id
stringlengths
14
15
text
stringlengths
623
1.87k
source
stringclasses
1 value
2d87191d9618-0
ResearchArticle Chagas Parasite Detection in Blood Images Using AdaBoost Víctor Uc-Cetina,1Carlos Brito-Loeza,1and Hugo Ruiz-Piña2 1FacultaddeMatem ´aticas,UniversidadAut ´onomadeYucat ´an,AnilloPerif ´ericoNorte,TablajeCatastral,13615M ´erida,YUC,Mexico 2CentrodeInvestigacionesRegionalesDr.HideyoNoguchi,UniversidadAut ´onomadeYucat ´an,Avenida,Itz ´aesNo.490x59, ColoniaCentro,97000M ´erida,YUC,Mexico CorrespondenceshouldbeaddressedtoV ´ıctorUc-Cetina;uccetina@uady.mx Received17October2014;Revised20February2015;Accepted20February2015 AcademicEditor:IriniDoytchinova Copyright©2015V ´ıctorUc-Cetinaetal. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. TheChagasdiseaseisapotentiallylife-threateningillnesscausedbytheprotozoanparasite, Trypanosoma cruzi. Visualdetectionof suchparasitethroughmicroscopicinspectionisatediousandtime-consumingtask.Inthispaper,weprovideanAdaBoostlearning
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-1
solution to the task of Chagas parasite detection in blood images. We give details of the algorithm and our experimental setup.With this method, we get 100% and 93.25% of sensitivity and specificity, respectively. A ROC comparison with the method mostcommonlyusedforthedetectionofmalariaparasitesbasedonsupportvectormachines(SVM)isalsoprovided.Ourexperimental work shows mainly two things: (1) Chagas parasites can be detected automatically using machine learning methods with high accuracy and (2) AdaBoost + SVM provides better overall detection performance than AdaBoost or SVMs alone. Such resultsare the best ones known so far for the problem of automatic detection of Chagas parasites through the use of machine learning, computervision,andimageprocessingmethods. 1. Introduction The Chagas disease, also known as American trypanosomi- asis, is a potentially life-threatening illness caused by theprotozoan parasite, Trypanosoma cruzi (T. cruzi).According to the World Health Organization [ 1], it is found mainly in Latin America, where it is mostly transmitted to humans by thefaecesoftriatominebugs.Morethan25millionpeopleare at risk of the disease and an estimated 10 million people areinfected worldwide, mostly in Latin America where Chagasdiseaseisendemic.Approximately20,000deathsattributabletoChagasdiseaseoccurannually[ 2]. The Chagas disease presents itself in two phases. The
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-2
The Chagas disease presents itself in two phases. The initial,acutephaselastsforabouttwomonthsafterinfection.Duringtheacutephase,ahighnumberofparasitescirculatein the blood. When the Chagas disease is diagnosed earlyin this phase and a treatment is initiated, the patient can becured. During the chronic phase, the parasites are hiddenmainly in the heart and digestive muscle. In later years theinfectioncanleadtosuddendeathorheartfailurecausedbyprogressivedestructionoftheheartmuscle.Sometestscanbeusefulformakingadiagnosis,depend- ing on the phase of the disease. According to [ 3], the most typical tests used for the diagnosis of the Chagas disease are blood culture, chest X-ray echocardiogram, electrocar- diogram (ECG), enzyme-linked immunoassay (ELISA), and peripheralbloodsmears.Uptodate,oneofthemosteffective ways of detecting the Chagas disease in its initial phase isthrough the ELISA test. Another commonly used method is theChagasStat-Pakrapidimmunochromatographictest[ 4], which provides a performance comparable to that obtained withELISA. Screening blood donors for Chagas disease is of much concerninallLatinAmericancountries.AlthoughtheW orldHealth Organization (WHO) expert committee and some guidelines recommend a single ELISA test to screen blood donors[5],insomecountries,suchasBrazil,thereisamore restrictiveregulation,recommendingtwosimultaneoustests ofdifferenttechniques[ 6],performedinparallel.Oneofthe tests that can be performed in parallel is the inspection ofperipheralbloodsmears. Hindawi Publishing Corporation
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-3
Hindawi Publishing Corporation Computational and Mathematical Methods in Medicine Volume 2015, Article ID 139681, 13 pages http://dx.doi.org/10.1155/2015/1396812 ComputationalandMathematicalMethodsinMedicine A peripheral blood smear is basically a glass microscope s l i d ec o a t e do no n es i d ew i t hat h i nl a y e ro fv e n o u sb l o o d .The slide is stained with a dye, usually Wright’s stain, andexaminedunderamicroscope.Eventhoughvisualdetectionof the Chagas parasite through microscopic inspection ofp e r i p h e r a lb l o o ds m e a r si st h em o s tw i d e l yu s e dt e c h n i q u e for parasitemia determination, it is a time-consuming and laborious process. When the number of blood screeningsperformed in a laboratory increases, it becomes a problem.To cope with this problem, we introduce an automaticcomputational method for the detection of Chagas parasitesbasedonmachinelearningandimageprocessingalgorithms.Chagas detection using automatic image analysis is, to thebest of our knowledge, not yet studied as it is evidenced bythelackofpublicationsonthistopic. Currently, there is only a couple of papers reporting resultsonChagasparasitesdetectionusingmachinelearningmethods [ 8,9]. In the former, a Gaussian discriminant analysis is implemented and the resulting performance ratesare0.0167false-negatives,0.1563false-positives,0.8437true-negatives,and0.9833true-positives.Inthelatter,a 𝑘-nearest
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-4
neighbors binary classifier is trained and the performancesare 0.98 and 0.80 in sensitivity and specificity terms, respec-tively. The results reported in this paper using AdaBoost +SVMhaveasignificantimprovementoverthepreviousones. A p a r tf r o mt h e s et w op a p e r s ,n oo t h e rs t u d yu s i n g machine learning and computer vision methods has beenreported for the detection of the Chagas parasite, to the besto fo u rk n o w l e d g e .H o w e v e r ,s u c hk i n do ft e c h n i q u e sh a v ebeen extensively used for the detection of Malaria parasites [10–16]. All the approaches we reviewed utilize supervised and/orunsupervisedlearningmethodstodetect,classifyandq u a n t i f yt h en u m b e ro fm a l a r i ap a r a s i t e so nb l o o di m a g e s .The general process commonly reported can be divided inthreestages: (1)segmentation; (2)featuresextraction;and (3) classification. In the first stage, the segmentation is obtained through differentmethodsbasedonimagehistrogramcomputations.The second stage is performed through the computation ofdifferent features that can be classified in four categories[15]: texture features, color features, geometric features, and features obtained from human expert knowledge. The thirdand final stage makes the biggest difference among themethods, some applied neural networks [ 10], some others applied statistical measures [ 11,13], bayesian classifiers [ 12], 𝑘-means clustering [ 17], and other methods [ 14,16]. Brief
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-5
𝑘-means clustering [ 17], and other methods [ 14,16]. Brief summariesoftheapproachesmorerelevantforourownworkarepresentednext. In [10], an image classification system implementing a two-stage tree classifier using back-propogation neuralnetworks is introduced. Such a system identifies malariaparasites present in thin blood smears and classifies themaccording to their different species. Image features based oncolor,texture,andthegeometryallowthesystemtoperform morphological and threshold selection of possible parasites, distinguishingthemfromsaneerythrocytesandothercells. Similarly, the study presented in [ 11] provides a method for quantification and classification of erythrocytes instained thin blood films infected with malaria parasites.Thisapproachiscomposedofthreemainphases.First,there isapreprocessingstep,whichcorrectsluminancedifferences.Second,thereisasegmentationstepthatusesthenormalizedRGBcolorspaceforclassifyingpixelseitheraserythrocyteorbackground. Third, there is a two-step classification processidentifies infected erythrocytes and diagnoses the infection stage, using a trained bank of classifiers. An interesting characteristic of this method is that user intervention isallowed when the approach cannot make a proper decision.Automatic identification of infected erythrocytes showed aspecificity of99.7% and a sensitivityof94%. Meanwhile, theinfectionstagewasdeterminedwithanaveragesensitivityof78.8%andaveragespecificityof91.2%. In [12], the parasite detector uses a Bayesian pixel classi-
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-6
In [12], the parasite detector uses a Bayesian pixel classi- fier to mark stained pixels. The class conditional probabilitydensity functions of the stained and the nonstained classesareestimatedusinganonparametrichistogrammethod.Thestained pixels are further processed to extract features suchasHumoments,relativeshapemeasurements,andcolorforaparasiteornonparasiteclassifier. Finallya distanceweighted𝑘-nearest neighbor classifier is trained with the extracted features achieving 74% of sensitivity, 98% of specificity, 88%ofpositiveprediction,and95%ofnegativepredictionvaluesforthemalariaparasitedetection. Another malaria study based on pattern matching and parameter optimization is presented in [ 13]. In this case, the parasitaemia measurements are carried out by partitioningtheuninfectedandinfectedcellsusingunsupervisedmachinelearning. A comparison of the performance is done witha training-based method which improves the classification ratesgiving92%ofprecisionand95%ofrecall. Even when some machine learning methods have been successfully applied to the detection, classification, andquantification of Malaria parasites, these methods cannotbe employed in a straightforward way to detect Chagasparasites for one powerful reason: their morphology isdifferent. Malaria parasites have a ring shape ( Figure1(a) ) while Chagas parasites have a curved shape, similar to ashrimp (Figure1(b) ). This difference in their morphology prevents us from using exactly the same features during theveryimportantstageoffeaturesselection. In addition to detection algorithms, some authors have developed complete automated systems. In [ 18], an auto- mated system to identify and analyze parasite species onthick blood films by image analysis techniques is presented.The system comprises two main components: (1)image acquisition unit and (2)image analysis module. The authors
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-7
acquisition unit and (2)image analysis module. The authors have developed an image acquisition system that can beeasily mounted on most conventional light microscopes. Itautomatically controls the movement of microscope stagein 3-directional planes. The vertical adjustment (focusing)can be made in a nanometer range (7–9nm). Images areacquiredwithadigitalcamerathatisinstalledatthetopofthemicroscope. The captured images are analyzed by an image analysis software which utilizes computer vision algorithms todetectandidentifymalariaparasites. Otherworksrelatedtotheapplicationofpatternrecogni- tion methods for images taken with microscopes are those used for the detection of special types of cells such asComputationalandMathematicalMethodsinMedicine 3 (a) (b) (c) Figure 1: (a) Malaria parasites in early form (inside green squares); (b) Chagas parasite (inside red square); (c) blood cell that looks very similartoaChagasparasite(insidebluesquare). cancerous cells [ 19] and cervical cells [ 20]. In the first work t h ea u t h o r su s e dt h es h a p e ,s i z e ,a n dt e x t u r eo ft h ec e l l sto perform a classification, meanwhile in the second worka𝑘-means clustering algorithm for designing binary tree classifiers is used, along with the Bhattacharyya distancemetric. Inthispaper,wepresentacomparisonoftwoofthemost robust algorithms for binary classification. Both methodshavebeenextensivelystudiedandtestedwithdifficultpatternrecognition problems on images such as face detection [ 21, 22]. We provide an AdaBoost and a SVM learning solution
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-8
22]. We provide an AdaBoost and a SVM learning solution to the task of Chagas parasite detection in blood images.Our AdaBoost solution includes the definition of a newset of Haar-like features specially designed for learning theChagas parasite’s morphology pattern. We give details ofthe algorithms and our experimental setup. With the bestresulting method, AdaBoost, we get 100% and 93.25% ofsensitivity and specificity, respectively. Our experimentalwork shows mainly two things: (1)Chagas parasites can be detectedautomaticallyusingmachinelearningmethodsand(2)AdaBoost+SVMprovidesbetterdetectionperformance thanplainAdaBoostorSVMs. In the next section, we provide a brief review of the two compared machine learning methods: AdaBoost andSVM.Section3 describesthefourstagesrequiredtoanalyze t h ei m a g e sl o o k i n gf o rs u b w i n d o w sc o n t a i n i n gaC h a g a sparasite. The experimental work is detailed in Section4, which contains the experimental methodology we followedand the experimental results. In Section5,w ed i s c u s st h e results and point out the issues that need to be addressed ifthesystemistobeestablishedanddeployedinreality.Finally,ourconclusionsareexplainedin Section6. 2. AdaBoost and SVM AdaBoost is the short for Adaptive Boosting which is cur- r e n t l yt h em o s tu s e db o o s t i n gm e t h o d .Th eg o a lo fb o o s t i n gis to improve the accuracy of any given learning algorithm.Boosting creates an ensemble of classifiers by training andaddingonecomponentclassifieratatime.Eachnewclassifier
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-9
istrainedusingadifferentsubsetofexamples.Thenewtrain-ingsubsetcontainsexamplesthatareincorrectlyclassifiedbythe current ensemble. By doing such an iterative selection ofdifficult examples, boosting methods improve the accuracyof any supervised machine learning algorithm. Althougheach component classifier has an accuracy just greater thanaverage, the joint decision rule of the ensemble has a highaccuracywithallpreviouslyselectedtrainingexamples[ 23]. 2.1. Boosted Cascade of Simple Features. In 2001, Viola and Jones [7] introduced an extremely rapid approach for visual objectdetectionmotivatedbythetaskoffacedetection.Suchanapproachhasasitscentralcomponentanalgorithmbasedon the standard AdaBoost learning method. In addition tot h eu s eo fb o o s t i n g ,t w oa r et h ec h a r a c t e r i s t i c st h a tm a k ethis approach very fast and efficient. First, the use of theso-called integral image allows the features used by thedetector to be computed very quickly. Second, a methodfor combining the increasingly more complex classifiers in acascade,asillustratedin Figure2,allowsbackgroundregions o ft h ei m a g et ob eq u i c k l yd i s c a r d e dw h i l es p e n d i n gm o r e computingresourcesonpromisingregions. ThealgorithmintroducedbyViolaandJonesispresented next. (i) Given example images (𝑥 1,𝑦1) ,...,( 𝑥𝑛,𝑦𝑛)where 𝑦𝑖=0 , 1for negative and positive examples, respec- tively.
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-10
tively. (ii) Initializeweights 𝑤1,𝑖= 1/2𝑚, 1/2𝑙 for𝑦𝑖=0 , 1resp- ectively, where 𝑚and𝑙are the number of negatives andpositives,respectively. (iii) For 𝑡=1 ,...,𝑇 : (1) Normalizetheweights, 𝑤𝑡,𝑖←󳨀𝑤𝑡,𝑖 ∑𝑛 𝑗=1𝑤𝑡,𝑗(1) sothat 𝑤𝑡isaprobabilitydistribution.4 ComputationalandMathematicalMethodsinMedicine Cascade of classifiers All subwindows subwindows subwindowsAcceptedTrue True True False False False Discardedh1 h2 hT··· Figure 2: The detection cascade. Classifiers with increasing com- plexity are arranged in a cascade scheme to allow background regions of the image to be quickly discarded while spending morecomputingresourcesonpromisingregions. (2) For each feature, 𝑗, train a classifier ℎ𝑗which is restricted to using a single feature. The error isevaluatedwithrespectto 𝑤 𝑡,𝜖𝑗=∑𝑖𝑤𝑖|ℎ𝑗(𝑥𝑖)− 𝑦𝑖|. (3) Choosetheclassifier, ℎ𝑡,withthelowesterror 𝜖𝑡. (4) Updatetheweights:
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-11
(4) Updatetheweights: 𝑤𝑡+1,𝑖=𝑤𝑡,𝑖𝛽1−𝑒𝑖 𝑡, (2) where 𝑒𝑖=0if, for example, 𝑥𝑖is classified correctly, 𝑒𝑖=1otherwise,and 𝛽𝑡=𝜖𝑡/(1 − 𝜖𝑡). (iv) Thefinalstrongclassifieris ℎ(𝑥)={{ {{ {1𝑇 ∑ 𝑡=1𝛼𝑡ℎ𝑡(𝑥)≥1 2𝑇 ∑ 𝑡=1𝛼𝑡, 0otherwise ,(3) where 𝛼𝑡=log(1/𝛽𝑡). 2.2. Support Vector Machines. As u p p o r tv e c t o rm a c h i n e (SVM)isalearningtoolthatoriginatedinmodernstatisticallearning theory [ 24]. In recent years, SVM learning has found a wide range of real-world applications, includinghandwritten digit recognition, object recognition, speakeridentification, face detection in images, and text catego-rization. The formulation of SVM learning is based on theprinciple of structural risk minimization. SVM tends toperform well when applied to data outside the training setand it has been reported that SVM-based approaches areabletosignificantlyoutperformcompetingmethodsinmanyapplications. SVM achieves this by focusing on the trainingexamples that are most difficult to classify. These trainingexamplesarecalledthesupportvectors. 3. Chagas Parasites Detection
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-12
3. Chagas Parasites Detection The whole process of Chagas detection is divided into fourstages, as illustrated in Figure3. This process has been spe- cially designed to allow the automatization of the diagnosisas much as possible. The robustness of our methodologyrelies on the four modular stages that can be implemented Image acquisition Preprocessing Learning-based postprocessingAdaBoost detection Figure 3: The process of parasite detection consists of four stages: (1)image adquisition via microscope; (2)image conversion from RGBcolortograyscaleformat; (3)possibleparasitesdetectionusing AdaBoost; and (4)amount of DNA pixels used to further discard falsepositives. a ss e p a r a t e dp r o g r a m m i n gc l a s s e s ,w h e r ee v e np a r a l l e lp r o - gramming,thatis, usingGraphics Processing Units(GPUs),c a na l s ob ee m p l o y e dt or e d u c et h et i m eo fd e t e c t i o n[25]. This methodology is generic enough to be applied in otherobjectdetectionsystems,particularlyfromthebiologydomain where samples observed through a microscope arestained. The first step consists in obtaining a digital image using ac a m e r aa n dam i c r o s c o p e .O n c ew eh a v et h ei m a g ei nRGB format, in a second step we convert it to a grayscaleimagetoreducetheamountofinformationthattheAdaBoostalgorithm needs to process. In the third stage, the grayscale image is scanned using our previously trained AdaBoost
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-13
image is scanned using our previously trained AdaBoost classifier in order to detect subwindows of pixels containingpossibleparasites.Finally,inthefourthstep,weusethegreencomponent of the original RGB image to extract some threefeatures related to the number of pixels representing a highDNA (deoxyribonucleic acid) content. Based on the amountofDN Aitispossibletodiscardfalseparasites.ThislaststageisimplementedwithaSVManditisveryimportantbecauseitallowsustoinclude,aspartoftheclassificationprocedure,aprioriknowledgeabouttheDNAoftheparasites. It is important to mention that the DNA from both parasitesandnonparasitessuchaswhitebloodcellstendstoabsorbdifferentamountsofthestainemployed.However,thepattern generated by the stained DNA of parasites is clearlydifferent from the pattern generated by the stained DNAof white blood cells. Such a difference is recognized by thepatternrecognitionalgorithms. 3.1.Stage1:ImageAcquisition. Agroupofmicewereinfected with an inoculation of 5 × 104blood trypomastigotes of T. cruzivia intraperitoneal. Once the mice were infected, the parasitaemia detection started in average between 11 and 15daysafterwards.Atthistime,thebloodsmearswerepreparedandstainedusingWrightstain,whichallowstheobservationofthemorphologyofdifferentbloodcells,aswellasparasitesComputationalandMathematicalMethodsinMedicine 5 (a) (b) Figure4:(a)GenericHaar-likefeaturesinspiredbythoseproposedbyViolaandJones[ 7];(b)Haar-likefeaturesspeciallydesignedtocapture Chagasparasite’smorphology.
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-14
Chagasparasite’smorphology. such asT. Cruzi,Leishmania sp., andPlasmodium sp,. After the staining process the blood smears were placed verticallyand were left to dry. Finally, an optical Nikon Eclipse E600microscope was used to take images, first at 10x and then at100x,witharesultingsizeincreaseof1000times. 3.2. Stage 2: RGB to Grayscale Format. The conversion from an RGB image to a grayscale image involves a simple mani-pulation of matrices and is performed to reduce the amountofinformationinvolvedinthelearninganddetectionprocess.The RGB color model is an additive color model in whichred, green, and blue light are added together in differentproportions to produce a wide array of colors. A color inthe RGB color model is described by indicating how muchof each of the red, green, and blue is included. The color isexpressedasatriplet (𝑟, 𝑔, 𝑏)whereeachcomponentcanvary from zero to a defined maximum value, which in our casewas256.R G Bi m a g e sa r es t o r e di nm e m o r ya s 𝑤×ℎ×3 matrices, where 𝑤is the width and ℎis the height of the imagemeasuredinnumberofpixels.Thethirddimensionofthe matrix which is of size 3corresponds to the 3different
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-15
color components of the image: red, green, and blue. Toconvert any RGB image to a grayscale representation of itsluminance, first one must obtain the values of its red, green,andbluecomponents.Then,weneedtoaddtogether30%oft h er e dv a l u e ,5 9 %o ft h eg r e e nv a l u e ,a n d1 1 %o ft h eb l u e value. These percentages are regarded as typical values. The resultingimageisstoredina 𝑤×ℎimagecontainingpixelsof differentintensitiesofthegraycolor. 3.3. Stage 3: Parasites Detection. The Chagas detection pro- cess employs a set of Haar-like features and one AdaBoost binary classifier. Haar-like features provide informationabout clear and dark regions in images. Such informationis very useful to detect types of objects that share thesame morphological pattern. Since Chagas parasites can beconsidered objects that share the same shrimp-like shape,t h e yc a nb ed e t e c t e du s i n gt h ea p p r o p r i a t es e to fH a a r - l i k efeatures. Givenadetectionwindowtakenfromaspecificlocation ofthegrayscaleimage,aHaar-likefeatureconsidersadjacentrectangularregions,sumsupthepixelintensitiescontainedineachrectangularregionandcalculatesthedifferencebetweenthese sums. This difference is then compared to alearnedthreshold that separates nonparasites from parasites. The rectangular regions are defined in such a way that relevantdark pixels fall into one same region, meanwhile clearpixels fallintoanotherone. Figure4illustratestheHaar-like featuresutilizedtodetectChagasparasites. The nine features in Figure4(a) were inspired by the
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-16
The nine features in Figure4(a) were inspired by the originalHaar-likefeaturesproposedbyViolaandJonesin[ 7]. ThesetoffourHaar-likefeaturesin Figure4(b) weredesigned specifically to represent the shrimp-like shape adopted byChagas parasites. After running ten times the AdaBoostalgorithm and examining the number of times that thealgorithmchosetheHaar-likefeaturesasweaklearners,itwaspossible to effectively discard those Haar-like features thatfailed to capture important patterns of the Chagas parasites,while keeping only the best ones. Figure5illustrates how specially designed Haar-like features adapt better to themorphologyoftheparasite. Moreover, the position of these rectangular regions is defined relative to a detection window that acts like abounding box to the target parasite. In the detection phase,a window of the target size is moved over the input image,andforeachsubsectionoftheimage,theHaar-likefeatureis computed. The Haar-like features can be computed very fast using the concept of integral image. Given an image 𝐼of size 𝑅× 𝐶,w h e r e 𝑅is the number of rows and 𝐶is the number of columns, the integral image of a point located in row 𝑟and column 𝑐isdefinedas 𝑖𝑖(𝑟, 𝑐)=∑ 𝑟󸀠≤𝑟,𝑐󸀠≤𝑐𝐼( 𝑟󸀠,𝑐󸀠),(4) where 𝐼(𝑟󸀠,𝑐󸀠)is the gray intensity value of pixels located in
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-17
row𝑟󸀠andcolumn 𝑐󸀠.Notethatthenumberofrowsincreases intheimagefromtoptobottomandthenumberofcolumnsincreasesfromlefttoright. Now, given the same image 𝐼and four points 𝑝 1(𝑟1,𝑐1), 𝑝2(𝑟2,𝑐2), 𝑝3(𝑟3,𝑐3),a n d 𝑝4(𝑟4,𝑐4)as illustrated inFigure6(a) , and using the definition of integral image of (4),w ec a nc o m p u t et h er e g i o n s 𝑆1,𝑆2,𝑆3,a n d 𝑆4as follows. To calculate the sum of pixels in subwindow 𝑆1, denoted by 𝜎(𝑆1),wesimplycompute 𝜎( 𝑆1)=𝑖𝑖 ( 𝑝1)=𝑖𝑖 ( 𝑟1,𝑐1). (5)6 ComputationalandMathematicalMethodsinMedicine Figure 5: Haar-like features on top can capture more information abou tthemorphology oftheChagas parasite,given their circular shape, thanthegenericHaar-likefeaturesproposedbyViolaandJones,shownonthebottom. S1S2 S3S4(r1,c1) (r2,c2) (r3,c3)( r4,c4)(1,C) (R,C ) (R,1)(1,1) (a) R1R2 R3 R4R5 (b)
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-18
(a) R1R2 R3 R4R5 (b) Figure6:(a)Genericintegralimagecomputation;(b)computationofoneChagas-specificintegralimage. For𝑆2and𝑆3wecompute,respectively, 𝜎( 𝑆2)=𝑖𝑖 ( 𝑟2,𝑐2)−𝑖𝑖 ( 𝑟1,𝑐1), 𝜎( 𝑆3)=𝑖𝑖 ( 𝑟3,𝑐3)−𝑖𝑖 ( 𝑟1,𝑐1).(6) Finally,for 𝑆4,wecalculate 𝜎( 𝑆4)=𝑖𝑖 ( 𝑟4,𝑐4)−𝑖𝑖 ( 𝑟3,𝑐3)−𝑖𝑖 ( 𝑟2,𝑐2)+𝑖𝑖 ( 𝑟1,𝑐1).(7) In general, for an image of size 𝑅×𝐶,thewho lein t egral image 𝑖𝑖(𝑅, 𝐶)c a nb ec o m p u t e di no n l yo n ep a s so v e rt h e imageandcanbestoredina2-dimensionalarray.Therefore,tocomputeasubwindowlike 𝑆 4inFigure6(a) ,weonlyneed toaccesstheintegralimagearrayfourtimes.Furthermore,tocomputetheintegralimageofourChagasHaar-likefeatures, wejustneedtosegmenteachHaar-likefeatureintheminimalnumber 𝑛 𝑟of subwindows required to cover it, this number
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-19
𝑟of subwindows required to cover it, this number being 𝑛𝑟=3for the simplest one and 𝑛𝑟=5for the most complex. The computation of one Chagas-specific integralimagewith5subwindowsisillustratedin Figure6(b) ,where theHaar-likefeaturetemplatehasbeendividedinto5regions,namely, 𝑅 1,...,𝑅5andinordertocom putethefea turevalue weneedtocalculate 𝜎(𝑅1)+𝜎 ( 𝑅2)+𝜎 ( 𝑅3)+𝜎 ( 𝑅4)−𝜎 ( 𝑅5). Inconclusion,thecomputationoftheintegralimageforourspeciallydesignedHaar-likefeaturesisstillveryfast. 3.4. Stage 4: Postprocessing. The cascade application of the AdaBoost algorithm together with the appropriate set ofComputationalandMathematicalMethodsinMedicine 7 DNA (a) 180 140 100 60 12 8 4 004812 (b) Figure7:(a)DarkspotofpixelsgeneratedbyanaccumulationofDNA;(b)aDNAspotseenasasurfaceof 11 × 11pixels. Haar-like features is highly effective in finding most of the trueChagasparasites.However,italsodetectsasmallnumberof false parasites. The postprocessing filter described in thissectionhelpstodiscardfalseparasitesanddecreasesthefalsepositiverateofourmethod. Thepostprocessingfilterisfocusedontheanalysisofone dark spot of pixels appearing in all parasites’ bodies. Thisspot, indicated by an arrow in Figure7(a) , corresponds to
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-20
an accumulation of DNA of the Chagas parasite. When the imageisplottedasasurface,itcanbeseenthatsuchDNAspothas one particular valley-like shape, shown in Figure7(b) . Sincetheshapeofthisdarkspotandthelowintensityvaluestakenbytheircorresponingpixelscreateapatternveryusefultodiscriminatebetweenparasitesandnonparasites,theywereusedasasourceforfeatureextractionasitisdescribednext. Given that the stained pixels have low intensities, these values being not greater than 80, in the green component oftheRGBimage,thefollowingthreefeatureswereusedtotrainaSVM. (i) Feature 1 . Given a subwindow detected by AdaBoost, the percentage of pixels that have intensities at most 80 wascomputed. This percentage, which represents the size of thestainedregionoftheparasite,isthefirstfeature. (ii) Feature 2 . The mean of the intensities of all pixels with individual intensities at most 80 was encountered in thesubwindowdetectedbyAdaBoost. (iii) Feature 3 . The standard deviation of the intensities of all pixelswithindividualintensitiesatmost80wasencounteredinthesubwindowdetectedbyAdaBoost. The choice of the green component over the red and the blue ones was decided after a careful visual examination ofa subset of images in the original RGB version and in all 3 individual components: red, green, and blue. After such visualexamination,itwasclearthatthegreencomponentwassuperiortotheothersintermsofdiscriminativeinformation.In other words, the information provided by the green channelallowsustoperformamoreaccurateclassificationofparasitesandnonparasites. 4. Experimental Work
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-21
4. Experimental Work 4.1.ExperimentalMethodology. W etestedtheproposedalgo- rithm using a data collected in the Instituto de Investiga-ciones Regionales at the Universidad Aut ´onomadeYucat ´an, M´exico. We had available for our study a total of 120 c o l o ri m a g e so fd i m e n s i o n2 5 6 ×256 pixels. Sixty of these images were specially selected to contain a Chagas parasite.Meanwhile, the other 60 remaining images were selectedto contain nonparasites. Machine learning was employed togenerate our basic classifier algorithm, which was then usedfor the more general Chagas parasite detection process. Theresultsweprovideinthispaperwereobtainedusingatypicalpattern recognition methodology [ 26], using training and testingsetsofimages. A 10-fold cross-validation procedure was used for train- ing and testing the AdaBoost and the SVM + FeatureExtraction learning methods. Since AdaBoost is sensible torotation, during the training phase, we generated a total of840 positive examples of parasites images by rotating the original images in increasing amounts of 15 degrees. For every experiment, we used 756 positive images for training,leaving 84 positive images for testing. Even though we werelimited by the number of positive examples available, thenumber of negative examples could be defined as desired.All negative examples were generated by random samplingfrom the original 256 × 256 pixels images. For training,
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-22
AdaBoost used as many negative examples as possible toreducethefalsepositiverate,andthefinallearnedmodelwastested with 10800 negative images. In the case of the SVM +Feature Extraction methods, we used 1296 negative imagesfortrainingand10800fortesting.Itisworthmentioningthatonce the classifiers have been trained, the proper analysis of8 ComputationalandMathematicalMethodsinMedicine a256 ×256image,whichmeansthatthealgorithmisactually lookingforparasites,isperformedinrealtime.Wearetalkingabout a few milliseconds, using a commercial laptop ordesktopcomputer.Therefore,severalhundredimagescanbeanalyzedinamatterofminutes.Moreover,ourprogramwasimplementedinMatlabusing2publiclibraries:fdtool,forthe AdaBoostimplementation,andlibsvm,forthesupportvector machineimplementation. 4.1.1.ComparisonwithSupportVectorMachines. Theapplica- tionofSVMlearningrequiresthecomputationofonefeaturevector for each subwindow that needs to be classified asparasiteornonparasite.Suchfeaturesvectorsarethenpassedto the SVM classifier which makes the final decision. Thes e l e c t i o no ft h ef e a t u r e si sak e yp a r to ft h ec l a s s i fi c a t i o nprocess, simply because by using the wrong features thelearning algorithm cannot extract the patterns required tobuild a confident classifier, resulting in a high classificationerror. In order to obtain the best results with the applicationof SVMs, we reviewed the most relevant published worksinvolving SVM learning and malaria parasite detection. Sothatwecanuseforourexperimentalworkthosefeaturesthathavebeenpreviouslyreportedtoprovidehighqualityresults. Color,shape,andtexturefeatureswereselectedtoforma
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-23
Color,shape,andtexturefeatureswereselectedtoforma 38-dimensional feature vector. The color and shape featuresare computed from 7 different histograms that containinformation about the red channel and green channel of theRGBimage,thehueandsaturationcomponentfromtheHSVimage, the grayscale pixel intensities and the result of the application of the sobel operator horizontally and vertically over the RGB image. Once the 7 histograms are computed,5 values of each histrogram form the first 35 features. Thevaluestakenfromthehistogramsarethemean,thestandarddeviation, kurtosis, skewness, and entropy. Finally, 3 morefeatures known as Tamura texture parameters are computedfrom the original RGB images. Those final parameters arecoarseness, contrast, and direction. All theses features areclassified in [ 15] aspartoftheirworkinautomaticdetection ofMalariaparasitesandtheyaredescribedin[ 27]. 4.1.2.Sensitivity,Specificity,and 𝐹-Measure. Inordertoeval- uate the performance of our implemented AdaBoost clas-sifier, we used three different statistical metrics: sensitivity,specificity, and effectiveness. Sensitivity is the probability ofa positive test given that the patient is ill and it is computedasfollows: sensitivity =number of true positives number of true positives +number of false negatives. (8) Specificityinturnistheprobabilityofanegativetestgiven thatthepatientiswellanditiscomputedasfollows: specificity =number of true negatives number of true negatives +number of false positives. (9)Due to the unbalanced class distribution between sub-
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-24
(9)Due to the unbalanced class distribution between sub- windowscontainingaChagasparasiteandsubwindowscon-tainingsomethingelse,theuseofstatisticalmetricsrelatedtotheeffectivenesswerealsoconsidered.Intheseexperiments,the𝐹-measureisused: 𝐹 𝛽=TPR∗PR 𝛽∗TPR+( 1−𝛽 )∗ PR(10) with TPR =TP TP+FN, PR=TP TP+FP,(11) where TP stands for the true positives, FN for the false negatives, and FP for the false positives. The 𝛽coefficient (0<𝛽<1 )allowstoassignrelativeweightstoboththetrue positiveandprecisionrates.Intheseexperiments, 𝛽 = 0.05is used,sothesearchwasaddressedtodetectionofTP. 4.2.ExperimentalResults 4.2.1. AdaBoost Results. Looking at the values provided in Table 2, we see that AdaBoost methods are as good as SVM methodsintermsofsensitivity,withmeanvalueof1andstan- dard deviation of 0. All five methods were trained to detect everyparasiteinthetestsets.Intermsofspecificity,AdaBoost+ Postprocessing is clearly the winner among the five meth-odsbeingcompared,withmeanvalueof0.9325andstandarddeviation of 0.0496, as given in Table 3. Tables4and5 also show that AdaBoost + Postprocessing is superior to theothers, in terms of 𝐹-measures. In Figure8,w ei l l u s t r a t e
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-25
t h er e s u l to fo u rA d a B o o s ta l g o r i t h mf o rC h a g a sp a r a s i t edetection.Theseresultswereobtainedwithacascadeofweakclassifiers of 11 stages. Figure9provides the ROC curves for SVMs using different degrees of polynomial kernels andFigure10 comparestheROCcurvesforAdaBoostandSVMs methodsusingdifferenttypesofkernels. 4.2.2. SVM + Feature Extraction Results. SVM results are good especially using linear and polynomial kernels. Fea-tures used to train the SVM proved to contain the neededinformation to build good classifiers. Those features werecreated from color histograms and some other values thatencapsulatetextureinformationoftheimage.Thebestresultsforeachtypeofkernelwereobtainedwiththelibrarylibsvm,finding first the best pair of parameters 𝛾and𝐶for our trainingexamples.Thoseparametersaregivenin Table 1.The AdaBoost + Postprocessing method was implemented usinga linear kernel SVM with parameters 𝛾 = 0.002 and𝐶= 0.5.Theseparameterswereobtainedthroughmultiplecross- validationexperimentsusingdifferentpairsofvaluesforthe𝛾and𝐶parameters. 4.2.3.AdaBoostversusSVM+FeatureExtractionComparison. Results obtained with AdaBoost are more robust than thoseobtained with SVM. However, the sensitivity and specificityComputationalandMathematicalMethodsinMedicine 9 (a) (b) (c) (d) (e) (f) (g) (h) (i) (j) (k) (l)
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-26
(h) (i) (j) (k) (l) Figure 8: Results of the Chagas parasite detection algorithm with a sample of images. Each column corresponds to a different example of detection. The first row contains the color images in RGB format; the second row shows the grayscale images once the parasites have been detected by the classifier; in the third row we show the result of the postprocessing stage; and the fourth row shows the final result of thedetectionsystems.10 ComputationalandMathematicalMethodsinMedicine 0 0.005 0.01 0.015 0.02 0.025 0.03 0.03500.10.20.30.40.50.60.70.80.91 False positive rateTrue positive ratePolynomial SVM ROC curves SVM_poly_ 2 SVM_poly_ 3 SVM_poly_ 4SVM_poly_ 5 SVM_poly_ 6 Figure 9: ROC curves with SVMs and polynomial kernels ranging fromdegrees2to6. 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 False positive rate SVM-linear SVM-polynomial SVM-RBFAdaBoost00.10.20.30.40.50.60.70.80.9True positive rate1 Ada+Postprocessing Figure 10: ROC curves using AdaBoost and SVMs with three differenttypesofkernel:linear,polynomial,andRBFs. reached by using SVM cannot be considered bad, given the difficultyofthetaskandtheirperformanceiscomparabletothe performance obtained in the malaria parasite detectiontask. 5. Discussion
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-27
5. Discussion B a s e do nt h er e s u l t so b t a i n e dw i t hb o t hm a c h i n el e a r n i n gmethods, AdaBoost and SVM, we can see that taking intoaccount features related to shape, color and texture seemsTable 1: Best 𝛾and𝐶parametersforSVMsaccordingtothetypeof kernelemployedandourtrainingdata. Kernel 𝛾𝐶 Linear 0.00003 8 Polynomial 0.0078 0.0312 Radialbasisfunction 0.00003 8192 Table 2: Sensitivity by method. Method Mean Std. Dev. AdaBoost 1 0 AdaBoost+Postprocessing 1 0 SVMlinear+FeatureExtraction 1 0 SVMpolynomial+FeatureExtraction 1 0SVMRBF+FeatureExtraction 1 0 T able3:Specificitybymethod. Method Mean Std. Dev. AdaBoost 0.7550 0.1820AdaBoost+Postprocessing 0.9325 0.0496SVMlinear+FeatureExtraction 0.8218 0.0118SVMpolynomial+FeatureExtraction 0.6691 0.0035 SVMRBF+FeatureExtraction 0.7111 0.0068 Table 4: 𝐹-measure( 𝛽=1)bymethod. Method Mean Std. Dev. AdaBoost 0.7550 0.1820 AdaBoost+Postprocessing 0.9325 0.0496 SVMlinear+FeatureExtraction 0.8218 0.0118
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-28
SVMlinear+FeatureExtraction 0.8218 0.0118 SVMpolynomial+FeatureExtraction 0.6691 0.0035SVMRBF+FeatureExtraction 0.7111 0.0068 Table 5: 𝐹-measure( 𝛽=1 0)bymethod. Method Mean Std. Dev. AdaBoost 0.9976 0.0018 AdaBoost+Postprocessing 0.9993 0.0004 SVMlinear+FeatureExtraction 0.9982 0.0001SVMpolynomial+FeatureExtraction 0.9967 0.0000 SVMRBF+FeatureExtraction 0.9971 0.0001 to be the most robust way to go when we target a parasite detection task. The main advantage in using AdaBoost withthe approach introduced by Viola and Jones is the fastcomputation of the features using the integral images trick,as opposed to any other method of image segmentation thatarecommonlyusedtogetherwiththeSVMapproach.Imagesegmentation is a time-consuming task and its use withoutparallel computation is almost impossible for applicationsthatrequiretoscanmanyimagesinashorttime. The consideration of DNA pixels through a trained SVM in the Ada + Postprocessing method is handled as apostprocessingprocedurebecauseitisalsoatime-consumingComputationalandMathematicalMethodsinMedicine 11 (a) (b) (c) (d) Figure11:Tworepresentativeimages.Inimage(a),thenumberofdarkpixelsmakesiteasyforthedetectionsystemtoconsideritasavalid
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-29
parasite. In image (b), the number of dark pixels is larger than average, which makes it difficult for the detection system to consider it as avalidparasite.Aneasybackgroundlookslike(c)andadifficultonelookslike(d). procedure which should be applied a reduced number of times. If we attempt to interwine this counting procedurewith the typical AdaBoost algorithm using integral images,we simply nullify the major advantage of the integral imagetrick,whichisafastcomputationofthepixelsaddition. InFigure11, we present some examples of images that arerepresentativecasesofeaseanddifficultyofclassification.When an image contains a high percentage of pixels stainedindarkpurple,asin Figure11(d) ,thealgorithmfindsitmore difficulttodiscardfalsepositives. P a r a s i t e sa n do t h e rs t a i n e do b j e c t sl i k eb l o o dc e l l sa r e nonrigid and therefore vary in shape and size. The color isimportantbutitsonlyusetodistinguishbetweentheChagasparasitesandotherthingsisnotenough.Ontheotherhand,rawimagescannotbeuseddirectlyasafeaturevectorfortwomain reasons: first the size of such a vector would increase unnecesarily the time of computation and second the high variationinshapeandsizewouldmakeitextremelydifficultfor any machine learning algorithm to learn the correctclassifier. In the ideal case, features must be able to capture shapeandcolorcharacteristicsoftheobjectstobeclassified. Failing to detect Chagas is certainly a more serious
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-30
Failing to detect Chagas is certainly a more serious situation than having a false positive error. Diagnosing apersonsufferingfromChagasasnothavingthediseasewouldp u th e rl i f ea tr i s k .Af a l s ep o s i t i v ea l a r mo nt h ec o n t r a r ywould only incur in a labor cost for the doctor in chargeof confirming whether the parasite is present in the bloodsample. Inpractice,thedetectionof T.cruziby analysis of digital images obtained from peripheral blood smears would bringm u l t i p l ebe n e fi t sa s soc i a t e dt oth eC h a g a sd i se a sed i a g n o s i s .First,itwouldmakepossiblethediagnosisofthediseaseinitsinitialphase,whichistheacutephase,becausethepresenceoftheparasiteinthebloodwouldindicatewithhighprobabilitythat the infection is present. As a consequence of the early diagnosis, the human cases could be treated with drugs such as Nifurtimox or Benznidazole, and the probability ofstopping the disease is very high. Second, given that human12 ComputationalandMathematicalMethodsinMedicine c a s e so fC h a g a sa r ev e r yd i ffi c u l tt ob ed i a g n o s e dd u r i n g the analysis routinely performed in national laboratories incountries such as Mexico, an automated Chagas detectionsystem would contribute to estimate the local impact of thediseaseandtodeterminetheexistenceofanyendemicregion. Moreover, the successful implementation of this algo- rithm, would motivate the laboratories and hospitals to
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-31
rithm, would motivate the laboratories and hospitals to reconsider the usefulness of the blood smears analysis andtotrustthistestasaconfidentlaboratorydiagnosis.Further-more, the laboratory technicians would reduce the numberof hours they need to sit in front of a microscope to analyzeso many samples day after day. Situation that is currentlyoriginatingthemproblemssuchasvisualfatigueandfrequentpainintheback. Finally, this algorithm can be integrated as part of one automated microscopic system for the adquisitionof imagesfrom blood smears. Such a device would be composed ofoneopticalmicroscopewithacontrollableelectromechanicals u r f a c es p e c i a l l yd e s i g n e dt om o v et h eb l o o ds m e a r sa sneeded by the scanning software. The microscope would beequipped with a high-resolution camera in order to take the images.Imageswouldbestoredinharddrivestoallowtheir posterioranalysiswithourChagasdetectionalgorithm.Thisdevice is currently being built as part of one CONACYTresearch project on Chagas led by Dr. Ruiz-Pi ˜na in Mexico (projectcode:salud-2009-01-113848). 6. Conclusions In this paper, we have provided an approach to the Chagasparasite detection problem based on AdaBoost learning. UsingtheapproachusedbyViolaandJonesforthetaskofface detection,weobtainedhighsensitivityandspecificityvalues. Chagasdetectionapproachesbasedonmachinelearningand computer vision methods have been barely studied as it is evidenced by the lack of literature on the topic. The mostpromising method proposed for the detection of malaria parasiteswasimplementedandcomparedtoAdaBoost.This
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-32
parasiteswasimplementedandcomparedtoAdaBoost.This method consists basically in the use of different kinds ofvalues taken from color histograms in order to form the f e a t u r ev e c t o rt h a ta r el a t e ru s e dt ot r a i nas u p p o r tv e c t o r m a c h i n e .B o t hs h a p ea n dc o l o rp r o v e dt ob ei m p o r t a n ti n the task of Chagas parasite detection when high level of sensitivityandspecificityarerequired.ApplyingSVMforthetask of Chagas parasite detection requires the computation of the 38-dimensional feature vector for each subwindow that we want to classify as being a parasite or nonparasite. However, if we proceed in this way with every subwindow of the original image of size 256 ×256, the whole process would be highly time-comsuming. One way to reduce thenumber of subwindows that are worth checking is by meansofsegmentation.Segmentationallowsustoidentifydifferentrelevantobjectsintheimageanddiscardthosethatdoesnotlook like parasites. Given that most of the objects detectedduringthesegmentationarediscarded,onlyaminornumberof subwindows are further check by the SVM. Although theuse of segmentation is very common for this kind of tasks,it can also be a time-consuming process. AdaBoost appliedwithintegralimagesdoesnotrequiresegmentation,because everysubwindowcanbecheckedinconstanttime,regardlessofitssize. Conflict of Interests The authors declare that there is no conflict of interestsregardingthepublicationofthispaper. Acknowledgment
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-33
Acknowledgment The authors thank the Universidad Autonoma de Yucatan,f o rs u p p o r t i n gt h er e s e a r c hp r o j e c to nC h a g a sp a r a s i t e sdectectionusingmachinelearningalgorithms. References [1] World Health Organization, “Chagas disease (American try- panosomiasis),” Fact Sheet 340, WHO, 2010, http://www.who .int/mediacentre/factsheets/fs340/en/ . [2] L. V. Kirchhoff, Chagas Disease (American Trypanosomiasis) , eMedicine,2010. [3] L. V. Kirchhoff, “Trypanosoma species (American trypanoso- miasis,Chagas’disease):biologyoftrypanosomes,”in Principles and Practice of Infectious Diseases ,G.L.M a ndell ,J .E.Bennet t, andR.Dolin,Eds.,ElsevierChurchillLivingstone,Philadelphia, Pa,USA,7thedition,2009. [4] C. Ponce, E. Ponce, E. Vinelli et al., “Validation of a rapid and reliable test for diagnosis of Chagas’ disease by detection ofTrypanosoma cruzi -specific antibodies in blood of donors and patients in Central America,” J o u r n a lo fC l i n i c a lM i c r o b i o l o g y , vol.43,no .10,pp .5065–5068,2005. [5] World Health Organization, “Control of Chagas disease: secondreportoftheWHOexpertcommittee,”WHOTechnical
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-34
secondreportoftheWHOexpertcommittee,”WHOTechnical ReportSeries905,WHO,2002, http://www.who.int/iris/handle/ 10665/42443 . [6] “Brazilian consensus on Chagas disease,” Revista da Sociedade BrasileiradeMedicinaTropical ,vol.38,no .3,pp .7 –29 ,2005. [7] P. Viola and M. Jones, “Rapid object detection using a boosted cascadeofsimplefeatures,”in ProceedingsoftheIEEEComputer SocietyConferenceonComputerVisionandPatternRecognition(CVPR’01) ,vol.1,pp.I-511–I-518,December2001. [8] V. Uc-Cetina, C. Brito-Loeza, and H. Ruiz-Pi ˜na, “Chagas par- asites detection through Gaussian discriminant analysis,” AbstractionandApplication ,vol.8,pp .6–17 ,2013. [9] R.Soberanis-Mukul,V .Uc-Cetina,C.Brito-Loeza,andH.Ruiz- Pi˜na,“Anautomaticalgorithmforthedetectionof Trypanosoma cruziparasitesinbloodsampleimages,” ComputerMethodsand ProgramsinBiomedicine ,vol.112,no .3,pp .633–639 ,2013. [10]M.I.Khan,B.Acharya,B.K.Singh,andJ.Soni,“Contentbased imageretrievalapproachesfordetectionofmalarialparasitein bloodimages,” InternationalJournalofBiometricsandBioinfor- matics,vol.5,no.2,pp.97–110,2011.
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-35
matics,vol.5,no.2,pp.97–110,2011. [11] G. D ´ıaz, F. A. Gonz ´alez, and E. Romero, “A semi-automatic method for quantification and classification of erythrocytes infectedwithmalariaparasitesinmicroscopicimages,” Journal ofBiomedicalInformatics ,vol.42,no .2,pp .296–307 ,2009 . [12] F. B. Tek, A. G. Dempster, and I. Kale, “Malaria parasite detec- tion in peripheral blood images,” in Proceedings of the 17th British Machine Vision Conference (BMVC ’06) ,p p .3 4 7 – 3 5 6 , September2006.ComputationalandMathematicalMethodsinMedicine 13 [13] S. Halim, T. R. Bretschneider, Y. Li, P. R. Preiser, and C. Kuss, “Estimating malaria parasitaemia from blood smear images,” inProceedings of the 9th International Conference on Control, Automation,RoboticsandVision(ICARCV’06) ,pp .1 –6,Singa- pore,December2006. [ 1 4 ]C .d iR u b e r t o ,A .D e m p s t e r ,S .K h a n ,a n dB .J a r r a ,“ A n a l y s i s of infected blood cell images using morphological operators,” Image and Vision Computing ,vol.20,no .2,pp .133–146,2002. [15] N.E.Ross,C.J.Pritchard,D.M.Rubin,andA.G.Dus ´e,“Auto-
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-36
mated image processing method for the diagnosis and classifi-cationofmalariaonthinbloodsmears,” Medical and Biological EngineeringandComputing ,vol.44,no .5,pp .427 –436,2006. [16] S. W. S. Sio, W. Sun, S. Kumar et al., “MalariaCount: an image analysis-based program for the accurate determination of par- asitemia,” Journal of Microbiological Methods ,v o l .68 ,n o .1 ,p p . 11–18,2007. [17] Y .Purwar,S.L.Shah,G.Clarke,A.Almugairi,andA.Muehlen- bachs, “Automated and unsupervised detection of malarialparasitesinmicroscopicimages,” MalariaJournal ,vol.10,article 364,2011. [18] S.Kaewkamnerd,A.Intarapanich,M.Pannarat,S.Chaotheing, C. Uthaipibull, and S. Tongsima, “Detection and classification deviceformalariaparasitesinthick-bloodfilms,”in Proceedings of the IEEE 6th International Conference on Intelligent Data AcquisitionandAdvancedComputingSystems(IDAACS’11) ,vol. 1,pp.435–438,Prague,CzechRepublic,September2011. [19] J.-P.ThiranandB.Macq,“Morphologicalfeatureextractionfor the classification of digital images of cancerous tissues,” IEEE TransactionsonBiomedicalEngineering ,vol.43,no.10,pp.1011– 1020,1996. [20] Y.K.LinandK.S.Fu,“Automaticclassificationofcervicalcells
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-37
[20] Y.K.LinandK.S.Fu,“Automaticclassificationofcervicalcells usingabinarytreeclassifier,” Pattern Recognition ,vol.16,no .1, pp .69–80,1983. [21] P. Viola and M. J. Jones, “Robust real-time face detection,” InternationalJournalofComputerVision ,vol.57 ,no .2,pp .137 – 154,2004. [22] M.R ¨atsch,S.Romdhani,andT.Vetter,“Efficientfacedetection byacascadedsupportvectormachineusinghaar-likefeatures,”inProceedings of the 26th Annual Symposium of the German A s s o c i a t i o nf o rP a t t e r nR e c o g n i t i o n , pp. 62–70, Tuebingen, Germany,2004. [23] R.E.Schapire,Y .Freund,P .Bartlett,andW .S.Lee,“Boostingthe margin: a new explanation for the effectiveness of votingmethods,”in Proceedingsofthe14thInternationalConferenceon MachineLearning ,1997 . [24] V.Vapnik, StatisticalLearningTheory ,Wiley,1998. [25] O.Mudanyali,C.Oztoprak,D.T seng,A.Erlinger,andA.Ozcan, “Detection of waterborne parasites using field-portable and cost-effective lensfree microscopy,” L a bo naC h i p :M i n i a t u r i - sationforChemistryandBiology ,vol.10,no.18,pp.2419–2423, 2010. [26] C.M.Bishop, PatternRecognitionandMachineLearning ,Sprin- ger,2006.
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-38
ger,2006. [27] S.TheodoridisandK.Koutroumbas, Pattern Recognition ,A ca- demicPress,SanDiego,Calif,USA,1999.Submit your manuscripts at http://www.hindawi.com Stem Cells International Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014M ED IATO R S INFLAM MATIONof Hindawi Publishing Corporation http://www.hindawi.com Volume 2014Behavioural Neurology EndocrinologyInternational Journal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014Disease Markers Hindawi Publishing Corporation http://www.hindawi.com Volume 2014BioMed Research International OncologyJournal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014Oxidative Medicine and Cellular Longevity Hindawi Publishing Corporation http://www.hindawi.com Volume 2014PPAR Research The Scientific World Journal Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Immunology Research Hindawi Publishing Corporation http://www.hindawi.com Volume 2014Journal of ObesityJournal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Computational and Mathematical Methods in Medicine OphthalmologyJournal of Hindawi Publishing Corporation
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf
2d87191d9618-39
Mathematical Methods in Medicine OphthalmologyJournal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Diabetes ResearchJournal of Hindawi Publishing Corporation http://www.hindawi.com Volume 2014 Hindawi Publishing Corporation http://www.hindawi.com Volume 2014Research and TreatmentAIDS Hindawi Publishing Corporation http://www.hindawi.com Volume 2014Gastroenterology Research and Practice Hindawi Publishing Corporation http://www.hindawi.com Volume 2014Parkinson’s Disease Evidence-Based Complementary and Alternative Medicine Volume 2014Hindawi Publishing Corporation http://www.hindawi.com
https://downloads.hindawi.com/journals/cmmm/2015/139681.pdf