Research Article Classifying Dementia Using Local …downloads.hindawi.com/journals/ijbi/2015/572567.pdfResearch Article Classifying Dementia Using Local Binary ... e paper is organised

Research ArticleClassifying Dementia Using Local Binary Patterns fromDifferent Regions in Magnetic Resonance Images

Ketil Oppedal12 Trygve Eftestoslashl1 Kjersti Engan1 Mona K Beyer34 and Dag Aarsland25

1Department of Electrical Engineering and Computer Science University of Stavanger 4036 Stavanger Norway2Centre for Age-Related Medicine Stavanger University Hospital Stavanger Norway3Department of Radiology and Nuclear Medicine Oslo University Hospital Oslo Norway4Department of Life Sciences and Health Faculty of Health Sciences Oslo and Akershus University College of Applied SciencesOslo Norway5Alzheimerrsquos Disease Research Centre Karolinska Institutet (KI) Stockholm Sweden

Correspondence should be addressed to Ketil Oppedal ketiloppedalgmailcom

Received 18 December 2014 Revised 26 February 2015 Accepted 2 March 2015

Academic Editor Yantian Zhang

Copyright copy 2015 Ketil Oppedal et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Dementia is an evolving challenge in society and no disease-modifying treatment exists Diagnosis can be demanding and MRimaging may aid as a noninvasive method to increase prediction accuracy We explored the use of 2D local binary pattern (LBP)extracted from FLAIR and T1 MR images of the brain combined with a Random Forest classifier in an attempt to discern patientswith Alzheimerrsquos disease (AD) Lewy body dementia (LBD) and normal controls (NC) Analysis was conducted in areas with whitematter lesions (WML) and all of white matter (WM) Results from 10-fold nested cross validation are reported as mean accuracyprecision and recall with standard deviation in bracketsThe best result we achieved was in the two-class problemNC versus AD +LBD with total accuracy of 098 (004) In the three-class problem AD versus LBD versus NC and the two-class problem AD versusLBD we achieved 087 (008) and 074 (016) respectivelyThe performance using 3DT1 images was notably better than when usingFLAIR imagesThe results from theWMregion gave similar results as in theWML region Our study demonstrates that LBP textureanalysis in brain MR images can be successfully used for computer based dementia diagnosis

1 Introduction

Dementia Is an Evolving Challenge As a result of increasingage dementia is an evolving challenge in society The annualhealth care costs related to dementia were estimated to $604billion worldwide in 2010 [1] Alzheimerrsquos disease (AD) isthemost common neurodegenerative dementia and accountsfor 50ndash60 of people with dementia [2] The classicalneuropathological signs of AD are amyloid plaques andneurofibrillary tangles [3] No efficient disease-modifyingtreatment for AD exists today Dementia with Lewy-bodies(DLB) together with dementia associated with Parkinsonrsquosdisease (PDD) account for 15ndash20 of people with dementia[2] The defining pathological feature for these patients isLewy-body degeneration in brain stem forebrain and limbicand cortical structures and the DLB and PDD are thereforeoften combined into a Lewy-body dementia group (LBD)[4 5] However the relationship between localization and

density of Lewy-bodies with clinical dementia symptoms isnot strong [6] suggesting that other pathologies contribute aswell such as AD pathology and vascular brain changes seenas white matter hyperintensities (WML) or lacunar infarctswhichmay contribute to the clinical presentation of LBD Forexample vascular changes in the basal ganglia are commonin the elderly and may cause parkinsonism and cognitiveimpairment [7]

Early Diagnosis Is Important AD and LBD are very complexdiseases making them difficult to be prevented delayed orcured Current therapy focuses on many approaches forexample helping patients maintain an acceptable mentalfunctioning managing typical behavioural changes andslowing symptom progression Early intervention is impor-tant and the ability to identify these types of dementiaand healthy controls early in the disease course may beessential for successful patient care Differentiating between

Hindawi Publishing CorporationInternational Journal of Biomedical ImagingVolume 2015 Article ID 572567 14 pageshttpdxdoiorg1011552015572567

2 International Journal of Biomedical Imaging

AD and LBD is also important since they differ in prog-nosis and response to drug treatment Currently the onlyavailable method to differentiate between AD and LBD isthe dopamine transporter scan which is expensive and notreadily available at all centres

Neuroimaging in Dementia Neuroimaging is an importanttool for studying dementia and cognitive deterioration Sev-eral excellent reviews are available [8ndash10] In [11] Malloyet al review available methods for quantitative imaging ofwhitematter anatomy andpathology aswell as recent findingsin ageing and dementia They state that computer aidedquantification offers better statistical power compared tovisual rating scales and that diffusion imaging is able to detectabnormalities not recognised in conventional acquisitionsequences Early detection of disease and relevant functionalconnections between brain areas are important benefits

Computer Aided Diagnosis in Dementia Computer aideddiagnosis (CAD) can be a helpful tool to pinpoint diagnosisearly in the disease course in a cost-effective manner andunbiased to human inconsistencies [12] Recent advances inthe field have focused especially on AD and patients withmild cognitive impairment (MCI) which are considered aprecursor to AD [13ndash16] Less attention has been put intodeveloping CAD systems for LBD As mentioned above LBDhave high prevalence and accurate clinical diagnosis dependson little available and expensive dopamine transporter scanand postmortem histology Few papers report high accuracydiscerning patients with AD and LBD or other dementias [1718] A promising approach is reported in [19] where Lebedevet al use sparse partial least squares (SPLS) classification ofcortical thickness measurements reporting a sensitivity of944 and a specificity of 8889 discerning AD from LBD

White Matter and White Matter Lesions in Dementia Whitematter (WM) comprises approximately half the brain volumeand provides connectivity between the two brain hemi-spheres as well as ensuring efficient transfer of neural activitycomplementing information processing in the gray matter(GM) WM neuropathology is often diffuse and affects manyneuronal networks which can be disturbed simultaneouslyresulting in a multidomain syndrome In [20] Filley empha-sizes the contribution of white matter disease (WMD) inmild cognitive dysfunction cognitive ageing and dementiaBartzokis [21] proposes a hypothesis for AD called the ldquomye-lin modelrdquo where axonal transport disruption formationof axonal swellings neuritic plaques and proteinaceousdeposits such as A120573 and tau are by-products of homeostaticmyelin repair processes Gunning-Dixon et al [22] reviewresults ofMRI studies of whitematter changes that occurwithnormal ageing and the relationship of age-associated changesin white matter to age-related declines in cognitive abilities

White matter lesions or white matter hyperintensities(WML) are among the neuroimaging expressions of cerebralsmall-vessel disease and are associated with various distur-bances with poor prognosis [23] WML are localized areasof increased signal intensities in the white matter of thebrain visible on T2-weighted MR images The underlyingpathology of WML is heterogeneous ranging from mild

demyelination to incomplete subcortical infarctionsThey aretypically seen around the ventricles (periventricular WML)but also as focal lesions in the deep white matter In theelderly WML usually represent small-vessel cerebrovasculardisease (CVD) [24] WML becomes more abundant withincreasing age in healthy subjects [25] but they are alsofound to be associated with AD [26] and other dementias[27 28] Clinical symptoms associated with WML includegait disturbances [29] depression [27 30] and cognitiveimpairment [31] although the exactmechanisms are not fullyunderstood In [32] Tuladhar et al concludes that corticalchanges mediated by WML and vascular risk factors mightlead to cognitive decline and dementia MunozManiega et al[33] write that age-related deterioration of normal appearingwhitematter (NAWM) is strongly associatedwith the severityof WML indicating that WML is important in dementiaresearch Fujishima et al conclude in [34] that mild cognitiveimpairment poor episodic memory and late-life depressionare associated with cerebral cortical thinning and WML

Texture Analysis in Neuroimaging Harrison [35] extensivelyreviews the use of texture analysis in a clinical contextanalysing MR images in non-Hodgkin lymphoma mildtraumatic brain injury and multiple sclerosis She concludesthat ldquonon visible lesions and physiological changes as well asvisible focal lesions of different aetiologies could be detectedand characterized by texture analysis of routine clinical 15Tesla scansrdquo The application of texture analysis in a machinelearning (ML) environment has shown success in discern-ing different dementias from each other and from healthycontrols Freeborough and Fox reported a classification rateof 91 discerning AD from healthy controls using measuresfrom a spatial gray-level dependent method applied in astepwise discriminant analysis approach in [36] de Olivieiraet al [37] found statistical significant differences in graylevel cooccurrence matrix measurements in subjects withmild AD amnestic mild cognitive impairment (aMCI) andhealthy controls Zhang et al [38] performed 3D textureanalysis in MR images of the hippocampus and entorhinalcortex in AD patients and achieved a classification accuracybetween 643 and 964 Sivapriya et al showed in [39]that texture analysis in brain MRI gave high classificationaccuracy in AD As of the authors knowledge the only paperconsidering texture analysis as an approach to distinguishAD LBD and NC is [40] where Kodama and Kawaseperformed discriminant analysis on features extracted froma cooccurrence matrix and a run-length matrix with anaccuracy of 917 700 and 880 respectively

Local binary pattern (LBP) was introduced by Ojala et al[41 42] as a texture descriptor It is a simple yet very efficienttexture operator which labels the pixels of an image bythresholding the neighbourhood of each pixel and considersthe result as a binary number Unay et al [43] showed thatthe rotation invariant LBP is invariant to some commonMRIartefacts which makes it a robust texture feature when usedin brain MR image analysis

Aims We have earlier shown that there were no differencesin WML volume between patients with AD and LBD or

International Journal of Biomedical Imaging 3

between a combined dementia group (AD + LBD) andhealthy controls in the DemVest study [44] Now we wantto test if the WML regions inherit textural information inan extent that can be used to classify dementia patientsfrom normal controls and AD from LBD As the detectionof textural information in WML might not be dependenton an exact delineation of WML we also want to test if acomparable classification accuracy can be achieved using allof WM as ROI since WM segmentation is more availableand only a 3DT1 MR image is needed which is commonlyacquired in a clinical setting

Earlier we have shown that using LBP texture analysis inWML regions in FLAIR MR images in a machine learning(ML) context can discern patients with dementia fromhealthy controls with high accuracy [45] We want to testdifferent types of LBP calculations together with a contrastmeasure (C) calculated from FLAIR and 3DT1 MR imagesfrom a cohort study (the DemWest and ParkWest study) andon a subset containing data from one scanner only

Because of the challenging situation with imbalanceddata having different numbers of subjects in the representedgroups in the abovementioned cohorts we want to test howthe use of resampling of instances affects classification results

Organisation of Paper The paper is organised as followsSection 2 describes the data material Section 3 describesthe image preprocessing procedures followed by Section 4which describes the image processing methods and Section 5describing the experimental setup Section 6 reveals theresults Section 7 discusses the results and ends the paperwitha conclusion

2 Material

21 Subjects MR images of dementia subjects included inthis study were drawn from the DemWest cohort StavangerNorway and MR images of the healthy controls from theParkWest cohort Stavanger Norway Inclusion and exclusioncriteria can be found in [2] and [46] respectively Thedementia and healthy control subjects were matched for sexage and years of education

The Regional Committee for Medical Research EthicsWestern Norway approved the study All participants signedinformed consent to participate in the study after the studyprocedures had been explained in detail to the patient and acaregiver usually the spouse or offspring

22MRI Thedementia patients were scanned at three differ-ent sites Stavanger University Hospital Stavanger NorwayHaugesund Hospital Haugesund Norway and HaraldsplassDeaconess Hospital Bergen Norway A 15 T scanner wasused in all three centres (Philips Intera in Stavanger andHaugesund and GE Signa Excite in Bergen) using the samescanner in each centre during the entire study period and acommon study imaging protocol

The NCs were scanned at four different sites They werescanned on the same scanners as the patients in Stavangerand Haugesund Additionally MR images of NC subjectswere acquired from Soslashrlandet Hospital Arendal Arendal

Norway (10T Philips Intera) and Unilabs Bergen Norway(15T Siemens Symphony)

After visual inspection some patient scans were excludeddue to either insufficient image quality not having bothFLAIR and T1 images for the patient or movement and otherartefacts

A total of 73 mild dementia subjects 57 with AD and16 with LBD had MRI scans of sufficient quality and wereincluded in this study as well as 36 healthy controls In [44]further clinical details as well as MR imaging parameters canbe found

To ensure high reliability between scans acquired atdifferent centres and at different time points three volunteerswere scanned at all centres using the same scanners andprotocols Details of the procedure can be found in [44]Cronbachs alpha between MR scanners at different centreswas calculated based on total brain volume and was reportedto be 0958 Cronbachs alpha between two time points variedbetween 0982 and 0995 indicating excellent reliabilitiesboth between centres and between different time pointsSimilar results were reported for the MR images of the NCsfrom the ParkWest study

3 Image Preprocessing

31 Region of Interest Extraction Two common approachesfor MR image segmentation of the brain are tissue classifi-cation and template registration In the tissue classificationapproach voxels are assigned to a class based on the classvoxel intensity distribution In the template registrationapproach a template image with predefined classes is warpedto the actual MR image In our study WM partitions weresegmented using the common functions in SPM8 on the T1imagesTheprocedure unifies a tissue segmentation approachwith a template registration method see [47] for furtherdetails

WML segmentation was performed according to amethod developed and previously published by Firbank etal in Newcastle England [48] The method is based ondetermining a threshold value from the image gray scaleintensity values and then classifying the hyperintense voxelsas WML Briefly the nonbrain regions were removed fromthe T1 image using the segmentation routines in the softwarepackage SPM5 [49] After transforming to the image space ofthe FLAIR image the segmented T1 imagewas used as amaskfor scull stripping of the FLAIR image Then the WML weresegmented automatically on a slice-by-slice basis from theFLAIR images with the images in native space A scale factordetermined experimentally wasmultiplied by themode of thehistogram of pixel intensities for each image slice and usedas a threshold value for WML segmentation To explore theregional distribution of WML throughout the brain a regionof interest (ROI) template in standard MNI space [50] wasused This ROI template was transformed from MNI spaceto the image space (FLAIR) of each subject by use of thenormalization routines in SPM5 and the volumes of WMLin each ROI were calculated The ROI map was based on theBrodmann template Further details can be found in [51]


Because of the variability in MR image quality acquiredfrom the different centres participating in this study a scalefactor that gave an overestimation of the lesion load in everysubject was selected and manual editing was then done tocorrect this by removing excess pixels using FSLView [52]a medical image-editing program being a part of the FSLsoftware bundle A medical doctor did the manual editingafter training by a consultant radiologist who is experiencedat evaluation of WML We performed inter- and intraraterreliability testing between the two raters to ensure good qual-ity They both edited the same 10 data sets twice once in thebeginning to ensure good interrater reliability and a secondtime at the end to ensure that similar reliability still persistedand to evaluate intrarater reliability Intraclass correlationcoefficient (ICC) was 0998 for interrater reliability and 0964for intrarater reliability

4 Image Processing Methods

41 LBP Ojala et al [41 42] introduced LBP as a textureoperator Since its discriminative power is high and at thesame time computationally simple LBP is a popular texturedescriptor used in various applications and unifies tradi-tionally divergent statistical and structural models of textureanalysis Adding an image contrast measure (C) calculatingthe local variance in the pixel neighbourhood as well as vary-ing the texture neighbourhood enhances the discriminativepower of the LBP feature even further In [43] Unay et aldemonstrated that the rotation invariant LBP is invariant tosome common MRI artefacts that is the bias field

The derivation of the gray scale and rotation invarianttexture operator LBP starts by defining texture 119879 in a localneighbourhood of a monochrome texture image as the jointdistribution of the gray levels of 119875 (119875 gt 1) image pixels

119879 = 119905 (119892119888 1198920 119892

119875minus1) (1)

where gray value 119892119888corresponds to the gray value of the

center pixel of the local neighbourhood and119892119901(119901 = 0 119875minus

1) corresponds to the gray value of 119875 equally spaced pixels ona circle of radius 119877 (119877 gt 0) that form a circularly symmetricneighbour set When the coordinates of 119892

119888are (0 0) the

coordinates of119892119901are given by (minus119877 sin(2120587119901119875) 119877 cos(2120587119901119875))

and the gray values of neighbours which do not fall exactly inthe center of pixels are estimated by interpolation

To achieve gray-scale invariance the gray value of thecenter pixel (119892

119888) is subtracted from the gray values of the

circular symmetric neighbourhood 119892119901(119901 = 0 119875 minus 1)

giving119879 = 119905 (119892

119888 1198920minus 119892119888 1198921minus 119892119888 119892

119875minus1minus 119892119888) (2)

By assuming that differences 119892119901minus 119892119888are independent of 119892

119888

and thereby factorizing we get119879 asymp 119905 (119892

119888) 119905 (1198920minus 119892119888 1198921minus 119892119888 119892

119875minus1minus 119892119888) (3)

The distribution 119905(119892119888) describes the overall luminance of the

image and is unrelated to local image texture and is removedThe approximated distribution

119879 asymp 119905 (1198920minus 119892119888 1198921minus 119892119888 119892

119875minus1minus 119892119888) (4)

conveysmany of the textural characteristics from the original

By considering just the signs of the differences instead oftheir exact values invariance with respect to gray-scale shiftsis achieved

119879 asymp 119905 (119904 (1198920minus 119892119888) 119904 (119892

1minus 119892119888) 119904 (119892

119875minus1minus 119892119888)) (5)

where

119904 (119909) = 1 119909 ge 0

0 119909 lt 0(6)

Each sign 119904(119892119901minus 119892119888) is assigned a binomial factor 2119901

such that 119879 is transformed into a unique LBP119875119877

number thatcharacterizes the spatial structure of the local image texture

LBP119875119877

=119875minus1

sum119901=0

119904 (119892119901minus 119892119888) 2119875 (7)

See also Figure 1To assign a unique identifier to each rotation invariant

local binary pattern LBP119903119894119875119877

is defined as

LBP119903119894119875119877

= min ROR (LBP119875119877 119894) | 119894 = 0 1 119875 minus 1 (8)

where ROR(119909 119894) performs a circular bitwise right shift on the119875-bit number 119909 119894 times

Certain local binary patterns are fundamental propertiesof texture ldquoUniformrdquo patterns are circular structures thatcontain very few spatial transitions They function as tem-plates for microstructures such as bright spot flat area darkspot and edges of varying positive and negative curvatureThe uniformity relates to the number of spatial transitions(ie bitwise 01 changes) in the LBP pattern for exam-ple 00000000

2and 11111111

2have a uniformity value

119880(ldquo119901119886119905119905119890119903119899rdquo) of 0 and 000000112and 10000111

2of 1 and

2 respectively Patterns that have a 119880 value of at most 2 aredesignated as ldquouniformrdquo A gray-scale rotation invariant anduniform LBP texture operator are defined as follows

LBP1199031198941199062119875119877

=

119875minus1

sum119901=0

119904 (119892119901minus 119892119888) if 119880(LBP

119875119877) le 2

119875 + 1 otherwise(9)

where119880 (LBP

119875119877) =

1003816100381610038161003816119904 (119892119875minus1 minus 119892119888) minus 119904 (1198920 minus 119892119888)1003816100381610038161003816

+119875minus1

sum119901=1

10038161003816100381610038161003816119904 (119892119901 minus 119892119888) minus 119904 (119892119901minus1 minus 119892119888)10038161003816100381610038161003816

(10)

Superscript 1199031198941199062 reflects the use of rotation invariant ldquouni-formrdquo patterns that have a119880 value of at most 2 By definitionexactly 119875 + 1 ldquouniformrdquo binary patterns can occur in acircularly symmetric neighbour set of 119875 pixels whereas theldquononuniformrdquo patterns are grouped under a miscellaneouslabel (119875 + 1)

42 Contrast The LBP119903119894119875119877

and LBP1199031198941199062119875119877

operators are excellentmeasures of spatial patterns but discard contrast If gray-scaleinvariance is not required the contrast (119862) of local image


The value of the LBP code of a pixel (xc yc) is given by

LBPPR =Pminus1

sump=0

s(gp minus gc)2p

Sample Threshold

x = 1 lowast 20+ 1 lowast 2

1+ 0 lowast 2

2+ 0 lowast 2

3+ 0 lowast 2

4+ 1 lowast 2

5+ 0 lowast 2

6+ 1 lowast 2

7= 163

Multiply by powers of two and sum

s(x) = 1 if x ⩾ 0

0 otherwise

83

52

94

88

97

61

14

77 82

0

1

1

1

0

0

0

1

x

(a)

52

94

88

97

61

14

77

83

82

5279

94

97

88

73

67

5461

54

14

4

23

23

83

23

82

(b)

Figure 1 (a) Demonstrates LBP thresholding A neighbour to a center pixel is set to one if it has equal or higher pixel value and zero if ithas lower pixel value In an anticlockwise manner every neighbour is multiplied by powers of two and summed as demonstrated in (9) (b)demonstrates how the radius and number of samples can be varied in the choice of neighbourhood (b) Left figure with small radius and 8samples Right figure with large radius and 16 samples

texture can be measured with a rotation invariant measureof local variance defined as

VAR119875119877

=1

119875

119875minus1

sum119901=0

(119892119901minus 120583)2

where 120583 = 1

119875

119875minus1

sum119901=0

119892119901 (11)

which is invariant against shifts in gray-scaleThe LBP and 119862 values are calculated for every voxel in

the specified region of interest creating an LBP- and a 119862-valued image Typically the LBP and 119862 values are collectedand represented as a histogram for each instance in thedata set The histogram can be used as a vector of featuresOther approaches include calculating new features from thehistogram

5 Proposed Method and Experimental Setup

51 Overview of Proposed Method A computer based systemfor classification of AD LBD and healthy controls basedon texture analysis was applied Firstly the two regions of

interestWML andWMwere extracted from theMR imagesTheWMregionswere segmented using common functions inSPM8 and theWMLwere segmented from the FLAIR imagesusing the thresholding technique proposed by Firbank et al[48] as briefly described in Section 31 See also Block 1 inFigure 2

Secondly rotation invariant 2D LBP and contrast wereextracted voxel-wise for the two different ROIs using differentcombinations of neighbourhood radii and number of sam-ples The 2D LBP and contrast texture analysis were doneboth on the FLAIR and the T1 MR images (see Section 41for information concerning the calculation of the LBP texturefeature and Section 42 for the contrast measure and Block 2of Figure 2) Statistical features were calculated from all theLBP and 119862 values in each ROI were then calculated

Eventually a combined feature selection and classificationprocedure were applied using a Random Forest [53] classifiertogether with a nested cross validation procedure [54] SeeBlock 3 in Figure 2


(1)ROI extraction

(2)Feature extraction by LBP

and C

(3)Feature selection and classificationby Random Forest

Figure 2 Overview of proposed method See Section 51 for details

52 Texture Feature Extraction For the 2D texture analysisapproach the LBP values as well as the119862measure were calcu-lated from every voxel in the selected ROI andMR image typefor all subjects in the data set using Matlab [55] Three differ-ent combinations of neighbourhood radius (119877) and numberof samples (119875) namely 119877 = 1 and 119875 = 8 119877 = 2 and 119875 = 12and 119877 = 4 and 119875 = 16 were used Mean standard deviationvariationmedian interquartile range entropy skewness andkurtosis of the ROI-wise collected LBP and 119862 values werecalculated to be used as a descriptor of the distributions of theLBP and 119862 values These features were subjected to furtherselection and classification resulting in 8 features for eachof the three combinations of 119877 and 119875 for both LBP and 119862resulting in a total of 48 features for each subject See Figure 3for an example of the FLAIR andT1MR images and theWMLsegmentation results See also Figure 4 for an example of LBP-and119862-valued images based on the FLAIR and T1MR images

53 Feature Selection and Classification A challenge in thedeveloped machine learning task was the high number offeatures calculated compared to the size of the data setSince the data were collected in a cohort study and therebyinexpedient to expand a method for feature subset selectionwas needed A method combining feature selection andclassification using two nested cross validation loops togetherwith a Random Forest classifier was chosen an inner CVscheme for classification parameter and feature selection andan outer CV scheme for final model testing see Figure 5 fordetails Such an approach prevents the improper procedure ofusing the complete data set for supervised feature selectionahead of using cross validation for performance evaluationThe latter approach would give an overly optimistic result

Image data were selected with stratification during boot-strap rounds in the cross validation procedure meaning thatthe relative representation of instances in each class was keptintact

Feature selection and classification were done using a10-tree Random Forest classifier and 10-fold nested crossvalidation for performance evaluation Search method wasbest first start set with no attributes search direction forward

Table 1

Actual positive Actual negativePredicted positive TP FPPredicted negative FN TN

stale search after five node expansions subset evaluation f-measure and number of folds for accuracy estimation was 10

Pretesting was done using different classifiers includingsupport vector machines Random Forests and a Bayesiannetwork classifier The Random Forest classifier outper-formed the other classifiers and thus all experiments pre-sented in this work are conducted using Random Forestclassifiers

To give a fairly acceptable graphic display of the selectedfeatures the feature and model parameter selection wereeventually performed on the complete data set and a matrixof scatter plots displaying the five selected features pairwiseagainst each other was made see Figure 6 Note that this wasonly done for the sake of practical graphical display and as anexample of which features that typically would be selected

Random Forest is a classifier based on ensembles ofdecision trees developed by Breiman [53] Many decisiontrees are built using bootstrap aggregation (bagging) andrandomized feature subset sampling where the mode of theclasses output by individual trees is voted for

Three separate tests were explored a three-class approachclassifying NC versus AD versus LBD a two-class approachclassifying a NC versus a combined dementia group (AD +LBD) and another two-class approach classifying AD versusLBD

54 Classification Accuracy Precision for a class is the frac-tion of instances that are correctly classified to all instancesthat are classified as this class and is also known as positivepredictive value Recall for a class is the fraction of instancesthat are correctly classified to all the instances that reallybelong to this class and is also known as true positive rate orsensitivity

In the context of a two-class problem where one class isthe positive class and the other is the negative class the truepositives (TP) are the instances that are correctly classified asbelonging to the positive class and the false positives (FP) arethe instances that are classified as the positive class but reallybelong to the negative classThe true negatives (TN) and falsenegatives (FN) can be explained similarly An overview ofresults can be presented as a confusion matrix see Table 1

Precision is then defined as Precision = TP(TP+FP) andrecall is defined as Recall = TP(TP + FN) Total accuracy(119879) precision (119875) and recall (119877) were calculated for eachof the ten folds in the cross validation procedure resultingin ten values for each (119879

1 1198792 119879

10) (1198751 1198752 119875

10) and

(1198771 1198772 119877

10) Empirical mean over the ten values was

calculated using the equation below

119898119909=1

119899

119899

sum119896=1

119909119896 where 0 le 119909

119896le 1 0 le 119898

119909le 1 (12)


(a) (b)

(c) (d)

Figure 3 Overview ofMR images and the ROIs used for feature extraction (a) in the top left corner shows an example of an axial FLAIRMRimage The white matter lesions are possible to see as hyperintense areas (b) in the top right shows the segmented voxels labelled as WMLoverlayed on the FLAIRMR image seen in (a) (c) in the bottom left corner shows the segmentedWML voxels found from the correspondingFLAIR overlayed on the T1 MR image (d) in the bottom right corner shows the segmented WM voxels overlayed on the T1 MR image

where119909 is either119879119875 or119877 and 119899 = 10The empirical standarddeviation was calculated as below

119904119909= (

1

119899 minus 1

119899

sum119896=1

(119909119896minus 119898119909119896)2

)

12

where 0 le 119909119896le 1 0 le 119904

119909le 1

(13)

where 119909 is either 119879 119875 or 119877 119899 = 10 and 119898119909is defined as in

(12) 119879 119875 and 119877 are reported as119898(119904) over 10-fold CV

55 Imbalanced Data Set The data set used in this study wasdrawn from a cohort A common drawback is the problem of

imbalanced data meaning that the data set contains groupsof different sizes Typically machine learning algorithms willperform poorly under such circumstances As a measure toprevent such a problem a resampling technique was used toeven out the sizes of the groups

All tests were done using the Synthetic Minority Over-sampling Technique (SMOTE) [56 57] to resample data suchthat all classes had the same number of instances and aresimilar to the largest class in the original data Similar testswere done without resampling as well and in all of thecases the classification accuracy for the LBD class improvedusing SMOTE at the expense of classification accuracy forthe other classes Total accuracy was either improved or at


(a) (b)

(c) (d)

Figure 4 Overview of LBP texture- and contrast-valued images based on the FLAIR and T1 images (a) in the top left corner shows anexample of an LBP-valued image calculated from a FLAIR MR image (b) in the top right corner shows a contrast-valued image calculatedfrom FLAIRMRI (c) in the bottom left corner shows an LBP-valued image calculated from a T1 image (d) in the bottom right corner showsa contrast-valued image calculated from a T1 MR image

least preserved In conclusion balancing out the numberof instances in each class in the data set balanced out theclassification performance for each class as well

6 Results

61 Three-Class Problem NC versus AD versus LBD Resultsfor the three-class problemwith class 0 beingNC class 1 beingAD and class 2 being LBD are shown in detail in Table 2 119879is the total accuracy for all three classes 1198750 is the precisionfor the NC group 1198751 is the precision for the AD group 1198752 isthe precision for the LBD group 1198770 is the recall for the NCgroup 1198771 is the recall for the AD group and 1198772 is the recallfor the LBD group

The first test named FLAIR-WML119903119894indicates that the

FLAIRMR image was used for calculation of LBP and119862 that

WMLwas theROI and that the rotational invariant variant ofthe LBP feature was used The second test named T1WML

119903119894

indicates that the T1 MR images were used for calculation ofLBP and 119862 that WML was the ROI and that the rotationalinvariant variant of the LBP feature was used The third testnamed T1WM

119903119894indicates that the T1 MR images were used

for calculation of the LBP and 119862 that the WM was the ROIand that the rotational invariant variant of the LBP featurewas used

The total accuracy showed great variation throughout thedifferent tests ranging from 06 (013) to 087 (008) Theperformance increased considerably when calculating theLBP and 119862 features from the T1 MR image as compared tothe FLAIRMR imageThe classification performance provedbest in the T1 case and when WML was used as ROI

For comments on the T1WMLsvg119903119894 test see Section 64


Outer loop Training data

Training data

Test data

Test dataInner loop

Test classifierparameters and

feature sets

Construct classifier model Predict

Buildclassifieron outertraining

andpredictoutertest

results

Select the best classifier modelparameters and final feature set

based on overlap over bootstrap roundsCollect success

measures

Figure 5 Nested cross validation in the inner loop the performance of different sets of classifier parameters and features is estimated basedon a bootstrap cross validation The optimal classifier parameters and features are selected based on the performance evaluation over severalbootstrap rounds In the outer loop model performance of the optimized classifier parameters and features is evaluated on the hold-out testset in the outer loop The outer loop is repeated several times every time with potentially different classifier parameters and features

Table 2 Results are reported as mean with standard deviation in brackets 119898(119904) over 10-fold cross validation classifying NC versus ADversus LBD 119879 = total accuracy 119877 = recall and 119875 = precision 0 for class NC 1 for class AD and 2 for class LBD ROI is either WM for whitematter or WML for white matter lesion area

Test 1198791198750 1198751 1198752

1198770 1198771 1198772

FLAIR-WML119903119894

060 (013) 071 (028) 061 (014) 033 (041)048 (025) 077 (028) 020 (035)

T1WML119903119894

082 (012) 096 (010) 080 (011) 058 (049)098 (008) 088 (018) 025 (035)

T1WMLSMOTEri 087 (008) 097 (007) 081 (017) 085 (011)

100 (000) 082 (016) 078 (020)

T1WM119903119894

082 (009) 096 (008) 081 (011) 042 (049)100 (000) 088 (016) 020 (035)

T1WMSMOTE119903119894

075 (013) 090 (012) 066 (016) 070 (021)100 (000) 072 (019) 055 (022)

T1WMLSMOTEsvg119903119894 091 (015) 100 (000) 100 (000) 087 (022)

100 (000) 077 (042) 100 (000)

62 Two-Class Problem NC versus AD+ LBD Results for thetwo-class problemwith class 0 being NC and class 1 being ADand LBD together are shown in detail in Table 3 119879 is the totalaccuracy for the two classes 1198750 is the precision for the NCgroup and 1198751 is the precision for the combined AD and LBDgroup 1198770 is the recall for the NC group and 1198771 is the recallfor the combined AD and LBD group

In addition to the abovementioned tests another testnamed T1WML

1199031198941199062was applied to assess whether the classi-

fication performance would differ when rotational invariantLBP were calculated alone or in combination with selectionof uniform LBP values only

Total accuracy is generally higher in the T1 case (rangingfrom 097 (004) to 098 (004)) compared to the FLAIR case


80 100 120 140 160

80

100

120

140

160

Var(

LBP)

R1P

8Var(LBP) R1P8

80 100 120 140 160

06

08

1

12

Skew(LBP) R1P8

80 100 120 140 160

028

03

032

034

036

038

Ent(LBP) R4P16

80 100 120 140 160

1

2

3

4

5

Mean(C) R4P16

06 08 1 12

80

100

120

140

160

Skew

(LBP

) R1P

8

06 08 1 12

06

08

1

12

06 08 1 12

028

03

032

034

036

038

06 08 1 12

1

2

3

4

5

028 032 036

80

100

120

140

160

Ent(L

BP) R

4P16

028 032 036

06

08

1

12

028 032 036

028

03

032

034

036

038

028 032 036

1

2

3

4

5

1 2 3 4 5

80

100

120

140

160

Mea

n(C

) R4P

16

1 2 3 4 5

06

08

1

12

1 2 3 4 5

028

03

032

034

036

038

1 2 3 4 5

1

2

3

4

5

times104

times104

times104

times104

times104

times104

times104

times104

Figure 6 Matrix of scatter plots displaying the five selected features pairwise against each other Blue depicts normal controls and red depictsdementia

Table 3 Results are reported as mean with standard deviation inbrackets 119898(119904) over 10-fold cross validation classifying NC versusAD + LBD 119879 = total accuracy 119877 = recall and 119875 = precision 0 forclass NC and 1 for class AD + LBD ROI is either WM for whitematter or WML for white matter lesion area

Test 1198791198750 1198751

1198770 1198771

FLAIR-WML119903119894

080 (012) 069 (020) 087 (011)072 (023) 084 (012)

T1WMLri 098 (004) 098 (006) 099 (004)098 (008) 099 (005)

T1WM119903119894

097 (004) 096 (008) 099 (004)098 (008) 097 (006)

T1WML1199031198941199062

098 (004) 096 (008) 100 (000)100 (000) 097 (006)

T1WMLsvg119903119894 100 (000) 100 (000) 100 (000)100 (000) 100 (000)

(080 (012)) but approximately similar to the two differentROIs when T1 MR images are used Precision for class 0 ishigher in the case of LBP and 119862 calculated in the WML areaof the T1 image (098 (006)) as compared to all of the WMarea (096 (008)) Recall for class 0 is similar for both ROIsThis is also the case for precision for class 1 (099 (004)) butrecall for class 1 is higher when LBP and 119862 are calculated in

the WML region 099 (005) as compared to the WM region(097 (006))

When the rotational invariant calculation of LBP iscombined with selection of the uniform values only the1198750 and 1198771 are similar to the 119903119894-case The 1199031198941199062-case hadmarginally higher values for total accuracy 1198751 and 1198770


63 Two-Class Problem AD versus LBD Results for the two-class problemwith class 1 being AD and class 2 being LBD areshown in detail in Table 4

Classification performance was highest in the T1 casewhen WM was used as ROI

64 Stavanger Data Only In both the three-class problemand the two-class problem NC versus AD + LBD a fifth testwas run namedT1WMLsvg119903119894 which indicates that the T1MRimages were used for calculation of the LBP and 119862 that theWM was the ROI and that only data from the MR scannerlocated at Stavanger University Hospital were used Thisexperiment was done to assess the robustness of the methodThe rotational invariant variant of the LBP feature was usedin this test An even better performance was reached in bothcases In the three-class problem a total accuracy of 091 (015)was achieved and all of the cases in the data set were classifiedcorrectly in the two-class problem An implication of thisis that between-centre noise falsely reduces classification


Table 4 Results are reported as mean with standard deviation inbrackets 119898(119904) over 10-fold cross validation classifying AD versusLBD 119879 = total accuracy 119877 = recall and 119875 = precision 1 for class ADand 2 for class LBD ROI is eitherWM for white matter orWML forwhite matter lesion area

Test 1198791198751 1198752

1198771 1198772

FLAIR-WML119903119894

073 (015) 078 (011) 020 (045)091 (012) 010 (032)

T1WML119903119894

066 (017) 074 (010) 000 (000)084 (018) 000 (000)

T1WMLSMOTE119903119894

073 (016) 072 (018) 076 (017)075 (020) 071 (019)

T1WM119903119894

074 (016) 080 (009) 045 (051)075 (020) 071 (019)

T1WMSMOTE119903119894

068 (014) 067 (014) 075 (021)069 (029) 068 (014)

accuracy and that the developed method shows even higherperformance when all data come from the same scanner

7 Discussion

Our results improved doing LBP texture analysis in 3DT1image rather than the FLAIR image indicating that thereexistsmore textural information in the 3DT1 image comparedto the FLAIR image relevant to our problem formulation Inthe three-class problem as well as in the two-class problemNC versus AD + LBD our results indicate that there existssimilar amount of relevant textural information regardingdementia classification using all of WM as ROI compared tousing onlyWMLThis could be a benefitWML segmentationis unsatisfactorily developed and very often demandingman-ual outlining is required as well as a FLAIRMR image whereWML is hyperintense while WM segmentation is readilyavailable from many well known and freely downloadablesoftware packages needing only a 3DT1 MR image which isa common part of a clinical MR protocol In addition recentfocus on diffusion tensor imaging (DTI) in vascular dis-ease [58] amnestic mild cognitive impairment (aMCI) [59]and dementia [60ndash62] strengthens the view that age-relatedchanges inWMplay an important role in the development ofdementia DTI is nevertheless not sufficiently available andat the same time is costly making other approaches for WManalysis like ours a valuable addition

In the two-class problem AD versus LBD we did notreach a comparable classification result compared to theAD + LBD versus NC case There probably exist severalexplanations for that one of themost obvious being the smallsample size in the LBD class compared to the other classesThe LBD subjects are mainly classified as AD subjects indi-cating that the two groups experience similarities concerningour methods Even though the two groups show differentneurological etiologies they do not differ equally regardingvascular changes Having few subjects in the LBD group thecalculated texture features may not represent the group with

proper specificity or generality Another explanation could berelated to the commonbasis for neurodegenerative dementiaspointed out by Bartzokis in [21] or Schneiderrsquos observationsabout mixed brain pathologies in dementia [63]

In the three-class problemNCversusADversus LBD theaccuracy for the LBD class is improved showing a precisionof 085 (011) and recall of 078 (020) When doing the sametest on the data from Stavanger only even better results wereachieved with a precision of 087 (022) and a recall of 100(000) for the LBD class Vemuri et al [18] used atrophymapsand a 119896-means clustering approach to diagnose AD with asensitivity of 907 and a specificity of 84 LBD with asensitivity of 786 and specificity of 988 and FTLD witha sensitivity of 844 and a specificity of 938 A strength oftheir study was that they only used MR images of later histo-logically confirmed LBD patientsThey also report sensitivityand specificity for the respective clinical diagnoses AD witha sensitivity of 895 and a specificity of 821 LBD with asensitivity of 700 and specificity of 1000 and FTLD witha sensitivity of 830 and a specificity of 956 Compared tothe reported sensitivity and specificity for clinical diagnosisour method shows substantial higher accuracy for LBD andcomparable accuracy for AD A limitation is the use of differ-ent measures of goodness to the classification results and thatdifferent data is used In Kodama and Kawase [40] a clas-sification accuracy of 70 for the LBD group from AD andNC is reported Burton et al report a sensitivity of 91 and aspecificity of 94 using calculations of medial temporal lobeatrophy assessing diagnostic specificity of AD in a sample ofpatients with AD LBD and vascular cognitive impairmentbut do not report results for the LBD group [17] In [19] Lebe-dev et al use sparse partial least squares (SPLS) classificationof cortical thickness measurements reporting a sensitivity of944 and a specificity of 8889 discerning AD from LBD

To verify that the classification results are not driven bydifferences in the local variation of signal intensities (the119862 values) between centres used during collection of MRdata in the study the test T1WMLsvg119903119894 was conducted onthe Stavanger data only The results showed an increase inclassification performance which gives us reason to believethat the results reflect real diagnostic differences

LBP is based on local gradients and is therefore prone tonoise and could be a limitation to our approach LBP valuescalculated in a noisy neighbourhood would be recognised bymany transitions between 0s and 1s We performed a test theT1WML

1199031198941199062test where only rotational invariant and uniform

LBP values showing a maximum of two transitions between0s and 1s are collected The result showed identical results asthe T1WML

119903119894test indicating that noise does not constitute a

severe problem in our method Even though noise reductionprocedures can be useful in the application of for examplesegmentation a noise reduction approach could removerelevant textures The contrast measure is invariant to shiftsin gray-scale but not invariant to scaling We do not use anynormalization of the images prior to the feature calculationThus one could argue that different patients are scaleddifferently making the contrast measure less trustworthy Onthe other hand if a normalization is done for example basedon a maximum intensity value this could indeed change the


local subtle textures and affect the contrastmeasures possiblyin a negative way In the present work we have investigatedthe discriminating power of the features calculated withoutany smoothing or normalization since the effect of suchoperators is not clear for this application In future workwe want to investigate the use of different preprocessingsteps using both denoising and normalization and comparethe discrimination power of the features with and withoutpreprocessing The improvement in results when using datafrom one centre only (Stavanger) indicates lack of robustnesswhich can be related to the facts mentioned above

As mentioned in Section 22 Cronbachrsquos alpha was calcu-lated using total brain volume to ensure that our datamaterialwas consistent even though it was collected from differentcentres spanning a time scale Texture features can be exposedto noise and a limitation to our study is the lack of usingtexture features for the reliability analysis

Another limitation to our study is the lack of clinicalinterpretation of texture features which is difficult in ourcase since brain regional information is lost in the processof feature calculation

71 Conclusion This study demonstrates that LBP texturefeatures combined with the contrast measure 119862 calculatedfrom brain MR images are potent features used in a machinelearning context for computer based dementia diagnosisThe results discerning AD + LBD from NC are especiallypromising potentially adding value to the clinical diagnoseIn the three-class problem the classification performanceexceeded the accuracy of clinical diagnosis for the LBDgroupat the same time keeping the classification accuracy for theAD group comparable to the clinical diagnoses A loweraccuracy was achieved when classifying AD from LBD in thetwo-class problem AD versus LBD We considered it goodnews that the results using WM as ROI gave almost equallygood classification performance as WML since the WMsegmentation routine is much more accessible compared toWML segmentationThe performance using 3DT1 images fortexture analysis was notably better than when using FLAIRimages which is an advantage since most common MRprotocols include a 3DT1 image

For future work we will look into texture features cal-culated in a 3D neighbourhood 3D texture features haveshown to be an important step towards better discriminationin machine learning systems when the images are intrinsicthree-dimensional likemanyMR sequences are [64] In addi-tion we will perform correlation analysis between texturefeatures and cognition since that could improve the clinicalvalue of our work

Conflict of Interests

Author ldquoD Aarslandrdquo received honoraria from NovartisLundbeck GSK Merck Serono DiaGenic GE Health andOrionPharmaceuticals andhas provided research support forMerck Serono Lundbeck and Novartis No other author hasanything to disclose

Acknowledgments

The authors want to thank principal investigators of theParkWest [46] cohort JP Larsen and OB Tysnes for givingaccess to MR images of the normal controls in the studyWe also want to thankTheWestern Norway Regional HealthAuthority for providing funding by grant 911546

References

[1] A Wimo and M Prince World Alzheimer Report 2010 TheGlobal Economic Impact of Dementia Alzheimerrsquos DiseaseInternational 2010 httpwwwalzcoukresearchfilesWorld-AlzheimerReport2010pdf

[2] D Aarsland A Rongve S P Nore et al ldquoFrequency and caseidentification of dementia with Lewy bodies using the revisedconsensus criteriardquoDementia andGeriatric Cognitive Disordersvol 26 no 5 pp 445ndash452 2008

[3] D P Perl ldquoNeuropathology of Alzheimerrsquos diseaserdquoTheMountSinai Journal of Medicine vol 77 no 1 pp 32ndash42 2010

[4] D Aarsland C G Ballard and G Halliday ldquoAre Parkinsonrsquosdisease with dementia and dementia with Lewy bodies the sameentityrdquo Journal of Geriatric Psychiatry andNeurology vol 17 no3 pp 137ndash145 2004

[5] C F Lippa J E Duda M Grossman et al ldquoDLB and PDDboundary issues diagnosis treatment molecular pathologyand biomarkersrdquo Neurology vol 68 no 11 pp 812ndash819 2007

[6] I G McKeith D W Dickson J Lowe et al ldquoDiagnosis andmanagement of dementia with Lewy bodies third report ofthe DLB consortiumrdquo Neurology vol 65 no 12 pp 1863ndash18722005

[7] B Thanvi N Lo and T Robinson ldquoVascular parkinsonismmdashan important cause of parkinsonism in older peoplerdquo Age andAgeing vol 34 no 2 pp 114ndash119 2005

[8] M Filippi and F Agosta ldquoStructural and functional networkconnectivity breakdown in Alzheimerrsquos disease studied withmagnetic resonance imaging techniquesrdquo Journal of AlzheimerrsquosDisease vol 24 no 3 pp 455ndash474 2011

[9] J A Bertelson and B Ajtai ldquoNeuroimaging of dementiardquo Neu-rologic Clinics vol 32 no 1 pp 59ndash93 2014

[10] D Wang S C Hui L Shi et al ldquoApplication of multimodalMR imaging on studying Alzheimerrsquos disease a surveyrdquoCurrentAlzheimer Research vol 10 no 8 pp 877ndash892 2013

[11] P Malloy S Correia G Stebbins and D H Laidlaw ldquoNeu-roimaging of white matter in aging and dementiardquoThe ClinicalNeuropsychologist vol 21 no 1 pp 73ndash109 2007

[12] J Stoitsis I Valavanis S G Mougiakakou S Golemati ANikita and K S Nikita ldquoComputer aided diagnosis based onmedical image processing and artificial intelligence methodsrdquoNuclear Instruments and Methods in Physics Research SectionA Accelerators Spectrometers Detectors and Associated Equip-ment vol 569 no 2 pp 591ndash595 2006

[13] K R Gray P Aljabar R A Heckemann A Hammers and DRueckert ldquoRandom forest-based similarity measures for multi-modal classification of Alzheimerrsquos diseaserdquo NeuroImage vol65 pp 167ndash175 2013

[14] M Liu D Zhang and D Shen ldquoEnsemble sparse classificationofAlzheimerrsquos diseaserdquoNeuroImage vol 60 no 2 pp 1106ndash11162012

[15] R Cuingnet E Gerardin J Tessieras et al ldquoAutomatic clas-sification of patients with Alzheimerrsquos disease from structural


MRI a comparison of ten methods using the ADNI databaserdquoNeuroImage vol 56 no 2 pp 766ndash781 2011

[16] S Kloppel C M Stonnington J Barnes et al ldquoAccuracy ofdementia diagnosismdasha direct comparison between radiologistsand a computerized methodrdquo Brain vol 131 no 11 pp 2969ndash2974 2008

[17] E J Burton R Barber E BMukaetova-Ladinska et al ldquoMedialtemporal lobe atrophy on MRI differentiates Alzheimerrsquos dis-ease from dementia with Lewy bodies and vascular cognitiveimpairment a prospective study with pathological verificationof diagnosisrdquo Brain vol 132 no 1 pp 195ndash203 2009

[18] P Vemuri G Simon K Kantarci et al ldquoAntemortem differ-ential diagnosis of dementia pathology using structural MRIdifferential-STANDrdquo NeuroImage vol 55 no 2 pp 522ndash5312011

[19] A V Lebedev E Westman M K Beyer et al ldquoMultivariateclassification of patients with Alzheimerrsquos and dementia withLewy bodies using high-dimensional cortical thickness mea-surements anMRI surface-basedmorphometric studyrdquo Journalof Neurology vol 260 no 4 pp 1104ndash1115 2013

[20] C M Filley ldquoWhite matter dementiardquoTherapeutic Advances inNeurological Disorders vol 5 no 5 pp 267ndash277 2012

[21] G Bartzokis ldquoAlzheimerrsquos disease as homeostatic responses toage-related myelin breakdownrdquo Neurobiology of Aging vol 32no 8 pp 1341ndash1371 2011

[22] F M Gunning-Dixon A M Brickman J C Cheng and G SAlexopoulos ldquoAging of cerebral white matter a review of MRIfindingsrdquo International Journal of Geriatric Psychiatry vol 24no 2 pp 109ndash117 2009

[23] A Poggesi L Pantoni D Inzitari et al ldquo2001ndash2011 a decade ofThe Ladis (Leukoaraiosis and Disability) study what have welearned about white matter changes and small-vessel diseaserdquoCerebrovascular Diseases vol 32 no 6 pp 577ndash588 2011

[24] V G Young G M Halliday and J J Kril ldquoNeuropathologiccorrelates of white matter hyperintensitiesrdquo Neurology vol 71no 11 pp 804ndash811 2008

[25] F-E de Leeuw J C de Groot E Achten et al ldquoPrevalence ofcerebral white matter lesions in elderly people a populationbased magnetic resonance imaging study The Rotterdam ScanStudyrdquo Journal of Neurology Neurosurgery amp Psychiatry vol 70no 1 pp 9ndash14 2001

[26] MYoshita E FletcherDHarvey et al ldquoExtent and distributionof white matter hyperintensities in normal aging MCI andADrdquo Neurology vol 67 no 12 pp 2192ndash2198 2006

[27] R Barber P Scheltens A Gholkar et al ldquoWhite matter lesionson magnetic resonance imaging in dementia with Lewy bodiesAlzheimerrsquos disease vascular dementia and normal agingrdquoJournal of Neurology Neurosurgery and Psychiatry vol 67 no1 pp 66ndash72 1999

[28] N D Prins E J van Dijk T den Heijer et al ldquoCerebral whitematter lesions and the risk of dementiardquo Archives of Neurologyvol 61 no 10 pp 1531ndash1534 2004

[29] H Baezner C Blahak A Poggesi et al ldquoAssociation of gait andbalance disorders with age-related white matter changes theLADIS Studyrdquo Neurology vol 70 no 12 pp 935ndash942 2008

[30] H Soennesyn K Oppedal O J Greve et al ldquoWhite matterhyperintensities and the course of depressive symptoms inelderly people with mild dementiardquo Dementia and GeriatricCognitive Disorders Extra vol 2 no 1 pp 97ndash111 2012

[31] N D Prins E J van Dijk T den Heijer et al ldquoCerebral small-vessel disease and decline in information processing speed

executive function andmemoryrdquoBrain vol 128 no 9 pp 2034ndash2041 2005

[32] A M Tuladhar A T Reid E Shumskaya et al ldquoRelationshipbetween white matter hyperintensities cortical thickness andcognitionrdquo Stroke vol 46 no 2 pp 425ndash432 2015

[33] S MunozManiega M C Valdes Hernandez and J D ClaydenldquoWhite matter hyperintensities and normal-appearing whitematter integrity in the aging brainrdquo Neurobiology of Aging vol36 no 2 pp 909ndash918 2015

[34] M Fujishima N Maikusa K Nakamura M Nakatsuka HMatsuda and K Meguro ldquoMild cognitive impairment poorepisodic memory and late-life depression are associated withcerebral cortical thinning and increased white matter hyperin-tensitiesrdquo Frontiers in Aging Neuroscience vol 6 2014

[35] L Harrison Clinical applicability of mri texture analysis[PhD thesis] University of Tampere Tampere Finland 2011httptampubutafibitstreamhandle1002466779978-951-44-8527-5pdfsequence=1

[36] P A Freeborough and N C Fox ldquoMR image texture analysisapplied to the diagnosis and tracking of alzheimerrsquos diseaserdquoIEEE Transactions on Medical Imaging vol 17 no 3 pp 475ndash479 1998

[37] M S de Oliveira M L F Balthazar A DrsquoAbreu et al ldquoMRimaging texture analysis of the corpus callosum and thalamusin amnestic mild cognitive impairment and mild Alzheimerdiseaserdquo The American Journal of Neuroradiology vol 32 no1 pp 60ndash66 2011

[38] J Zhang C Yu G Jiang W Liu and L Tong ldquo3D textureanalysis on MRI images of Alzheimerrsquos diseaserdquo Brain Imagingand Behavior vol 6 no 1 pp 61ndash69 2012

[39] T R Sivapriya V Saravanan and P R J Thangaiah ldquoTextureanalysis of brain MRI and classification with BPN for the diag-nosis of dementiardquo Communications in Computer and Informa-tion Science vol 204 pp 553ndash563 2011

[40] N Kodama and Y Kawase ldquoComputerized method for classi-fication between dementia with Lewy bodies and Alzheimerrsquosdisease by use of texture analysis on brain MRIrdquo in Proceedingsof the World Congress on Medical Physics and BiomedicalEngineering pp 319ndash321 Munich Germany September 2009

[41] T Ojala M Pietikainen and D Harwood ldquoPerformance evalu-ation of texture measures with classification based on kullbackdiscrimination of distributionsrdquo in Proceedings of the 12thInternational Conference on Pattern Recognition vol 1 pp 582ndash585 Jerusalem Israel 1994

[42] T Ojala M Pietikainen and D Harwood ldquoA comparativestudy of texture measures with classification based on featuredistributionsrdquo Pattern Recognition vol 29 no 1 pp 51ndash59 1996

[43] D Unay A Ekin M Cetin R Jasinschi and A Ercil ldquoRobust-ness of local binary patterns in brain MR image analysisrdquo inProceedings of the 29th Annual International Conference of theIEEE Engineering in Medicine and Biology Society (EMBS 07)pp 2098ndash2101 Lyon France August 2007

[44] K Oppedal D Aarsland M J Firbank et al ldquoWhite matterhyperintensities in mild Lewy body dementiardquo Dementia andGeriatric Cognitive Disorders Extra vol 2 no 1 pp 481ndash4952012

[45] K Oppedal K Engan D Aarsland M Beyer O B Tysnes andT Eftestoslashl ldquoUsing local binary pattern to classify dementia inMRIrdquo in Proceedings of the 9th IEEE International Symposiumon Biomedical Imaging FromNano toMacro (ISBI rsquo12) pp 594ndash597 May 2012


[46] GAlves BMuller KHerlofson et al ldquoIncidence of Parkinsonrsquosdisease in Norway the Norwegian ParkWest studyrdquo Journal ofNeurology Neurosurgery and Psychiatry vol 80 no 8 pp 851ndash857 2009

[47] J Ashburner and K J Friston ldquoUnified segmentationrdquo Neu-roImage vol 26 no 3 pp 839ndash851 2005

[48] M J Firbank A J Lloyd N Ferrier and J T OrsquoBrien ldquoA volu-metric study of MRI signal hyperintensities in late-life depres-sionrdquo The American Journal of Geriatric Psychiatry vol 12 no6 pp 606ndash612 2004

[49] SPM5 Statistical Parametric Mapping 2005 httpwwwfilionuclacukspm

[50] MNI Montreal Neurological Institute 1995 httpwwwnilwustledulabskevinmananswersmnispacehtml

[51] M J Firbank TMinett and J T OrsquoBrien ldquoChanges inDWI andMRS associated with white matter hyperintensities in elderlysubjectsrdquo Neurology vol 61 no 7 pp 950ndash954 2003

[52] FSLView v31 ldquoFMRIB Software Libary v40rdquo August 2007httpfslfmriboxacukfslfslview

[53] L Breiman ldquoRandom forestsrdquo Tech Rep Statistics Depart-ment University of California Berkely Calif USA 2001 httpozberkeleyeduusersbreimanrandomforest2001pdf

[54] T Hastie R Tibshirani and J FriedmanThe Elements of Statis-tical Learning vol 2 Springer 2009

[55] Matlab R2012b ldquoMathworksrdquo October 2012 httpwwwmath-worksseindexhtml

[56] N V Chawla K W Bowyer L O Hall and W P KegelmeyerldquoSMOTE synthetic minority over-sampling techniquerdquo Journalof Artificial Intelligence Research vol 16 pp 321ndash357 2002

[57] H He and E A Garcia ldquoLearning from imbalanced datardquo IEEETransactions on Knowledge and Data Engineering vol 21 no 9pp 1263ndash1284 2009

[58] G S Alves F K Sudo C E D O Alves et al ldquoDiffusion tensorimaging studies in vascular disease a review of the literaturerdquoDementia e Neuropsychologia vol 6 no 3 pp 158ndash163 2012

[59] M Dyrba M Ewers M Wegrzyn et al ldquoPredicting prodromalAlzheimerrsquos disease in people with mild cognitive impairmentusing multicenter diffusion-tensor imaging data and machinelearning algorithmsrdquo Alzheimerrsquos amp Dementia vol 9 no 4 p426 2013

[60] M Naik A Lundervold H Nygaard and J-T Geitung ldquoDif-fusion tensor imaging (DTI) in dementia patients with frontallobe symptomsrdquo Acta Radiologica vol 51 no 6 pp 662ndash6682010

[61] M J Firbank AM Blamire A Teodorczuk E Teper DMitraand J T OrsquoBrien ldquoDiffusion tensor imaging in Alzheimerrsquosdisease and dementia with Lewy bodiesrdquo Psychiatry ResearchNeuroimaging vol 194 no 2 pp 176ndash183 2011

[62] R Watson A M Blamire S J Colloby et al ldquoCharacterizingdementia with Lewy bodies by means of diffusion tensorimagingrdquo Neurology vol 79 no 9 pp 906ndash914 2012

[63] J A Schneider Z Arvanitakis W Bang and D A BennettldquoMixed brain pathologies account for most dementia cases incommunity-dwelling older personsrdquo Neurology vol 69 no 24pp 2197ndash2204 2007

[64] V A Kovalev F Kruggel H J Gertz and D Y von CramonldquoThree-dimensional texture analysis of MRI brain datasetsrdquoIEEE Transactions on Medical Imaging vol 20 no 5 pp 424ndash433 2001

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Active and Passive Electronic Components

Control Scienceand Engineering

Journal of



RotatingMachinery


Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics


Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation



DistributedSensor Networks



AD and LBD is also important since they differ in prog-nosis and response to drug treatment Currently the onlyavailable method to differentiate between AD and LBD isthe dopamine transporter scan which is expensive and notreadily available at all centres

Neuroimaging in Dementia Neuroimaging is an importanttool for studying dementia and cognitive deterioration Sev-eral excellent reviews are available [8ndash10] In [11] Malloyet al review available methods for quantitative imaging ofwhitematter anatomy andpathology aswell as recent findingsin ageing and dementia They state that computer aidedquantification offers better statistical power compared tovisual rating scales and that diffusion imaging is able to detectabnormalities not recognised in conventional acquisitionsequences Early detection of disease and relevant functionalconnections between brain areas are important benefits

Computer Aided Diagnosis in Dementia Computer aideddiagnosis (CAD) can be a helpful tool to pinpoint diagnosisearly in the disease course in a cost-effective manner andunbiased to human inconsistencies [12] Recent advances inthe field have focused especially on AD and patients withmild cognitive impairment (MCI) which are considered aprecursor to AD [13ndash16] Less attention has been put intodeveloping CAD systems for LBD As mentioned above LBDhave high prevalence and accurate clinical diagnosis dependson little available and expensive dopamine transporter scanand postmortem histology Few papers report high accuracydiscerning patients with AD and LBD or other dementias [1718] A promising approach is reported in [19] where Lebedevet al use sparse partial least squares (SPLS) classification ofcortical thickness measurements reporting a sensitivity of944 and a specificity of 8889 discerning AD from LBD

White Matter and White Matter Lesions in Dementia Whitematter (WM) comprises approximately half the brain volumeand provides connectivity between the two brain hemi-spheres as well as ensuring efficient transfer of neural activitycomplementing information processing in the gray matter(GM) WM neuropathology is often diffuse and affects manyneuronal networks which can be disturbed simultaneouslyresulting in a multidomain syndrome In [20] Filley empha-sizes the contribution of white matter disease (WMD) inmild cognitive dysfunction cognitive ageing and dementiaBartzokis [21] proposes a hypothesis for AD called the ldquomye-lin modelrdquo where axonal transport disruption formationof axonal swellings neuritic plaques and proteinaceousdeposits such as A120573 and tau are by-products of homeostaticmyelin repair processes Gunning-Dixon et al [22] reviewresults ofMRI studies of whitematter changes that occurwithnormal ageing and the relationship of age-associated changesin white matter to age-related declines in cognitive abilities

White matter lesions or white matter hyperintensities(WML) are among the neuroimaging expressions of cerebralsmall-vessel disease and are associated with various distur-bances with poor prognosis [23] WML are localized areasof increased signal intensities in the white matter of thebrain visible on T2-weighted MR images The underlyingpathology of WML is heterogeneous ranging from mild

demyelination to incomplete subcortical infarctionsThey aretypically seen around the ventricles (periventricular WML)but also as focal lesions in the deep white matter In theelderly WML usually represent small-vessel cerebrovasculardisease (CVD) [24] WML becomes more abundant withincreasing age in healthy subjects [25] but they are alsofound to be associated with AD [26] and other dementias[27 28] Clinical symptoms associated with WML includegait disturbances [29] depression [27 30] and cognitiveimpairment [31] although the exactmechanisms are not fullyunderstood In [32] Tuladhar et al concludes that corticalchanges mediated by WML and vascular risk factors mightlead to cognitive decline and dementia MunozManiega et al[33] write that age-related deterioration of normal appearingwhitematter (NAWM) is strongly associatedwith the severityof WML indicating that WML is important in dementiaresearch Fujishima et al conclude in [34] that mild cognitiveimpairment poor episodic memory and late-life depressionare associated with cerebral cortical thinning and WML

Texture Analysis in Neuroimaging Harrison [35] extensivelyreviews the use of texture analysis in a clinical contextanalysing MR images in non-Hodgkin lymphoma mildtraumatic brain injury and multiple sclerosis She concludesthat ldquonon visible lesions and physiological changes as well asvisible focal lesions of different aetiologies could be detectedand characterized by texture analysis of routine clinical 15Tesla scansrdquo The application of texture analysis in a machinelearning (ML) environment has shown success in discern-ing different dementias from each other and from healthycontrols Freeborough and Fox reported a classification rateof 91 discerning AD from healthy controls using measuresfrom a spatial gray-level dependent method applied in astepwise discriminant analysis approach in [36] de Olivieiraet al [37] found statistical significant differences in graylevel cooccurrence matrix measurements in subjects withmild AD amnestic mild cognitive impairment (aMCI) andhealthy controls Zhang et al [38] performed 3D textureanalysis in MR images of the hippocampus and entorhinalcortex in AD patients and achieved a classification accuracybetween 643 and 964 Sivapriya et al showed in [39]that texture analysis in brain MRI gave high classificationaccuracy in AD As of the authors knowledge the only paperconsidering texture analysis as an approach to distinguishAD LBD and NC is [40] where Kodama and Kawaseperformed discriminant analysis on features extracted froma cooccurrence matrix and a run-length matrix with anaccuracy of 917 700 and 880 respectively

Local binary pattern (LBP) was introduced by Ojala et al[41 42] as a texture descriptor It is a simple yet very efficienttexture operator which labels the pixels of an image bythresholding the neighbourhood of each pixel and considersthe result as a binary number Unay et al [43] showed thatthe rotation invariant LBP is invariant to some commonMRIartefacts which makes it a robust texture feature when usedin brain MR image analysis

Aims We have earlier shown that there were no differencesin WML volume between patients with AD and LBD or






2 Material

















119879 = 119905 (119892119888 1198920 119892

119875minus1) (1)




119888are (0 0) the






giving119879 = 119905 (119892

119888 1198920minus 119892119888 1198921minus 119892119888 119892

119875minus1minus 119892119888) (2)


119888


119888) 119905 (1198920minus 119892119888 1198921minus 119892119888 119892

119875minus1minus 119892119888) (3)




119875minus1minus 119892119888) (4)



119879 asymp 119905 (119904 (1198920minus 119892119888) 119904 (119892

1minus 119892119888) 119904 (119892

119875minus1minus 119892119888)) (5)

where

119904 (119909) = 1 119909 ge 0

0 119909 lt 0(6)




LBP119875119877

=119875minus1

sum119901=0

119904 (119892119901minus 119892119888) 2119875 (7)



is defined as

LBP119903119894119875119877

= min ROR (LBP119875119877 119894) | 119894 = 0 1 119875 minus 1 (8)



2and 11111111



2of 1 and


LBP1199031198941199062119875119877

=

119875minus1

sum119901=0

119904 (119892119901minus 119892119888) if 119880(LBP

119875119877) le 2

119875 + 1 otherwise(9)

where119880 (LBP

119875119877) =


+119875minus1

sum119901=1


(10)


42 Contrast The LBP119903119894119875119877

and LBP1199031198941199062119875119877




LBPPR =Pminus1

sump=0

s(gp minus gc)2p

Sample Threshold


1+ 0 lowast 2

2+ 0 lowast 2

3+ 0 lowast 2

4+ 1 lowast 2

5+ 0 lowast 2

6+ 1 lowast 2

7= 163


s(x) = 1 if x ⩾ 0

0 otherwise

83

52

94

88

97

61

14

77 82

0

1

1

1

0

0

0

1

x

(a)

52

94

88

97

61

14

77

83

82

5279

94

97

88

73

67

5461

54

14

4

23

23

83

23

82

(b)



VAR119875119877

=1

119875

119875minus1

sum119901=0

(119892119901minus 120583)2

where 120583 = 1

119875

119875minus1

sum119901=0

119892119901 (11)









(1)ROI extraction


and C







Table 1










1 1198792 119879

10) (1198751 1198752 119875

10) and

(1198771 1198772 119877



119898119909=1

119899

119899

sum119896=1

119909119896 where 0 le 119909

119896le 1 0 le 119898

119909le 1 (12)


(a) (b)

(c) (d)



119904119909= (

1

119899 minus 1

119899

sum119896=1

(119909119896minus 119898119909119896)2

)

12

where 0 le 119909119896le 1 0 le 119904

119909le 1

(13)







(a) (b)

(c) (d)



6 Results





119903119894








Training data

Test data

Test dataInner loop


feature sets



andpredictoutertest

results



measures



Test 1198791198750 1198751 1198752

1198770 1198771 1198772

FLAIR-WML119903119894

060 (013) 071 (028) 061 (014) 033 (041)048 (025) 077 (028) 020 (035)

T1WML119903119894

082 (012) 096 (010) 080 (011) 058 (049)098 (008) 088 (018) 025 (035)

T1WMLSMOTEri 087 (008) 097 (007) 081 (017) 085 (011)

100 (000) 082 (016) 078 (020)

T1WM119903119894

082 (009) 096 (008) 081 (011) 042 (049)100 (000) 088 (016) 020 (035)

T1WMSMOTE119903119894

075 (013) 090 (012) 066 (016) 070 (021)100 (000) 072 (019) 055 (022)

T1WMLSMOTEsvg119903119894 091 (015) 100 (000) 100 (000) 087 (022)

100 (000) 077 (042) 100 (000)







80 100 120 140 160

80

100

120

140

160

Var(

LBP)

R1P

8Var(LBP) R1P8

80 100 120 140 160

06

08

1

12

Skew(LBP) R1P8

80 100 120 140 160

028

03

032

034

036

038

Ent(LBP) R4P16

80 100 120 140 160

1

2

3

4

5

Mean(C) R4P16

06 08 1 12

80

100

120

140

160

Skew

(LBP

) R1P

8

06 08 1 12

06

08

1

12

06 08 1 12

028

03

032

034

036

038

06 08 1 12

1

2

3

4

5

028 032 036

80

100

120

140

160

Ent(L

BP) R

4P16

028 032 036

06

08

1

12

028 032 036

028

03

032

034

036

038

028 032 036

1

2

3

4

5

1 2 3 4 5

80

100

120

140

160

Mea

n(C

) R4P

16

1 2 3 4 5

06

08

1

12

1 2 3 4 5

028

03

032

034

036

038

1 2 3 4 5

1

2

3

4

5

times104

times104

times104

times104

times104

times104

times104

times104



Test 1198791198750 1198751

1198770 1198771

FLAIR-WML119903119894

080 (012) 069 (020) 087 (011)072 (023) 084 (012)

T1WMLri 098 (004) 098 (006) 099 (004)098 (008) 099 (005)

T1WM119903119894

097 (004) 096 (008) 099 (004)098 (008) 097 (006)

T1WML1199031198941199062

098 (004) 096 (008) 100 (000)100 (000) 097 (006)

T1WMLsvg119903119894 100 (000) 100 (000) 100 (000)100 (000) 100 (000)










Test 1198791198751 1198752

1198771 1198772

FLAIR-WML119903119894

073 (015) 078 (011) 020 (045)091 (012) 010 (032)

T1WML119903119894

066 (017) 074 (010) 000 (000)084 (018) 000 (000)


073 (016) 072 (018) 076 (017)075 (020) 071 (019)

T1WM119903119894

074 (016) 080 (009) 045 (051)075 (020) 071 (019)

T1WMSMOTE119903119894

068 (014) 067 (014) 075 (021)069 (029) 068 (014)


7 Discussion



















Acknowledgments


References







































































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation














2 Material

















119879 = 119905 (119892119888 1198920 119892

119875minus1) (1)




119888are (0 0) the






giving119879 = 119905 (119892

119888 1198920minus 119892119888 1198921minus 119892119888 119892

119875minus1minus 119892119888) (2)


119888


119888) 119905 (1198920minus 119892119888 1198921minus 119892119888 119892

119875minus1minus 119892119888) (3)




119875minus1minus 119892119888) (4)



119879 asymp 119905 (119904 (1198920minus 119892119888) 119904 (119892

1minus 119892119888) 119904 (119892

119875minus1minus 119892119888)) (5)

where

119904 (119909) = 1 119909 ge 0

0 119909 lt 0(6)




LBP119875119877

=119875minus1

sum119901=0

119904 (119892119901minus 119892119888) 2119875 (7)



is defined as

LBP119903119894119875119877

= min ROR (LBP119875119877 119894) | 119894 = 0 1 119875 minus 1 (8)



2and 11111111



2of 1 and


LBP1199031198941199062119875119877

=

119875minus1

sum119901=0

119904 (119892119901minus 119892119888) if 119880(LBP

119875119877) le 2

119875 + 1 otherwise(9)

where119880 (LBP

119875119877) =


+119875minus1

sum119901=1


(10)


42 Contrast The LBP119903119894119875119877

and LBP1199031198941199062119875119877




LBPPR =Pminus1

sump=0

s(gp minus gc)2p

Sample Threshold


1+ 0 lowast 2

2+ 0 lowast 2

3+ 0 lowast 2

4+ 1 lowast 2

5+ 0 lowast 2

6+ 1 lowast 2

7= 163


s(x) = 1 if x ⩾ 0

0 otherwise

83

52

94

88

97

61

14

77 82

0

1

1

1

0

0

0

1

x

(a)

52

94

88

97

61

14

77

83

82

5279

94

97

88

73

67

5461

54

14

4

23

23

83

23

82

(b)



VAR119875119877

=1

119875

119875minus1

sum119901=0

(119892119901minus 120583)2

where 120583 = 1

119875

119875minus1

sum119901=0

119892119901 (11)









(1)ROI extraction


and C







Table 1










1 1198792 119879

10) (1198751 1198752 119875

10) and

(1198771 1198772 119877



119898119909=1

119899

119899

sum119896=1

119909119896 where 0 le 119909

119896le 1 0 le 119898

119909le 1 (12)


(a) (b)

(c) (d)



119904119909= (

1

119899 minus 1

119899

sum119896=1

(119909119896minus 119898119909119896)2

)

12

where 0 le 119909119896le 1 0 le 119904

119909le 1

(13)







(a) (b)

(c) (d)



6 Results





119903119894








Training data

Test data

Test dataInner loop


feature sets



andpredictoutertest

results



measures



Test 1198791198750 1198751 1198752

1198770 1198771 1198772

FLAIR-WML119903119894

060 (013) 071 (028) 061 (014) 033 (041)048 (025) 077 (028) 020 (035)

T1WML119903119894

082 (012) 096 (010) 080 (011) 058 (049)098 (008) 088 (018) 025 (035)

T1WMLSMOTEri 087 (008) 097 (007) 081 (017) 085 (011)

100 (000) 082 (016) 078 (020)

T1WM119903119894

082 (009) 096 (008) 081 (011) 042 (049)100 (000) 088 (016) 020 (035)

T1WMSMOTE119903119894

075 (013) 090 (012) 066 (016) 070 (021)100 (000) 072 (019) 055 (022)

T1WMLSMOTEsvg119903119894 091 (015) 100 (000) 100 (000) 087 (022)

100 (000) 077 (042) 100 (000)







80 100 120 140 160

80

100

120

140

160

Var(

LBP)

R1P

8Var(LBP) R1P8

80 100 120 140 160

06

08

1

12

Skew(LBP) R1P8

80 100 120 140 160

028

03

032

034

036

038

Ent(LBP) R4P16

80 100 120 140 160

1

2

3

4

5

Mean(C) R4P16

06 08 1 12

80

100

120

140

160

Skew

(LBP

) R1P

8

06 08 1 12

06

08

1

12

06 08 1 12

028

03

032

034

036

038

06 08 1 12

1

2

3

4

5

028 032 036

80

100

120

140

160

Ent(L

BP) R

4P16

028 032 036

06

08

1

12

028 032 036

028

03

032

034

036

038

028 032 036

1

2

3

4

5

1 2 3 4 5

80

100

120

140

160

Mea

n(C

) R4P

16

1 2 3 4 5

06

08

1

12

1 2 3 4 5

028

03

032

034

036

038

1 2 3 4 5

1

2

3

4

5

times104

times104

times104

times104

times104

times104

times104

times104



Test 1198791198750 1198751

1198770 1198771

FLAIR-WML119903119894

080 (012) 069 (020) 087 (011)072 (023) 084 (012)

T1WMLri 098 (004) 098 (006) 099 (004)098 (008) 099 (005)

T1WM119903119894

097 (004) 096 (008) 099 (004)098 (008) 097 (006)

T1WML1199031198941199062

098 (004) 096 (008) 100 (000)100 (000) 097 (006)

T1WMLsvg119903119894 100 (000) 100 (000) 100 (000)100 (000) 100 (000)










Test 1198791198751 1198752

1198771 1198772

FLAIR-WML119903119894

073 (015) 078 (011) 020 (045)091 (012) 010 (032)

T1WML119903119894

066 (017) 074 (010) 000 (000)084 (018) 000 (000)


073 (016) 072 (018) 076 (017)075 (020) 071 (019)

T1WM119903119894

074 (016) 080 (009) 045 (051)075 (020) 071 (019)

T1WMSMOTE119903119894

068 (014) 067 (014) 075 (021)069 (029) 068 (014)


7 Discussion



















Acknowledgments


References







































































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation














119879 = 119905 (119892119888 1198920 119892

119875minus1) (1)




119888are (0 0) the






giving119879 = 119905 (119892

119888 1198920minus 119892119888 1198921minus 119892119888 119892

119875minus1minus 119892119888) (2)


119888


119888) 119905 (1198920minus 119892119888 1198921minus 119892119888 119892

119875minus1minus 119892119888) (3)




119875minus1minus 119892119888) (4)



119879 asymp 119905 (119904 (1198920minus 119892119888) 119904 (119892

1minus 119892119888) 119904 (119892

119875minus1minus 119892119888)) (5)

where

119904 (119909) = 1 119909 ge 0

0 119909 lt 0(6)




LBP119875119877

=119875minus1

sum119901=0

119904 (119892119901minus 119892119888) 2119875 (7)



is defined as

LBP119903119894119875119877

= min ROR (LBP119875119877 119894) | 119894 = 0 1 119875 minus 1 (8)



2and 11111111



2of 1 and


LBP1199031198941199062119875119877

=

119875minus1

sum119901=0

119904 (119892119901minus 119892119888) if 119880(LBP

119875119877) le 2

119875 + 1 otherwise(9)

where119880 (LBP

119875119877) =


+119875minus1

sum119901=1


(10)


42 Contrast The LBP119903119894119875119877

and LBP1199031198941199062119875119877




LBPPR =Pminus1

sump=0

s(gp minus gc)2p

Sample Threshold


1+ 0 lowast 2

2+ 0 lowast 2

3+ 0 lowast 2

4+ 1 lowast 2

5+ 0 lowast 2

6+ 1 lowast 2

7= 163


s(x) = 1 if x ⩾ 0

0 otherwise

83

52

94

88

97

61

14

77 82

0

1

1

1

0

0

0

1

x

(a)

52

94

88

97

61

14

77

83

82

5279

94

97

88

73

67

5461

54

14

4

23

23

83

23

82

(b)



VAR119875119877

=1

119875

119875minus1

sum119901=0

(119892119901minus 120583)2

where 120583 = 1

119875

119875minus1

sum119901=0

119892119901 (11)









(1)ROI extraction


and C







Table 1










1 1198792 119879

10) (1198751 1198752 119875

10) and

(1198771 1198772 119877



119898119909=1

119899

119899

sum119896=1

119909119896 where 0 le 119909

119896le 1 0 le 119898

119909le 1 (12)


(a) (b)

(c) (d)



119904119909= (

1

119899 minus 1

119899

sum119896=1

(119909119896minus 119898119909119896)2

)

12

where 0 le 119909119896le 1 0 le 119904

119909le 1

(13)







(a) (b)

(c) (d)



6 Results





119903119894








Training data

Test data

Test dataInner loop


feature sets



andpredictoutertest

results



measures



Test 1198791198750 1198751 1198752

1198770 1198771 1198772

FLAIR-WML119903119894

060 (013) 071 (028) 061 (014) 033 (041)048 (025) 077 (028) 020 (035)

T1WML119903119894

082 (012) 096 (010) 080 (011) 058 (049)098 (008) 088 (018) 025 (035)

T1WMLSMOTEri 087 (008) 097 (007) 081 (017) 085 (011)

100 (000) 082 (016) 078 (020)

T1WM119903119894

082 (009) 096 (008) 081 (011) 042 (049)100 (000) 088 (016) 020 (035)

T1WMSMOTE119903119894

075 (013) 090 (012) 066 (016) 070 (021)100 (000) 072 (019) 055 (022)

T1WMLSMOTEsvg119903119894 091 (015) 100 (000) 100 (000) 087 (022)

100 (000) 077 (042) 100 (000)







80 100 120 140 160

80

100

120

140

160

Var(

LBP)

R1P

8Var(LBP) R1P8

80 100 120 140 160

06

08

1

12

Skew(LBP) R1P8

80 100 120 140 160

028

03

032

034

036

038

Ent(LBP) R4P16

80 100 120 140 160

1

2

3

4

5

Mean(C) R4P16

06 08 1 12

80

100

120

140

160

Skew

(LBP

) R1P

8

06 08 1 12

06

08

1

12

06 08 1 12

028

03

032

034

036

038

06 08 1 12

1

2

3

4

5

028 032 036

80

100

120

140

160

Ent(L

BP) R

4P16

028 032 036

06

08

1

12

028 032 036

028

03

032

034

036

038

028 032 036

1

2

3

4

5

1 2 3 4 5

80

100

120

140

160

Mea

n(C

) R4P

16

1 2 3 4 5

06

08

1

12

1 2 3 4 5

028

03

032

034

036

038

1 2 3 4 5

1

2

3

4

5

times104

times104

times104

times104

times104

times104

times104

times104



Test 1198791198750 1198751

1198770 1198771

FLAIR-WML119903119894

080 (012) 069 (020) 087 (011)072 (023) 084 (012)

T1WMLri 098 (004) 098 (006) 099 (004)098 (008) 099 (005)

T1WM119903119894

097 (004) 096 (008) 099 (004)098 (008) 097 (006)

T1WML1199031198941199062

098 (004) 096 (008) 100 (000)100 (000) 097 (006)

T1WMLsvg119903119894 100 (000) 100 (000) 100 (000)100 (000) 100 (000)










Test 1198791198751 1198752

1198771 1198772

FLAIR-WML119903119894

073 (015) 078 (011) 020 (045)091 (012) 010 (032)

T1WML119903119894

066 (017) 074 (010) 000 (000)084 (018) 000 (000)


073 (016) 072 (018) 076 (017)075 (020) 071 (019)

T1WM119903119894

074 (016) 080 (009) 045 (051)075 (020) 071 (019)

T1WMSMOTE119903119894

068 (014) 067 (014) 075 (021)069 (029) 068 (014)


7 Discussion



















Acknowledgments


References







































































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











LBPPR =Pminus1

sump=0

s(gp minus gc)2p

Sample Threshold


1+ 0 lowast 2

2+ 0 lowast 2

3+ 0 lowast 2

4+ 1 lowast 2

5+ 0 lowast 2

6+ 1 lowast 2

7= 163


s(x) = 1 if x ⩾ 0

0 otherwise

83

52

94

88

97

61

14

77 82

0

1

1

1

0

0

0

1

x

(a)

52

94

88

97

61

14

77

83

82

5279

94

97

88

73

67

5461

54

14

4

23

23

83

23

82

(b)



VAR119875119877

=1

119875

119875minus1

sum119901=0

(119892119901minus 120583)2

where 120583 = 1

119875

119875minus1

sum119901=0

119892119901 (11)









(1)ROI extraction


and C







Table 1










1 1198792 119879

10) (1198751 1198752 119875

10) and

(1198771 1198772 119877



119898119909=1

119899

119899

sum119896=1

119909119896 where 0 le 119909

119896le 1 0 le 119898

119909le 1 (12)


(a) (b)

(c) (d)



119904119909= (

1

119899 minus 1

119899

sum119896=1

(119909119896minus 119898119909119896)2

)

12

where 0 le 119909119896le 1 0 le 119904

119909le 1

(13)







(a) (b)

(c) (d)



6 Results





119903119894








Training data

Test data

Test dataInner loop


feature sets



andpredictoutertest

results



measures



Test 1198791198750 1198751 1198752

1198770 1198771 1198772

FLAIR-WML119903119894

060 (013) 071 (028) 061 (014) 033 (041)048 (025) 077 (028) 020 (035)

T1WML119903119894

082 (012) 096 (010) 080 (011) 058 (049)098 (008) 088 (018) 025 (035)

T1WMLSMOTEri 087 (008) 097 (007) 081 (017) 085 (011)

100 (000) 082 (016) 078 (020)

T1WM119903119894

082 (009) 096 (008) 081 (011) 042 (049)100 (000) 088 (016) 020 (035)

T1WMSMOTE119903119894

075 (013) 090 (012) 066 (016) 070 (021)100 (000) 072 (019) 055 (022)

T1WMLSMOTEsvg119903119894 091 (015) 100 (000) 100 (000) 087 (022)

100 (000) 077 (042) 100 (000)







80 100 120 140 160

80

100

120

140

160

Var(

LBP)

R1P

8Var(LBP) R1P8

80 100 120 140 160

06

08

1

12

Skew(LBP) R1P8

80 100 120 140 160

028

03

032

034

036

038

Ent(LBP) R4P16

80 100 120 140 160

1

2

3

4

5

Mean(C) R4P16

06 08 1 12

80

100

120

140

160

Skew

(LBP

) R1P

8

06 08 1 12

06

08

1

12

06 08 1 12

028

03

032

034

036

038

06 08 1 12

1

2

3

4

5

028 032 036

80

100

120

140

160

Ent(L

BP) R

4P16

028 032 036

06

08

1

12

028 032 036

028

03

032

034

036

038

028 032 036

1

2

3

4

5

1 2 3 4 5

80

100

120

140

160

Mea

n(C

) R4P

16

1 2 3 4 5

06

08

1

12

1 2 3 4 5

028

03

032

034

036

038

1 2 3 4 5

1

2

3

4

5

times104

times104

times104

times104

times104

times104

times104

times104



Test 1198791198750 1198751

1198770 1198771

FLAIR-WML119903119894

080 (012) 069 (020) 087 (011)072 (023) 084 (012)

T1WMLri 098 (004) 098 (006) 099 (004)098 (008) 099 (005)

T1WM119903119894

097 (004) 096 (008) 099 (004)098 (008) 097 (006)

T1WML1199031198941199062

098 (004) 096 (008) 100 (000)100 (000) 097 (006)

T1WMLsvg119903119894 100 (000) 100 (000) 100 (000)100 (000) 100 (000)










Test 1198791198751 1198752

1198771 1198772

FLAIR-WML119903119894

073 (015) 078 (011) 020 (045)091 (012) 010 (032)

T1WML119903119894

066 (017) 074 (010) 000 (000)084 (018) 000 (000)


073 (016) 072 (018) 076 (017)075 (020) 071 (019)

T1WM119903119894

074 (016) 080 (009) 045 (051)075 (020) 071 (019)

T1WMSMOTE119903119894

068 (014) 067 (014) 075 (021)069 (029) 068 (014)


7 Discussion



















Acknowledgments


References







































































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(1)ROI extraction


and C







Table 1










1 1198792 119879

10) (1198751 1198752 119875

10) and

(1198771 1198772 119877



119898119909=1

119899

119899

sum119896=1

119909119896 where 0 le 119909

119896le 1 0 le 119898

119909le 1 (12)


(a) (b)

(c) (d)



119904119909= (

1

119899 minus 1

119899

sum119896=1

(119909119896minus 119898119909119896)2

)

12

where 0 le 119909119896le 1 0 le 119904

119909le 1

(13)







(a) (b)

(c) (d)



6 Results





119903119894








Training data

Test data

Test dataInner loop


feature sets



andpredictoutertest

results



measures



Test 1198791198750 1198751 1198752

1198770 1198771 1198772

FLAIR-WML119903119894

060 (013) 071 (028) 061 (014) 033 (041)048 (025) 077 (028) 020 (035)

T1WML119903119894

082 (012) 096 (010) 080 (011) 058 (049)098 (008) 088 (018) 025 (035)

T1WMLSMOTEri 087 (008) 097 (007) 081 (017) 085 (011)

100 (000) 082 (016) 078 (020)

T1WM119903119894

082 (009) 096 (008) 081 (011) 042 (049)100 (000) 088 (016) 020 (035)

T1WMSMOTE119903119894

075 (013) 090 (012) 066 (016) 070 (021)100 (000) 072 (019) 055 (022)

T1WMLSMOTEsvg119903119894 091 (015) 100 (000) 100 (000) 087 (022)

100 (000) 077 (042) 100 (000)







80 100 120 140 160

80

100

120

140

160

Var(

LBP)

R1P

8Var(LBP) R1P8

80 100 120 140 160

06

08

1

12

Skew(LBP) R1P8

80 100 120 140 160

028

03

032

034

036

038

Ent(LBP) R4P16

80 100 120 140 160

1

2

3

4

5

Mean(C) R4P16

06 08 1 12

80

100

120

140

160

Skew

(LBP

) R1P

8

06 08 1 12

06

08

1

12

06 08 1 12

028

03

032

034

036

038

06 08 1 12

1

2

3

4

5

028 032 036

80

100

120

140

160

Ent(L

BP) R

4P16

028 032 036

06

08

1

12

028 032 036

028

03

032

034

036

038

028 032 036

1

2

3

4

5

1 2 3 4 5

80

100

120

140

160

Mea

n(C

) R4P

16

1 2 3 4 5

06

08

1

12

1 2 3 4 5

028

03

032

034

036

038

1 2 3 4 5

1

2

3

4

5

times104

times104

times104

times104

times104

times104

times104

times104



Test 1198791198750 1198751

1198770 1198771

FLAIR-WML119903119894

080 (012) 069 (020) 087 (011)072 (023) 084 (012)

T1WMLri 098 (004) 098 (006) 099 (004)098 (008) 099 (005)

T1WM119903119894

097 (004) 096 (008) 099 (004)098 (008) 097 (006)

T1WML1199031198941199062

098 (004) 096 (008) 100 (000)100 (000) 097 (006)

T1WMLsvg119903119894 100 (000) 100 (000) 100 (000)100 (000) 100 (000)










Test 1198791198751 1198752

1198771 1198772

FLAIR-WML119903119894

073 (015) 078 (011) 020 (045)091 (012) 010 (032)

T1WML119903119894

066 (017) 074 (010) 000 (000)084 (018) 000 (000)


073 (016) 072 (018) 076 (017)075 (020) 071 (019)

T1WM119903119894

074 (016) 080 (009) 045 (051)075 (020) 071 (019)

T1WMSMOTE119903119894

068 (014) 067 (014) 075 (021)069 (029) 068 (014)


7 Discussion



















Acknowledgments


References







































































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b)

(c) (d)



119904119909= (

1

119899 minus 1

119899

sum119896=1

(119909119896minus 119898119909119896)2

)

12

where 0 le 119909119896le 1 0 le 119904

119909le 1

(13)







(a) (b)

(c) (d)



6 Results





119903119894








Training data

Test data

Test dataInner loop


feature sets



andpredictoutertest

results



measures



Test 1198791198750 1198751 1198752

1198770 1198771 1198772

FLAIR-WML119903119894

060 (013) 071 (028) 061 (014) 033 (041)048 (025) 077 (028) 020 (035)

T1WML119903119894

082 (012) 096 (010) 080 (011) 058 (049)098 (008) 088 (018) 025 (035)

T1WMLSMOTEri 087 (008) 097 (007) 081 (017) 085 (011)

100 (000) 082 (016) 078 (020)

T1WM119903119894

082 (009) 096 (008) 081 (011) 042 (049)100 (000) 088 (016) 020 (035)

T1WMSMOTE119903119894

075 (013) 090 (012) 066 (016) 070 (021)100 (000) 072 (019) 055 (022)

T1WMLSMOTEsvg119903119894 091 (015) 100 (000) 100 (000) 087 (022)

100 (000) 077 (042) 100 (000)







80 100 120 140 160

80

100

120

140

160

Var(

LBP)

R1P

8Var(LBP) R1P8

80 100 120 140 160

06

08

1

12

Skew(LBP) R1P8

80 100 120 140 160

028

03

032

034

036

038

Ent(LBP) R4P16

80 100 120 140 160

1

2

3

4

5

Mean(C) R4P16

06 08 1 12

80

100

120

140

160

Skew

(LBP

) R1P

8

06 08 1 12

06

08

1

12

06 08 1 12

028

03

032

034

036

038

06 08 1 12

1

2

3

4

5

028 032 036

80

100

120

140

160

Ent(L

BP) R

4P16

028 032 036

06

08

1

12

028 032 036

028

03

032

034

036

038

028 032 036

1

2

3

4

5

1 2 3 4 5

80

100

120

140

160

Mea

n(C

) R4P

16

1 2 3 4 5

06

08

1

12

1 2 3 4 5

028

03

032

034

036

038

1 2 3 4 5

1

2

3

4

5

times104

times104

times104

times104

times104

times104

times104

times104



Test 1198791198750 1198751

1198770 1198771

FLAIR-WML119903119894

080 (012) 069 (020) 087 (011)072 (023) 084 (012)

T1WMLri 098 (004) 098 (006) 099 (004)098 (008) 099 (005)

T1WM119903119894

097 (004) 096 (008) 099 (004)098 (008) 097 (006)

T1WML1199031198941199062

098 (004) 096 (008) 100 (000)100 (000) 097 (006)

T1WMLsvg119903119894 100 (000) 100 (000) 100 (000)100 (000) 100 (000)










Test 1198791198751 1198752

1198771 1198772

FLAIR-WML119903119894

073 (015) 078 (011) 020 (045)091 (012) 010 (032)

T1WML119903119894

066 (017) 074 (010) 000 (000)084 (018) 000 (000)


073 (016) 072 (018) 076 (017)075 (020) 071 (019)

T1WM119903119894

074 (016) 080 (009) 045 (051)075 (020) 071 (019)

T1WMSMOTE119903119894

068 (014) 067 (014) 075 (021)069 (029) 068 (014)


7 Discussion



















Acknowledgments


References







































































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b)

(c) (d)



6 Results





119903119894








Training data

Test data

Test dataInner loop


feature sets



andpredictoutertest

results



measures



Test 1198791198750 1198751 1198752

1198770 1198771 1198772

FLAIR-WML119903119894

060 (013) 071 (028) 061 (014) 033 (041)048 (025) 077 (028) 020 (035)

T1WML119903119894

082 (012) 096 (010) 080 (011) 058 (049)098 (008) 088 (018) 025 (035)

T1WMLSMOTEri 087 (008) 097 (007) 081 (017) 085 (011)

100 (000) 082 (016) 078 (020)

T1WM119903119894

082 (009) 096 (008) 081 (011) 042 (049)100 (000) 088 (016) 020 (035)

T1WMSMOTE119903119894

075 (013) 090 (012) 066 (016) 070 (021)100 (000) 072 (019) 055 (022)

T1WMLSMOTEsvg119903119894 091 (015) 100 (000) 100 (000) 087 (022)

100 (000) 077 (042) 100 (000)







80 100 120 140 160

80

100

120

140

160

Var(

LBP)

R1P

8Var(LBP) R1P8

80 100 120 140 160

06

08

1

12

Skew(LBP) R1P8

80 100 120 140 160

028

03

032

034

036

038

Ent(LBP) R4P16

80 100 120 140 160

1

2

3

4

5

Mean(C) R4P16

06 08 1 12

80

100

120

140

160

Skew

(LBP

) R1P

8

06 08 1 12

06

08

1

12

06 08 1 12

028

03

032

034

036

038

06 08 1 12

1

2

3

4

5

028 032 036

80

100

120

140

160

Ent(L

BP) R

4P16

028 032 036

06

08

1

12

028 032 036

028

03

032

034

036

038

028 032 036

1

2

3

4

5

1 2 3 4 5

80

100

120

140

160

Mea

n(C

) R4P

16

1 2 3 4 5

06

08

1

12

1 2 3 4 5

028

03

032

034

036

038

1 2 3 4 5

1

2

3

4

5

times104

times104

times104

times104

times104

times104

times104

times104



Test 1198791198750 1198751

1198770 1198771

FLAIR-WML119903119894

080 (012) 069 (020) 087 (011)072 (023) 084 (012)

T1WMLri 098 (004) 098 (006) 099 (004)098 (008) 099 (005)

T1WM119903119894

097 (004) 096 (008) 099 (004)098 (008) 097 (006)

T1WML1199031198941199062

098 (004) 096 (008) 100 (000)100 (000) 097 (006)

T1WMLsvg119903119894 100 (000) 100 (000) 100 (000)100 (000) 100 (000)










Test 1198791198751 1198752

1198771 1198772

FLAIR-WML119903119894

073 (015) 078 (011) 020 (045)091 (012) 010 (032)

T1WML119903119894

066 (017) 074 (010) 000 (000)084 (018) 000 (000)


073 (016) 072 (018) 076 (017)075 (020) 071 (019)

T1WM119903119894

074 (016) 080 (009) 045 (051)075 (020) 071 (019)

T1WMSMOTE119903119894

068 (014) 067 (014) 075 (021)069 (029) 068 (014)


7 Discussion



















Acknowledgments


References







































































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











Training data

Test data

Test dataInner loop


feature sets



andpredictoutertest

results



measures



Test 1198791198750 1198751 1198752

1198770 1198771 1198772

FLAIR-WML119903119894

060 (013) 071 (028) 061 (014) 033 (041)048 (025) 077 (028) 020 (035)

T1WML119903119894

082 (012) 096 (010) 080 (011) 058 (049)098 (008) 088 (018) 025 (035)

T1WMLSMOTEri 087 (008) 097 (007) 081 (017) 085 (011)

100 (000) 082 (016) 078 (020)

T1WM119903119894

082 (009) 096 (008) 081 (011) 042 (049)100 (000) 088 (016) 020 (035)

T1WMSMOTE119903119894

075 (013) 090 (012) 066 (016) 070 (021)100 (000) 072 (019) 055 (022)

T1WMLSMOTEsvg119903119894 091 (015) 100 (000) 100 (000) 087 (022)

100 (000) 077 (042) 100 (000)







80 100 120 140 160

80

100

120

140

160

Var(

LBP)

R1P

8Var(LBP) R1P8

80 100 120 140 160

06

08

1

12

Skew(LBP) R1P8

80 100 120 140 160

028

03

032

034

036

038

Ent(LBP) R4P16

80 100 120 140 160

1

2

3

4

5

Mean(C) R4P16

06 08 1 12

80

100

120

140

160

Skew

(LBP

) R1P

8

06 08 1 12

06

08

1

12

06 08 1 12

028

03

032

034

036

038

06 08 1 12

1

2

3

4

5

028 032 036

80

100

120

140

160

Ent(L

BP) R

4P16

028 032 036

06

08

1

12

028 032 036

028

03

032

034

036

038

028 032 036

1

2

3

4

5

1 2 3 4 5

80

100

120

140

160

Mea

n(C

) R4P

16

1 2 3 4 5

06

08

1

12

1 2 3 4 5

028

03

032

034

036

038

1 2 3 4 5

1

2

3

4

5

times104

times104

times104

times104

times104

times104

times104

times104



Test 1198791198750 1198751

1198770 1198771

FLAIR-WML119903119894

080 (012) 069 (020) 087 (011)072 (023) 084 (012)

T1WMLri 098 (004) 098 (006) 099 (004)098 (008) 099 (005)

T1WM119903119894

097 (004) 096 (008) 099 (004)098 (008) 097 (006)

T1WML1199031198941199062

098 (004) 096 (008) 100 (000)100 (000) 097 (006)

T1WMLsvg119903119894 100 (000) 100 (000) 100 (000)100 (000) 100 (000)










Test 1198791198751 1198752

1198771 1198772

FLAIR-WML119903119894

073 (015) 078 (011) 020 (045)091 (012) 010 (032)

T1WML119903119894

066 (017) 074 (010) 000 (000)084 (018) 000 (000)


073 (016) 072 (018) 076 (017)075 (020) 071 (019)

T1WM119903119894

074 (016) 080 (009) 045 (051)075 (020) 071 (019)

T1WMSMOTE119903119894

068 (014) 067 (014) 075 (021)069 (029) 068 (014)


7 Discussion



















Acknowledgments


References







































































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










80 100 120 140 160

80

100

120

140

160

Var(

LBP)

R1P

8Var(LBP) R1P8

80 100 120 140 160

06

08

1

12

Skew(LBP) R1P8

80 100 120 140 160

028

03

032

034

036

038

Ent(LBP) R4P16

80 100 120 140 160

1

2

3

4

5

Mean(C) R4P16

06 08 1 12

80

100

120

140

160

Skew

(LBP

) R1P

8

06 08 1 12

06

08

1

12

06 08 1 12

028

03

032

034

036

038

06 08 1 12

1

2

3

4

5

028 032 036

80

100

120

140

160

Ent(L

BP) R

4P16

028 032 036

06

08

1

12

028 032 036

028

03

032

034

036

038

028 032 036

1

2

3

4

5

1 2 3 4 5

80

100

120

140

160

Mea

n(C

) R4P

16

1 2 3 4 5

06

08

1

12

1 2 3 4 5

028

03

032

034

036

038

1 2 3 4 5

1

2

3

4

5

times104

times104

times104

times104

times104

times104

times104

times104



Test 1198791198750 1198751

1198770 1198771

FLAIR-WML119903119894

080 (012) 069 (020) 087 (011)072 (023) 084 (012)

T1WMLri 098 (004) 098 (006) 099 (004)098 (008) 099 (005)

T1WM119903119894

097 (004) 096 (008) 099 (004)098 (008) 097 (006)

T1WML1199031198941199062

098 (004) 096 (008) 100 (000)100 (000) 097 (006)

T1WMLsvg119903119894 100 (000) 100 (000) 100 (000)100 (000) 100 (000)










Test 1198791198751 1198752

1198771 1198772

FLAIR-WML119903119894

073 (015) 078 (011) 020 (045)091 (012) 010 (032)

T1WML119903119894

066 (017) 074 (010) 000 (000)084 (018) 000 (000)


073 (016) 072 (018) 076 (017)075 (020) 071 (019)

T1WM119903119894

074 (016) 080 (009) 045 (051)075 (020) 071 (019)

T1WMSMOTE119903119894

068 (014) 067 (014) 075 (021)069 (029) 068 (014)


7 Discussion



















Acknowledgments


References







































































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











Test 1198791198751 1198752

1198771 1198772

FLAIR-WML119903119894

073 (015) 078 (011) 020 (045)091 (012) 010 (032)

T1WML119903119894

066 (017) 074 (010) 000 (000)084 (018) 000 (000)


073 (016) 072 (018) 076 (017)075 (020) 071 (019)

T1WM119903119894

074 (016) 080 (009) 045 (051)075 (020) 071 (019)

T1WMSMOTE119903119894

068 (014) 067 (014) 075 (021)069 (029) 068 (014)


7 Discussion



















Acknowledgments


References







































































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation

















Acknowledgments


References







































































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation
































































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









Documents

Research Article Classifying Dementia Using Local …downloads.hindawi.com/journals/ijbi/2015/572567.pdfResearch Article Classifying Dementia Using Local Binary ... e paper is organised