image and AUC with varyingregion sizes at pixel-based criterion

AUC Average time in secondsRegion size = 10 074 11623 sRegion size = 30 073 3805 sRegion size = 50 070 1037 s

Table 2 Performance results on DiaretDB1 dataset

Authors Sensitivity SpecificityProposed method 8330 9730Sanchez et al [17] 8769 9244Ravishankar et al [14] 9510 9050Jaafar et al [15] 9880 8620Roychowdhury et al [33] 7550 9373

ground truth provided in database is not very precise andalways larger than the true size of red lesions So when eachobtained red lesion map compares with its ground truth theimprecise ground truth will affect the methodrsquos performanceWhat is more choosing the smaller size of superpixel thesegmentation results are more accurate than the bigger onesand when the region size is set to 10 our proposed algorithmachieves the best performance (ie AUC = 074) as shownin Table 1 However smaller region size leads to highercomputing complexity When the region size = 10 theaverage execute time for each image approximately takes11623 s and the average computing time reaches about 1037 sfor the region size = 50 listed in Table 1 (the average runningtime of per image is evaluated using Matlab R2015a on aPC with Intel Core i5 running Windows 7 at the clock of330GHZ with 16GRAM)

In addition considering the clinical point of view andfor screening applications [31] choosing the bigger regionsize has completely attended their goals Taking the abovereasons into consideration we set the region size (superpixelsegmentation size) as 50 in our next experiment

In our last experiment we employ image-based evalu-ation method proposed by Kauppi et al [25] to verify theeffectiveness of the proposed method by comparing it withother state-of-the-artmethods According to [25] each imageneeds to provide a score and a high score means a highprobability that a lesion presents in corresponding imageWith the provided scores the sensitivity and specificitymeasures can be calculated By image-based comparison theproposed approach achieves a sensitivity of 8330 and aspecificity of 9730 Table 2 lists several comparison resultsobtained by the existing red lesion detection methods onDiaretDB1 database

From the comparison results of various algorithms illus-trated in Table 2 we can know that our proposed methodis more reliable than the other methods and achieves satis-factory result But there are still some points that need to bementioned as follows Firstly from Figures 8 and 9 it can befound that the residues of blood vessels close to red lesionscannot be removed from the final red lesion maps causingfalse positives Secondly when there are small size red lesions

contained in the retinal images they may not be completelydetected

4 Conclusion

To summarize we put forward a novel red lesion detectionbased on superpixel Multichannel Multifeature (MCMF)classification in color retinal images which is able to detectthe red lesions efficiently regardless of their variability inappearance and size Firstly the whole image is segmentedinto a series of candidates using superpixel segmentationAnd then multiple features from the multichannel imagesas well as the contextual feature are proposed for describingeach candidate Next FDA is introduced to classify thered lesions among the candidates Finally a postprocessingtechnique is applied to distinguish red lesions from bloodvessels and fovea Experiment results on DiaretDB1 databasedemonstrate that our proposed method is effective for redlesion detection

Since our proposed approach extracts a number offeatures for each superpixel complex relationships amongthe extracted features exist and other classifier (eg neuralnetwork or Extreme LearningMachine) may lead better clas-sification result which could be validated and researched infuture works In addition applying our proposed frameworkto other lesions detection is also another interesting topic forfuture study

Conflicts of Interest

All authors declare that the support for this research does notlead to any conflicts of interest regarding the publication ofthis paper

Acknowledgments

This work is supported by National Natural Science Foun-dation of China (nos 61471110 and 61602221) Founda-tion of Liaoning Educational Department (L2014090) andFundamental Research Funds for the Central Universities(N140403005 N162610004 and N160404003)

References

[1] P J Kertes and M T Johnson Evidence Based Eye CareLippincott Williams amp Wilkins Philadelphia Pa USA 1stedition 2007

[2] B E Klein S E Moss R Klein and T S Surawicz ldquoTheWisconsin epidemiologic study of diabetic retinopathy XIIIRelationship of serum cholesterol to retinopathy and hardexudatesrdquo Archives of Ophthalmology vol 102 no 4 pp 520ndash526 1991

[3] O Faust A U Rajendra E Y K Ng K-H Ng and J S SurildquoAlgorithms for the automated detection of diabetic retinopathyusing digital fundus images a reviewrdquo Journal of MedicalSystems vol 36 no 1 pp 145ndash157 2012

[4] M D Abramoff M Niemeijer M S A Suttorp-Schulten MA Viergever S R Russell and B Van Ginneken ldquoEvaluation ofa system for automatic detection of diabetic retinopathy from


color fundus photographs in a large population of patients withdiabetesrdquo Diabetes Care vol 31 no 2 pp 193ndash198 2008

[5] R P Crick and P T Khaw A Textbook of Clinical Ophthal-mology A Practical Guide to Disorders of the Eyes and TheirManagement World Scientific Singapore 1997

[6] R N Frank ldquoDiabetic retinopathyrdquoTheNew England Journal ofMedicine vol 350 no 1 pp 48ndash58 2004

[7] C E Baudoin B J Lay and J C Klein ldquoAutomatic detectionof microaneurysms in diabetic fluorescein angiographyrdquo RevuedrsquoEpidemiologie et de Sante Publique vol 32 no 3-4 pp 254ndash261 1984

[8] T Spencer J A Olson K C McHardy P F Sharp and J V For-rester ldquoAn image-processing strategy for the segmentation andquantification of microaneurysms in fluorescein angiograms ofthe ocular fundusrdquo Computers and Biomedical Research vol 29no 4 pp 284ndash302 1996

[9] A J Frame P E Undrill M J Cree et al ldquoA comparison ofcomputer based classification methods applied to the detectionof microaneurysms in ophthalmic fluorescein angiogramsrdquoComputers in Biology and Medicine vol 28 no 3 pp 225ndash2381998

[10] L A Yannuzzi K T Rohrer L J Tindel et al ldquoFluoresceinangiography complication surveyrdquo Ophthalmology vol 93 no5 pp 611ndash617 1986

[11] G Quellec M Lamard P M Josselin G Cazuguel B Coch-ener andC Roux ldquoOptimal wavelet transform for the detectionof microaneurysms in retina photographsrdquo IEEE Transactionson Medical Imaging vol 27 no 9 pp 1230ndash1241 2008

[12] B Zhang X Wu J You Q Li and F Karray ldquoDetectionof microaneurysms using multi-scale correlation coefficientsrdquoPattern Recognition vol 43 no 6 pp 2237ndash2248 2010

[13] S B Junior and D Welfer ldquoAutomatic detection of microa-neurysms and hemorrhages in color eye fundus imagesrdquoInternational Journal of Computer Science and InformationTechnology vol 5 no 5 pp 21ndash37 2013

[14] S Ravishankar A Jain and A Mittal ldquoAutomated featureextraction for early detection of diabetic retinopathy in fundusimagesrdquo in Proceedings of the IEEE Computer Society Conferenceon Computer Vision and Pattern Recognition Workshops (CVPRrsquo09) pp 210ndash217 June 2009

[15] H F Jaafar A K Nandi andW Al-Nuaimy ldquoAutomated detec-tion of red lesions from digital colour fundus photographsrdquo inProceedings of the Annual International Conference of the IEEEEngineering inMedicine and Biology Society vol 2011 pp 6232ndash6235 2011

[16] M Niemeijer B Van Ginneken J Staal M S A Suttorp-Schulten and M D Abramoff ldquoAutomatic detection of redlesions in digital color fundus photographsrdquo IEEE Transactionson Medical Imaging vol 24 no 5 pp 584ndash592 2005

[17] C I Sanchez R Hornero A Mayo and M Garcıa ldquoMixturemodel-based clustering and logistic regression for automaticdetection of microaneurysms in retinal imagesrdquo in MedicalImaging 2009 Computer-Aided Diagnosis vol 7260 of Proceed-ings of SPIE Lake Buena Vista Fla USA February 2009

[18] B Zhang F Karray Q Li and L Zhang ldquoSparse representationclassifier for microaneurysm detection and retinal blood vesselextractionrdquo Information Sciences vol 200 no 1 pp 78ndash90 2012

[19] W Zhou C Wu D Chen Z Wang Y Yi and W DuldquoAutomatic microaneurysms detection based on multifeaturefusion dictionary learningrdquo Computational and MathematicalMethods inMedicine vol 2017 Article ID 2483137 11 pages 2017

[20] W Zhou C Wu D Chen Z Wang Y Yi and W Du ldquoAuto-matic microaneurysm detection using the sparse principalcomponent analysis based unsupervised classificationmethodrdquoIEEE Access vol 5 no 1 pp 2169ndash3536 2017

[21] R Achanta A Shaji K Smith A Lucchi P Fua and SSusstrunk ldquoSLIC superpixels compared to state-of-the-artsuperpixel methodsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 34 no 11 pp 2274ndash2281 2012

[22] A Levinshtein A Stere K N Kutulakos D J Fleet S JDickinson and K Siddiqi ldquoTurboPixels fast superpixels usinggeometric flowsrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 31 no 12 pp 2290ndash2297 2009

[23] O Veksler Y Boykov and P Mehrani ldquoSuperpixels and super-voxels in an energy optimization frameworkrdquo in Proceedingsof the European Conference on Computer Vision pp 211ndash224Crete Greece September 2010

[24] V N Vapnik ldquoAn overview of statistical learning theoryrdquo IEEETransactions on Neural Networks vol 10 no 5 pp 988ndash9991999

[25] T Kauppi V Kalesnykiene J-K Kamarainen et al ldquoTheDIARETDB1 diabetic retinopathy database and evaluation pro-tocolrdquo in Proceedings of the British Machine Vision Conference(BMVC rsquo07) pp 252ndash261 University ofWarwickWarwick UK2007

[26] K Zuiderveld ldquoContrast limited adaptive histogram equaliza-tionrdquo in Graphics Gems pp 474ndash485 1994

[27] T Chanwimaluang and G Fan ldquoAn efficient blood vesseldetection algorithm for retinal images using local entropythresholdingrdquo in Proceedings of the International Symposium onCircuits and Systems vol 5 pp 21ndash24 2003

[28] K Fukunaga Introduction to Statistical Pattern RecognitionAcademic Press 1972

[29] M Goldbaum S Moezzi A Taylor et al ldquoAutomated diagnosisand image understanding with object extraction object classi-fication and inferencing in retinal imagesrdquo in Proceedings of theIEEE International Conference on Image Processing (ICIP rsquo96)vol 3 pp 695ndash698 September 1996

[30] L Giancardo F Meriaudeau T P Karnowski Y Li KW TobinJr and E Chaum ldquoAutomatic retina exudates segmentationwithout a manually labelled training setrdquo in Proceedings of the8th IEEE International Symposium on Biomedical Imaging FromNano to Macro (ISBI rsquo11) vol 7906 pp 1396ndash1400 April 2011

[31] X Zhang G Thibault E Decenciere et al ldquoExudate detectionin color retinal images for mass screening of diabetic retinopa-thyrdquoMedical Image Analysis vol 18 no 7 pp 1026ndash1043 2014

[32] O D Richard E H Peter and G S David Pattern Classifica-tion John Wiley amp Sons New York NY USA 2001

[33] S Roychowdhury D D Koozekanani and K K Parhi ldquoScreen-ing fundus images for diabetic retinopathyrdquo Signals Systems ampComputers vol 43 no 4 pp 1641ndash1645 2012

Submit your manuscripts athttpswwwhindawicom

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


MEDIATORSINFLAMMATION

of


Behavioural Neurology

EndocrinologyInternational Journal of



Disease Markers


BioMed Research International

OncologyJournal of



Oxidative Medicine and Cellular Longevity


PPAR Research

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

ObesityJournal of



Computational and Mathematical Methods in Medicine

OphthalmologyJournal of


Diabetes ResearchJournal of



Research and TreatmentAIDS


Gastroenterology Research and Practice


Parkinsonrsquos Disease

Evidence-Based Complementary and Alternative Medicine

Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom


A optic discB foveaC maculaD blood vessels

E hemorrhageF microaneurysmG hard exudates

E

D

D

A BC

DF

G

Figure 1 A retinal image with different types of lesions and mainanatomical features

usage of intravenous contrast agents is not applicable foreveryone especially the pregnant woman [10] Thereforefluorescein angiograms cannot be widely used in public DRscreening programs and the adaption of digital color fundusphotographs is a better choice for screening purpose

A number of algorithms based on filtering have beenproposed for red lesion detection in digital color fundusphotographs Quellec et al [11] adopted a template matchingmethod based on subbands of wavelet transformed imagesfor microaneurysm detection which can solve the unevenillumination or high-frequency noise problems BesidesZhang et al [12] employed Multiscale Gaussian CorrelationCoefficients (MSCF) by computing the correlation coefficientbetween the grayscale distribution of microaneurysms andthe Gaussian function to detect MAs However MAs andhemorrhages have large variations in their appearance forexample shape and size Therefore how to design a suitabletemplate to match them becomes a vital problem

Apart from the above-described detection approachesseveral mathematical morphology based approaches havebeen proposed for the detection of red lesions Junior andWelfer [13] designed a five-stage detection approach usingmathematical morphology for detecting microaneurysmsand hemorrhages obtaining a sensitivity of 8769 and aspecificity of 9244 Ravishankar et al [14] put forwardan automatic feature extraction method for multiple lesionsdetection In their method geometrical relationships ofdifferent features and lesions can be used along with sim-ple morphological operations Jaafar et al [15] suggested acombination of mathematical morphology and rule-basedclassification approach for the detection of red lesions whichachieved a sensitivity of 8620 and a specificity of 9880However the above-mentioned mathematical morphologybased approaches heavily depend on the choosing of struc-turing elements and it may weaken the performance ofalgorithms when changing their size and shape To addressthis limitation some pixel classification based methods have

been proposed for the detection of red lesions in color fun-dus photographs [16ndash20] Niemeijer et al [16] incorporatedmorphological top-hat transform with a 119896NN classifier intoa unified framework for automatic detection of red lesionsAfter that Sanchez et al [17] employed the Gaussian mixturemodel and logistic regression classification for the detectionof red lesions Furthermore Zhang et al [18] designed amicroaneurysm detection approach which integrates thedictionary learning (DL) with Sparse Representation Clas-sification (SRC) The microaneurysms can be identified bycalculated reconstruction error Considering the importanceof multifeature Zhou et al [19] presented a novel approachwhich combines the multiple features with dictionary learn-ing for microaneurysm detection The total reconstructionerror of multifeature is used for microaneurysm classifica-tion Besides a microaneurysm detection approach usingSparse PCA and 1198792 statistic is developed in [20] Sincetheir approach just needs to learn one class data the classimbalance problem can be addressed

All the above-mentioned detection approaches regardimage pixels as the basic unit to distinguish red lesions fromnonred lesions However image pixels are a consequenceof the discrete representation of images but not naturalentities Compared to the pixel-based image representationsuperpixel-based image representation is more consistentwith human visual cognition and contains less redundancy[21ndash23] The reason lies in the fact that superpixel groupspixels into perceptually meaningful atomic regions with sim-ilar color and texture and has the ability to adhere to imageboundaries Therefore it can not only provide a convenientmanner to compute image features but also greatly reduce thecomplexity of subsequent image processing tasks Howeverhow to detect the true red lesions from a series of superpixelsegmentation results is still a problem

To overcome the above-mentioned issues in this paperwe propose a novel red lesion detection approach based onsuperpixel Multichannel Multifeature (MCMF) Two maincontributions are as follows on one hand a novel candidateextraction scheme based on superpixel segmentation is givenwhich is more consistent with human visual cognition andcontains less redundancy improving efficiency and accuracyof subsequent image processing tasks On the other handextensive features extracted from multiple-channel imageshave been proposed and introduced to our feature extractionprocess which can improve the performance of red lesiondetection

Our proposed approach consists of the following fivephases first of all preprocessing is used to make redlesions more visible Next candidates can be extracted byapplying superpixel segmentation in digital color fundusphotographs And then these candidates are characterizedusing not only intensity features in multichannel images(here ldquomultichannelrdquo means that a series of images canbe produced by different operations based on the originalimage) but also the contextual feature Besides Fisher Dis-criminant Analysis (FDA) [24] is used to perform the classi-fication of candidates Finally a postprocessing technique isapplied to distinguish red lesions from nonlesions appearingas red such as fovea and blood vessels Experiments on




Output









119863119894119896 = radic1198892119897119886119887 + (119898119878 )2 1198892119909119910


(1)










(a) (b)

(c) (d)












100381710038171003817100381710038171003817100381710038171003817100381710038172

2

119895 isin 119873 (119894)

(4)



(a) (b)

(c) (d) (e)




1198911015840119894 = 119891119894 minus 120583119894120590119894 (5)




119878119887 = 2sum119896=1

119899119896 (120583119896 minus 120583) (120583119896 minus 120583)119879

119878119908 = 2sum119896=1

sum119909119894isin119862119896

(119909119894 minus 120583119896) (119909119894 minus 120583119896)119879 (6)


(a) (b)

(c) (d)

(e) (f)

(g)




119882 = 119908119879119878119887119908119908119879119878119908119908 (7)


119878119887120593 = 120582119878119908120593 (8)



(a) (b)

(c) (d)




119888 = 119882119879 (1205831 + 1205832)2 (9)





1198911 = 120574(119870) (120599(119870) (119868CLAHE)) (10)


1198912 = 1198911 minus 119868CLAHE (11)


1198913 = 12059511198912 + 12059521198912 + sdot sdot sdot 120595121198912 (12)


1198914 = 119877 (1198913) (13)


(e)

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

+ + +

(a1) (b1) (c1) (d1)












(a) (b)

(c) (d)










(16)





(a) (b)









0010203040506070809

1

Sens

itivi




0010203040506070809

1

Sens

itivi

ty















4 Conclusion





Acknowledgments


References









































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of



















Output









119863119894119896 = radic1198892119897119886119887 + (119898119878 )2 1198892119909119910


(1)










(a) (b)

(c) (d)












100381710038171003817100381710038171003817100381710038171003817100381710038172

2

119895 isin 119873 (119894)

(4)



(a) (b)

(c) (d) (e)




1198911015840119894 = 119891119894 minus 120583119894120590119894 (5)




119878119887 = 2sum119896=1

119899119896 (120583119896 minus 120583) (120583119896 minus 120583)119879

119878119908 = 2sum119896=1

sum119909119894isin119862119896

(119909119894 minus 120583119896) (119909119894 minus 120583119896)119879 (6)


(a) (b)

(c) (d)

(e) (f)

(g)




119882 = 119908119879119878119887119908119908119879119878119908119908 (7)


119878119887120593 = 120582119878119908120593 (8)



(a) (b)

(c) (d)




119888 = 119882119879 (1205831 + 1205832)2 (9)





1198911 = 120574(119870) (120599(119870) (119868CLAHE)) (10)


1198912 = 1198911 minus 119868CLAHE (11)


1198913 = 12059511198912 + 12059521198912 + sdot sdot sdot 120595121198912 (12)


1198914 = 119877 (1198913) (13)


(e)

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

+ + +

(a1) (b1) (c1) (d1)












(a) (b)

(c) (d)










(16)





(a) (b)









0010203040506070809

1

Sens

itivi




0010203040506070809

1

Sens

itivi

ty















4 Conclusion





Acknowledgments


References









































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)

(c) (d)












100381710038171003817100381710038171003817100381710038171003817100381710038172

2

119895 isin 119873 (119894)

(4)



(a) (b)

(c) (d) (e)




1198911015840119894 = 119891119894 minus 120583119894120590119894 (5)




119878119887 = 2sum119896=1

119899119896 (120583119896 minus 120583) (120583119896 minus 120583)119879

119878119908 = 2sum119896=1

sum119909119894isin119862119896

(119909119894 minus 120583119896) (119909119894 minus 120583119896)119879 (6)


(a) (b)

(c) (d)

(e) (f)

(g)




119882 = 119908119879119878119887119908119908119879119878119908119908 (7)


119878119887120593 = 120582119878119908120593 (8)



(a) (b)

(c) (d)




119888 = 119882119879 (1205831 + 1205832)2 (9)





1198911 = 120574(119870) (120599(119870) (119868CLAHE)) (10)


1198912 = 1198911 minus 119868CLAHE (11)


1198913 = 12059511198912 + 12059521198912 + sdot sdot sdot 120595121198912 (12)


1198914 = 119877 (1198913) (13)


(e)

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

+ + +

(a1) (b1) (c1) (d1)












(a) (b)

(c) (d)










(16)





(a) (b)









0010203040506070809

1

Sens

itivi




0010203040506070809

1

Sens

itivi

ty















4 Conclusion





Acknowledgments


References









































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)

(c) (d) (e)




1198911015840119894 = 119891119894 minus 120583119894120590119894 (5)




119878119887 = 2sum119896=1

119899119896 (120583119896 minus 120583) (120583119896 minus 120583)119879

119878119908 = 2sum119896=1

sum119909119894isin119862119896

(119909119894 minus 120583119896) (119909119894 minus 120583119896)119879 (6)


(a) (b)

(c) (d)

(e) (f)

(g)




119882 = 119908119879119878119887119908119908119879119878119908119908 (7)


119878119887120593 = 120582119878119908120593 (8)



(a) (b)

(c) (d)




119888 = 119882119879 (1205831 + 1205832)2 (9)





1198911 = 120574(119870) (120599(119870) (119868CLAHE)) (10)


1198912 = 1198911 minus 119868CLAHE (11)


1198913 = 12059511198912 + 12059521198912 + sdot sdot sdot 120595121198912 (12)


1198914 = 119877 (1198913) (13)


(e)

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

+ + +

(a1) (b1) (c1) (d1)












(a) (b)

(c) (d)










(16)





(a) (b)









0010203040506070809

1

Sens

itivi




0010203040506070809

1

Sens

itivi

ty















4 Conclusion





Acknowledgments


References









































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)

(c) (d)

(e) (f)

(g)




119882 = 119908119879119878119887119908119908119879119878119908119908 (7)


119878119887120593 = 120582119878119908120593 (8)



(a) (b)

(c) (d)




119888 = 119882119879 (1205831 + 1205832)2 (9)





1198911 = 120574(119870) (120599(119870) (119868CLAHE)) (10)


1198912 = 1198911 minus 119868CLAHE (11)


1198913 = 12059511198912 + 12059521198912 + sdot sdot sdot 120595121198912 (12)


1198914 = 119877 (1198913) (13)


(e)

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

+ + +

(a1) (b1) (c1) (d1)












(a) (b)

(c) (d)










(16)





(a) (b)









0010203040506070809

1

Sens

itivi




0010203040506070809

1

Sens

itivi

ty















4 Conclusion





Acknowledgments


References









































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)

(c) (d)




119888 = 119882119879 (1205831 + 1205832)2 (9)





1198911 = 120574(119870) (120599(119870) (119868CLAHE)) (10)


1198912 = 1198911 minus 119868CLAHE (11)


1198913 = 12059511198912 + 12059521198912 + sdot sdot sdot 120595121198912 (12)


1198914 = 119877 (1198913) (13)


(e)

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

+ + +

(a1) (b1) (c1) (d1)












(a) (b)

(c) (d)










(16)





(a) (b)









0010203040506070809

1

Sens

itivi




0010203040506070809

1

Sens

itivi

ty















4 Conclusion





Acknowledgments


References









































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(e)

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

+ + +

(a1) (b1) (c1) (d1)












(a) (b)

(c) (d)










(16)





(a) (b)









0010203040506070809

1

Sens

itivi




0010203040506070809

1

Sens

itivi

ty















4 Conclusion





Acknowledgments


References









































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)

(c) (d)










(16)





(a) (b)









0010203040506070809

1

Sens

itivi




0010203040506070809

1

Sens

itivi

ty















4 Conclusion





Acknowledgments


References









































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)









0010203040506070809

1

Sens

itivi




0010203040506070809

1

Sens

itivi

ty















4 Conclusion





Acknowledgments


References









































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of























0010203040506070809

1

Sens

itivi




0010203040506070809

1

Sens

itivi

ty















4 Conclusion





Acknowledgments


References









































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


























4 Conclusion





Acknowledgments


References









































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of




















































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
















automated detection of red lesions using superpixel...

Documents