alzheimer’sdiseasediagnosisbasedoncorticaland...

Research ArticleAlzheimerrsquos Disease Diagnosis Based on Cortical andSubcortical Features

Yubraj Gupta12 Kun Ho Lee32 Kyu Yeong Choi2 Jang Jae Lee2 Byeong Chae Kim42

and Goo-Rak Kwon 12

1School of Information Communication Engineering Chosun University 309 Pilmun-Daero Dong-GuGwangju 61452 Republic of Korea2National Research Center for Dementia Chosun University 309 Pilmun-Daero Dong-Gu Gwangju 61452 Republic of Korea3Department of Biomedical Science College of Natural Sciences Chosun University 309 Pilmun-Daero Dong-GuGwangju 61452 Republic of Korea4Department of Neurology Chonnam National University Medical School Gwangju 61469 Republic of Korea

Correspondence should be addressed to Goo-Rak Kwon grkwonchosunackr

Received 19 July 2018 Revised 8 September 2018 Accepted 13 February 2019 Published 3 March 2019

Guest Editor Qiu-Hua Lin

Copyright copy 2019 Yubraj Gupta et al 0is 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

Alzheimerrsquos disease (AD) is a common neurodegenerative disease with an often seen prodromal mild cognitive impairment (MCI)phase wherememory loss is the main complaint progressively worsening with behavior issues and poor self-care However not allpatients clinically diagnosed with MCI progress to the AD Currently several high-dimensional classification techniques havebeen developed to automatically distinguish among AD MCI and healthy control (HC) patients based on T1-weighted MRIHowever thesemethod features are based on wavelets contourlets gray-level co-occurrencematrix etc rather than using clinicalfeatures which helps doctors to understand the pathological mechanism of the AD In this study a new approach is proposedusing cortical thickness and subcortical volume for distinguishing binary and tertiary classification of the National ResearchCenter for Dementia dataset (NRCD) which consists of 326 subjects Five classification experiments are performed binaryclassification ie AD vs HC HC vs mAD (MCI due to the AD) and mAD vs aAD (asymptomatic AD) and tertiary classificationie AD vs HC vs mAD and AD vs HC vs aAD using cortical and subcortical features Datasets were divided in a 7030 ratio andlater 70 were used for training and the remaining 30 were used to get an unbiased estimation performance of the suggestedmethods For dimensionality reduction purpose principal component analysis (PCA) was used After that the output of PCA waspassed to various types of classifiers namely softmax support vector machine (SVM) k-nearest neighbors and naıve Bayes (NB)to check the performance of the model Experiments on the NRCD dataset demonstrated that the softmax classifier is best suitedfor the AD vs HC classification with an F1 score of 9906 whereas for other groups the SVM classifier is best suited for the HC vsmAD mAD vs aAD and AD vs HC vs mAD classifications with the F1 scores being 9951 975 and 9999 respectively Inaddition for the AD vs HC vs aAD classification NB performed well with an F1 score of 9588 In addition to check our proposedmodel efficiency we have also used the OASIS dataset for comparing with 9 state-of-the-art methods

1 Introduction

Alzheimerrsquos disease (AD) is the most usual neurodegener-ative dementia and a rapidly growing health problem whichis the main reason for causing dementia in the elderlypopulation [1] A conclusive analysis can only be madepostmortem and requires histopathological confirmation ofneurofibrillary tangles and amyloid plaques Early and

correct diagnosis of AD is not only difficult but also vitalfrom the perspective of future treatments

Although there is currently no treatment for AD thereare several promising pharmacologic compounds in ad-vanced stages of development and it is expected that atreatment will be soon available One of the factors that havebeen speculated to be the cause of lack of success in de-veloping disease-modifying treatments thus far is the

HindawiJournal of Healthcare EngineeringVolume 2019 Article ID 2492719 13 pageshttpsdoiorg10115520192492719

inability to correctly differentiate patients at the mild cog-nitive impairment (MCI) stage who will progress on todevelop AD from those that have MCI symptoms that aredue to other causes A potentially promising treatment forADmay not be beneficial when administered to those whoseMCI symptoms are not due to AD causing an apparentlypromising treatment to be considered a failure Anotherreason for distinguishing those patients with prodromal ADat the MCI stage is that interventions early in the course ofthe disease may help delay the onset andor mitigate the riskof full-blown AD [2] whereas interventions later in thecourse of the disease may limit the disease but would not beable to reverse the pathology-induced neuronal loss after ithas already occurred Hence diagnosis of presymptomaticAD in the MCI stage is a highly important task which willbecome even more crucial and urgent when curativetreatment becomes available Such diagnosis can also bevaluable for those whose MCI is due to causes other thanAD as these causes may be easier to determine and manageHence neuroimaging improves the positive predictive valueof the diagnosis and involves measurements by structuralMRI to test medial temporal lobe atrophy (MTL) andpositron emission tomography using fluorodeoxyglucose(FDG) or amyloid markers MTL structures are fundamentalfor the creation of new memories and are centrally related tothe growth of AD [3] Specifically MTA atrophy and alliedepisodic memory impairment are trademark features of ADand they both progressively lessen over the course of thedisease [4] Normally MTA is determined by using vertex-based [5] voxel-based [6] and ROI-based [7] approaches

In this study the focus is on binary and tertiary clas-sification between AD HC mAD (MCI due to AD) andaAD (asymptomatic AD) using structural MRI By using theatrophy measure from sMRI scans the intensity and stage ofthe neurodegeneration can be determined 0ese studiesinclude morphometric approaches such as the region ofinterest (ROI) and volume of interest (VOI) of gray mattervoxels for the automatic segmentation of sMRI images andthe sMRI volume measurement of the hippocampus and themedial progressive lobe [8] Various machine-learningmethods have been used to differentiate binary and ter-tiary classification between AD HC mAD and aAD Assuggested in [9 10] only using amyloid imaging biomarkermay be less sensitive in tracing AD progression mainly inthe symptomatic stage recent consensus statements haveunderlined the importance of a biomarker of neuro-degeneration which is a critical component of AD patho-physiology in the prodromal and early dementia stages andalso from [11] atrophy on sMRI reflects on cumulative lossand shrinkage of the neuropil [12ndash14] 0ese indicate that avolumetric measure of cortical thickness and subcorticalvolume is one essential biomarker for early detection of theAD Cortical thickness analysis establishes as another widelyaccepted approach to measure gray matter atrophy in theAD and cortical thinning has been found in MCI and ADDespite evidence for subcortical amyloid and neurofibrillarytangle formation in AD [15] MRI research has drawn itsattention to AD-related subcortical structure changes onlyrecently Advanced segmentation techniques now permit the

quantification of subcortical volumes and provide the basisfor subcortical shape analysis Measurement of structuralchanges based on brain MRI scans has previously been usedto classify AD patients versus cognitively normal (CN)subjects and to predict the risk of progression from MCI tothe AD Jha et al [16] proposed a method for diagnosis of theAD using complex dual-tree wavelet principal coefficientstransform (DTCWT) 0ey have used two datasets OASISand ADNI for the validation of their results An extremelearning machine was used as a classifier and an accuracy of9026 sensitivity of 9027 and specificity of 9027 wereachieved for the 2D ADNI dataset and an accuracy of9575 a sensitivity of 9659 and a specificity of 9303were achieved for the OASIS dataset Chyzhyk et al [17] usesa lattice independent component analysis (LICA) techniquefor the feature section stage with combine kernel trans-formation of the data 0eir approach improved the gen-eralization of the dendritic computing classifiers 0en theyapply their proposed model on the OASIS dataset forclassification of AD patients with normal subjects and theirmodel achieve 7475 accuracy 96 sensitivity and 525specificity Khajehnejad et al [18] uses the manifold-basedsemisupervised learning method for early diagnosis of theAD 0ey have used the OASIS dataset for experiment andtheir proposed method yield accuracy of 9386 sensitivityof 9465 and specificity of 9322 Jha and Kwon [19]proposed an AD detection method using sparse autoen-coder scale conjugate gradient and softmax output layerwith fine-tuning 0ey have used the OASIS dataset andachieved an accuracy of 916 9809 of sensitivity and8409 of specificity Islam and Zhan [20] proposed anensemble of a deep convolutional neural network for earlydetection of AD Here they have used DenseNet-121DenseNet-161 and DenseNet-169 deep learning modelsfor classification of the OASIS dataset and later they haveused their proposed ensemble model which was fused withdeep learning for classification of the same dataset 0eirproposed method yields 9318 of accuracy 94 of pre-cision 93 of sensitivity and 92 of F1 score Farhan et al[21] proposed a model which uses two extracted features ofthe brain first features are the volume of GMWM and CSFand second features are the (l hrh) area of the hippocampusIn addition later they have used four different classifiers(SVM MLP j48 and ensemble of classifiers) to evaluate theclassification accuracy An ensemble of classifiers has got thehigh accuracy compared to other classifiers with 9375accuracy for combined features Lama et al [22] proposed abinary and a tertiary classification between AD MCI andNC using sMRI data from the ADNI dataset A regularizedextreme learning machine (RELM) was used as a classifierand an accuracy of 7730 for binary classification (AD vsNC) and 7661 for tertiary classification (AD vs NC vsMCI) were obtained cortical thickness features were used inthe experiment Based on a large subgroup of ADNI dataCuingnet et al [23] compared ten sMRI-based feature ex-traction techniques and their ability to distinguish betweenclinically relevant subject groups 0ese methods comprisedfive voxel-based techniques three methods based on corticalthickness and two techniques based on the hippocampus

2 Journal of Healthcare Engineering

Optimum sensitivity and specificity values were (81 95)for AD vs HC (70 61) for S-MCI vs P-MCI and (7385) for HC vs P-MCI Zhang et al [24] proposed amultimodal classification approach by employing amultiple-kernel support vector machine (SVM) based onbiomarkers including sMRI PET and cerebrospinal fluid(CSF) to distinguish AD (or MCI) and normal control (NC)subjects For the binary classification (AD vs NC andMCI vsNC) results their suggested model had high accuracy for ADclassification whereas for MCI classification satisfactoryaccuracy was obtained More recently Cho et al [25]conducted an experiment on 72 MCIc and 131 MCIncsubjects using the incremental learning technique based onspatial frequency which shows the representation of corticalthickness data In addition their proposed method yieldedbetter results than the ten benchmark techniques for MCIcvs MCInc classification as reported in [23] and a sensitivityof 63 and a specificity of 76 were obtained Wolz et al[26] used four different automated feature extractiontechniques (namely hippocampal volume TBM corticalthickness and manifold-based learning) to analyze struc-tural MRI data of 834 ADNI AD MCI and healthy control(CTL) subjects 0e extracted features were used to comparethe performance of two classifiers namely LDA and SVMfor AD classification and MCI prediction 0e best accuracyfor AD versus CTL classification was obtained by combiningall extracted features and utilizing a LDA classifier that is anaccuracy of 89 (sensitivity and specificity of 93 and 85respectively) Similarly using combined features and theLDA classifier resulted in the highest accuracy of 68(sensitivity and specificity of 67 and 69 respectively) forclassification of MCI-converter and MCI-stable subjectsWhen different feature types were studied individually theTBM features yielded the best result

Compared to earlier work this study aims at establishingthe enhancement in accuracy and constancy that can beattained by combining more than one MR-based featureHere we are going to investigate a combined feature ofcortical thickness and subcortical volume in the AD andaADwith stable cognitive abilities compared toHC as well asin mAD which converted to AD NRCD data were used totest the potential combination of diverse MR-based featuresfor improving classification accuracy For evaluation all 326NRCD dataset images provided by the dementia center ofChosun University hospital were used and NeuroI [27]which was developed at NRCD was also used Specificallythe binary classifications AD vs HC HC vs mAD and mADvs aAD and the tertiary classifications AD vs HC vs mADand AD vs HC vs aAD were used To this end four rep-resentative classifiers are presented and compared using anefficient feature selection approach including SVM k-nearest neighbors (KNN) naıve Bayes (NB) and softmaxclassifiers for themulticlass classification of various stages ofthe AD progression

2 Materials and Methods

21 Subjects 0e data utilized in this research were obtainedfrom the National Research Center for Dementia (NRCD)

All the patients for whom preprocessed images were availablewere selected 0e dataset contained 326 subjects 81 ADsubjects (39 males 42 females ageplusmn SD 7186plusmn 709 yearsrange 56ndash83 years education level 734plusmn 488 range 0ndash18) 171 cognitively normal health control (HC) subjects (83males 88 females ageplusmn SD 7166plusmn 543 range 60ndash85education level 916plusmn 554 range 0ndash22) 39 patients withmAD (MCI who had converted to AD) (25 males 14 femalesageplusmn SD 7323plusmn 709 range 49ndash87 education level820plusmn 519 range 0ndash18) 35 patients with aAD (MCI whohad not converted) (15 males 20 females ageplusmn SD

7274plusmn 482 range 61ndash83 education level 788plusmn 630range 0ndash18)

Table 1 shows the demographics of the 326 study sub-jects Statistical significant differences in demographics andclinical variables between the study groups were measuredusing Studentrsquos unpaired t-test In this study the significancelevel was set to 005 which is a standard alpha value 0erewere more females in all groups except for the mAD group0e education level was highly dissimilar in the pairwisecomparisons between all other study groups Compared tothe patients in all other groups AD patients had significantlya lower education level

To obtain unbiased estimations of the performance theset of contestants was randomly split into two parts in a 7030 ratio for training and testing 0e algorithm was trainedon a training code and the measures of the group perfor-mance were evaluated in terms of accuracy sensitivityspecificity precision and F1 score using an independent testset

22MRIAcquisition Standard 3T T1-weighted images wereobtained using the volumetric 3D MPRAGE protocol withresolution 1mmtimes 1mmtimes 1mm (voxel size) 0ese were N4bias correction images

23 Feature Selection Subcortical volumetric and corticalthickness measures have been widely used for classificationpurposes From [28] we can say that cortical thickness is adirect index of atrophy and is therefore a potentiallypowerful candidate in the diagnosis of AD In this experi-ment cortical thickness and subcortical volume areextracted using Freesurfer (v53) In total 110 features wereextracted from 3D sMRI T1-weighted image Freesurfer is aset of tools for analysis and visualization of cortical (Fig-ure 1) and subcortical (Figure 2) brain imagining dataFreesurfer was designed around an automated workflow thatincludes several standard image processing phases which arenecessary to achieve a final brain parcellation image withinthe subjectrsquos space however manual image editing is alsoallowed after each stage to ensure quality control (httpsurfernmrmghharvardedufswiki) At first a good qualityof T1-weighted sMRI image volume is passed to the Free-surfer pipeline and after that Freesurfer performs a series ofsteps start from motion artifact correction affine trans-formation (12 degrees of freedom) to Talairach image spacenonuniform intensity normalization for intensity in-homogeneity correction and removal of nonbrain tissues

Journal of Healthcare Engineering 3

0e remaining brain image volume is intensity normalizedto tie with the Freesurfer atlas image intensity histogramwhich is followed by a nonlinear warping of the atlas brainimage to subject brain image A warped Atlas 3D brainimage in the subject space is utilized in atlas-based tissuesegmentation for labelling subcortical brain structures likebrain stem cerebellum and cerebral cortex 0e next step inFreesurfer is to generate topologically correct cortical surfacerepresentation per hemisphere like gray-white mattersegmentation [29] and the third segments of 34 ROIs basedon anatomic landmarks [30]

Freesurfer also provides the ability to construct asurface-based morphometry (SBM) for representations ofthe cortex from which neuroanatomic volume corticalthickness and surface area can be derived Cortical surfacelies either at the WMGM tissue interface or at the GMCSFtissue interface Each hemispherersquos cortical surface repre-sentation is mapped automatically to a standard sphericalcoordinate system Key components of the surface mappinginclude surface inflation with minimal metric distortiontopology correction projection to spherical coordinates andSBM warping to correctly align anatomical homologouspoints Mapping to the standard spherical coordinate systemdefined by Freesurfer atlas brain allows for automated an-atomical parcellation of cortex into gyral regions Surfaceparcellation is then extended to GM volume yielding par-cellation of GM tissue sheet and regional cortical volumes0e precise matching of a morphologically homologouscortical location for each patient was obtained by plottingthe atlas-based cortical surface on a sphere aligning thecortical patterns Table 2 presents an overview of the features

calculated for all 326 available NRCD images After thepreprocessing stage all the data extracted by Freesurfer werenormalized to zero mean and unit variance for each featureas shown in Figure 3 using standard scalar function of Scikit-learn library 0at is given the data matrix X where rowsrepresent subjects and columns represent features thenormalized matrix with elements x(i j) is given by

Xnorm(ij) x(ij) minusmean xj1113872 1113873

std xj1113872 1113873 (1)

where Xj is the jth column of the matrix (X) Subsequentlyprincipal component analysis (PCA) was performed [31]which is a dimensionality reduction technique whereby thefeatures are mapped onto a lower dimensional space PCA isa feature selection process generating new features that arelinear combinations of the initial features PCA maps thedata in a d-dimensional space to a new k-dimensionalsubspace with klt d 0e new k variables are called principalcomponents (PC) and each principal component hasmaximum variance eliminating the variance that is alreadyaccounted for in all succeeding components Consequentlythe first component has a larger variance than the com-ponents that follow 0e principal components can be de-fined by the following equation

PCi a1b1 + a2b2 + middot middot middot + adbd (2)

where PCi is a principal component in i xd is an originalfeature in d and ad is a numerical coefficient of xd 0enumber of principal components is always less than or equalto the number of original observations

0e obtained principal components for AD vs HC areshown in Figure 4 0e number of components was de-termined by maintaining the variance greater than 99 Intotal there were 110 features however not all features wererequired for convergence As can be seen in Figure 4 the firstprincipal component achieved 99 of the variance com-pared to the other features 0erefore the values that wereobtained first were taken as a principal component for AD vsHC Likewise for other binary and tertiary classification thesame procedure was followed

24 Classification Method Once the sMRI-based featureswere extracted from the automated toolbox in both corticaland subcortical regions feature vectors containing mean-centered voxel intensities were created combining allfeatures 0e classification approach we design is aimed tocombine two sources of features cortical thickness and

Table 1 Demographic characteristics of the studied population (from the NRCD database)

Group Number of subjects AgeGender

EducationM F

AD 81 7186plusmn 709 [56ndash83] 39 42 734plusmn 488 [0ndash18]mAD 39 7323plusmn 734 [49ndash87] 25 14 820plusmn 519 [0ndash18]aAD 35 7274plusmn 482 [61ndash83] 15 20 788plusmn 630 [0ndash18]HC 171 7166plusmn 543 [60ndash85] 83 88 916plusmn 554 [0ndash22]Values are specified as meanplusmn standard deviation [range]

Figure 1 Cortical thickness showing left and right hemisphere ofpial and white surface


subcortical volume ese features are used for globaldecision framework to discriminate AD with other groupse overall diagram of the approach is presented in Fig-ure 3 Here our aim is early classication of an AD fromother groups by using combined features of corticalthickness and subcortical volume of the same subjects Ouraim is to classify binary and tertiary groups Moreoverclassication is the problem of categorizing to which a newset of observation belongs based on a training datasetwhose group aliation is already known Moreover weknow that di erent classiers follow a di erent mathe-matical function for classication at is why we haveused four best machine-learning classiers which are in

practice like SVM and its variants KNN NB and softmaxclassiers to know which classier learns our model ac-curately to classify the di erent groups and we alsocompared their results All classication methods give adecision output

25 SVM is is the most common classication method toanalyze sMRI data [23] SVM [32] is the most commonlyused algorithm in AD research for multivariate classication[26]ismethod is based on choosing a critical point for theclassication task Support vectors are the elements of thedata that are relevant in the splitting into two classes e

Table 2 Features used in this study

Methods Number of features Descriptions

Cortical thickness and subcortical volumesegmentation 110

Based on group-level statistical analysis measure ofsubcortical segmentation volume and cortical

thickness

(a) (b) (c)

Figure 2 Subcortical brain segmentation

3D input image

Skull stripping and itsslices White and pial surface

Cortical thickness and subcortical volumesegmentation feature measure

Normalization

PCA

5-fold stratified K-fold CV

Classifier Cross-validation

Feature selection

Diagnosis output

AD

HCKNN

Naiumlve Bayes

Softmax classifier

SVMaAD

mAD

(e)(a)

(b)

(h) (g)

(d)(c)

(f)

+

Figure 3 Workow of proposed method (a) 3D input image (b) skull stripping (c) white and pial surface (d) cortical thickness andsubcortical volume segmentation features (e) feature selection (f) cross validation (g) various classiers and (h) diagnosis output


SVM algorithm determines the parameters of the decisionfunction that maximizes the margin between the trainingexamples and the class boundary as shown in Figure 5 Anumber of nonlinear SVM approaches have also been usedsuch as kernel SVM and multikernel SVM [33]

0e main concept of kernel techniques is to map theinput data which are linearly nonseparable into a higherdimensional feature space where they are more likely to belinearly separable

26 KNN Cover and Hart proposed KNN in 1968 [34]which is one of the simplest machine-learning algorithms Itis an extension of the simple nearest neighbor techniqueKNN classifies an unknown sample depending on the ldquovoterdquoof the k-nearest neighbors rather than the single nearestneighbor

0e main steps of KNN implementation are as follows

(1) Similarity assessment the similarity between thetest sample and each sample of the training set iscalculated In general the resemblance can bemeasured by for instance the Euclidean distancethe Manhattan distance the Jacquard similaritycoefficient and the correlation coefficient Amongthese the Euclidean distance method is the mostwidely used For a given feature sampletest(xj1 xj2 xji) and training set featuretrain(xj1 xj2 xjn) the Euclidean distance iscalculated as follows

dj

1113944

n

i1testji minus trainjn1113872 1113873

11139741113972

(3)

where n is the number of feature vectors j is thenumber of training and testing trials and dj is theEuclidean distance between the jth sample of thetraining set and a test sample

(2) In the second step the neighborrsquos nearest distancesshould be determined and sorted in ascending order

0ere are several methods (like elbow method andthumb rule) which help us to achieve the best k valuefor KNN 0e optimum k value will always varydepending on the dataset so it should be as big asthat its noises will not disturb the prediction highlyand as low as that one factor will not dominateanother 0en the selection of the value will directlyaffect the classification result and should thus becarefully made

(3) In the third step the ldquovote and classifyrdquo method isapplied whereby the test sample is classified to aclass according to the voting result of eachcategory

27 NB NB is a machine-learning technique that has beenutilized for over 50 years in biomedical informatics [35 36]0ese classifiers are a family of simple ldquoprobabilistic clas-sifiersrdquo based on Bayesrsquo theorem with strong independenceassumptions between their features An NB classificationmodel is easy to construct with no complex iterative pa-rameter approximation which makes it particularly usefulfor large datasets Despite its simplicity the NB model oftenperforms unexpectedly well and is widely used for classifi-cation because it often outperforms several more sophisti-cated classification approaches Bayesrsquo theorem is used todetermine the posterior probability p(cx) from p(c) p(x)and p(xc) 0e NB classifier assumes that the effect of theparameter of a predictor (x) on a specified class (c)

is autonomous compared with the other predictorrsquos pa-rameters 0is assumption is called class-conditionalindependence

pc

x1113874 1113875

p(xc)p(c)

p(x) (4)

where p(cx) p(x1c)lowastp(x2c)lowast middot middot middot lowastp(xnc)lowastp(c) middot

p(cx) is the posterior likelihood of the class (target) giventhe predictor (attribute) p(c) is the prior probability of theclass p(xc) is the probability of the predictor given theclass and p(x) is the prior probability of the predictor

12E + 12

1E + 12

8E + 11

6E + 11

No

of c

ompo

nent

s

4E + 11

2E + 11

0

1 5 9 13 17 21 25 29 33 37 41 45 49 53 57

No of variances

61 65 69 73 77 81 85 89 93 97 101

105

109

Figure 4 Number of principal components vs number of variances


28 Softmax Classifier Softmax regression is a generalizedform of logistic regression that can be used in multiclassclassification problems where the classes are mutually ex-clusive0e earlier linear regression model [19 37] producescontinuous or unbounded y values To proceed from ar-bitrary values y isin Rc to normalized probability valuesp isin Rc is estimated for a single instance and exponentiationand normalization are used as follows

pi expyi

1113936cc1expyc

(5)

where i c isin 1 C is the range over the classes pi refersto the probabilities and yi and yc refer to the value of asingle instance 0is is called the softmax function whichtakes the vector of arbitrary real-valued scores in y andtransforms it into a vector of values ranging between 0 and1 that later sum to 1 A model that converts the un-normalized values at the end of a linear regression tonormalized probabilities for classification is called softmaxclassifier Here the softmax layer takes the learned rep-resentation pi and interprets it to the output class Aprobability score pi is also assigned to the output class 0esoftmax classifier understands the scores inside the outputvector y as the un-normalized log likelihoods for eachspecific class and then replaces the hinge losses with across-entropy loss that has the form

Li minuslog pi( 1113857 (6)

where Li is the loss of cross entropy of the network Ex-ponentiation of these quantities yields the (un-normalized)likelihoods and their division performs the normalizationso that their probabilities sum to 1 In probabilistic termsthe negative log probability of the accurate class is mini-mized which can be regarded as performing maximumlikelihood estimation (MLE) A fine feature of this ob-servation is that the regularization term R (W) can now beinterpreted as the full loss function as approaching from aGaussian prior over the weight matrix W where we areexecuting the maximum a posteriori (MAP) estimationinstead of MLE

3 Results and Discussion

31 Background In this study the proposed method ispresented using SVM KNN NB and softmax classifiers0e details are shown in Figure 3 0e proposed idea is tocombine two extracted features namely cortical thicknessand subcortical volume from Freesurfer toolbox to dif-ferentiate between AD vs other groups For early pre-diction we perform a binary and tertiary classificationusing the AD HC mAD and aAD datasets to check howwell our proposed method has performed on the sMRI 3Dimage At the preprocessing stage normalization wasperformed for each case Moreover the PCA di-mensionality reduction technique is utilized to find theoptimal number of principal components for each classi-fication as shown in Figure 4 for AD vs HC Differentnumbers of principal components were obtained for dif-ferent cases In addition stratified K-Fold (SKF) cross-validation (CV) was used to confirm the robustness of theclassification results Four different types of classifiers wereused to obtain the test sample accuracy sensitivity spec-ificity precision and F1 score

32 Evaluation To obtain unbiased estimations of theperformance the set of subjects was randomly separated intotwo parts in a 7030 ratio for training and testing 0etraining set was utilized to determine the optimal hyper-parameters of each classifier or method and later to train theclassifier Subsequently the testing set was used to evaluatethe classification performance 0e training and testingdatasets were identical for all techniques On the training setSKF CV was used to estimate the optimal hyperparametersTwo kernel parameters c and c for radical basic functionSVM and one kernel value c for linear SVMwere required tobe determined according to the Libsvm library 0e SVMalgorithm performs poorly on the experimental data whenthe default parameter values are selected 0erefore the gridsearch method was utilized to determine the optimal pa-rameters for c and c before they were used for training 0epair (c and c) with the highest cross-validation accuracy was

f1

f2

Maxmarg

in

Maxmarg

in Optimal hyperplane

(a)

Not max

margin

Not max

margin

f1

f2

(b)

Figure 5 SVM optimal hyperplane


chosen as the best values In the present study the searchscales for these two values were set as follows c 1 to 10and c (1eminus4 1eminus2 00001) for all cases

In addition the obtained values of c and c for all cases areshown in Table 3 For each method the obtained best-suitedoptimized value was set to the hyperparameters which werelater utilized to train the classifier using the training groupsand the performance of each resulting classifier was thenassessed on the testing set 0ereby unbiased evaluations ofthe performances of eachmethod were obtained whereas forKNN classifier we have used the elbowmethod to decide thek value As we know that in the elbow method if one plotsthe percentage of variance explained by the cluster againstthe number of clusters the first cluster will add much in-formation (explain a lot of variances) but at some point themarginal gain will drop giving an angle in the graph So thenumber of clusters is chosen at this point First we plot theelbow curve using the Matplotlib library Moreover in ourcase we provide a range of 1 to 251 for AD vs HC classi-fication problem (which represents the number of subjectsfor each classification) 0erefore when we plot the graphthe k value from the graph where there is a bend tells us howmany clusters are there

0e performance of a binary and multiclass classifier canbe understood using confusion matrix as shown in Table 40e performance of the system was assessed using the SVMKNN NB and softmax classifiers for each precise test in-cluding binary and tertiary classification tasks 0e diagonalelements of the matrix show the number of correct pre-dictions made by the classifier0ese elements can be furtherseparated into true positive (TP) and true negative (TN)thus indicating the correctly identified controls Similarlythe number of incorrectly classified subjects may be char-acterized by false negative (FN) and false positive (FP)Accuracy measures the number of examples that werecorrectly labeled by the classifier that is

acc TP + TN

TP + TN + FP + FN (7)

However for a dataset with unstable class distributioncalculating only the accuracy may result in a misleadingestimation of the performance 0erefore four additionalperformance metrics should be calculated namely speci-ficity sensitivity precision and F1 score0ey are defined asfollows

sen TP

TP + FN (8)

spe TN

TN + FP (9)

ppv TP

TP + FP (10)

F1 score 2TP

2TP + FP + FN (11)

Sensitivity (8) indicates the accuracy of the predictiongroup and specificity (9) indicates the accuracy of the

prediction of the absence group Sensitivity measures thesuccess rate for a particular class ie within a class thepercentage of correctly determined subjects (by the classi-fication algorithm) to be in the class Specificity provides ameasure for those not in the class ie it is the percentage ofthose not in the class that were found not to be in the classPrecision (10) (which is also termed as positive predictivevalue (PPV)) is the fraction of relevant occurrences amongthe retrieved occurrences and F1 score (11) (which is alsocalled F score or F measure) is a quantity related to a testrsquosaccuracy

To evaluate whether each technique performs signifi-cantly better than a random classifier McNemarrsquos chi-squared test was used and its significance level was keptto 005 which is a benchmark value McNemarrsquos test ex-amines the difference between proportions in paired ob-servations It was used to evaluate the difference betweenthe proportions of accurately classified subjectsie accuracy 0e corresponding contingency chart isshown in Table 5

33 Classification Results Both binary and tertiary clas-sification methods were used to measure the classificationperformance in obtaining cortical and subcorticalfeatures

0e obtained results of the classification tests are shownin Tables 6ndash10 for AD vs HC HC vs mAD mAD vs aAD(binary) AD vs HC vs mAD and AD vs HC vs aAD(tertiary) respectively 0e classification report for eachcase and the performance are shown in Figure 6 Allprograms were executed in 64-bit Python 36 environmenton Intel(R) Core(TM) i3-7100 at 390Hz and 8GB of RAMrunning Ubuntu 1604 LTS 0e model may be imple-mented on any computer in which Python 36 iscompatible

331 Binary Classification AD vs HC HC vs mAD andmAD vs aAD

AD vs HC 0e classification results for AD vs HC aresummarized in Table 6 and Figure 6(a) For each case thedataset was divided into two parts in a 7030 ratio All

Table 3 Best obtained c and c values using the grid search method

Group C value Gamma valueAD vs HC 1 00025HC vs mAD 1 0005600000000000001mAD vs aAD 1 0005AD vs HC vs mAD 6 00096AD vs HC vs aAD 3 00088

Table 4 Confusion matrix

True classPredicted class

G2 G2G1 TP FNG1 FP TN


methods performed significantly better and their P valuewas smaller than the conventional 005 0e assumption wasthat there were significant dissimilarities between the twoparts 0e difference between AD and HC was 6447 with95 confidence interval from 5687 to 7208 and it washighly significant (Plt 0000001) Softmax classified AD vsHC with a high accuracy of 9934 sensitivity of 9814specificity of 100 precision of 100 and F1 score of 9906

HC vs mAD 0e classification report for HC vs mAD issummarized in Table 7 and Figure 6(b) In this case alltechniques achieved significantly better P value than chance(Plt 005) however SVM classified HC vs mAD highlysignificantly (Plt 0000001) with an accuracy of 992 sen-sitivity of 9902 specificity of 100 precision of 100 and F1score of 9951

mAD vs aAD0e classification results for mAD vs aADare summarized in Table 8 and Figure 6(c) Here threemethods performed significantly better than chance andtheir p value was less than the conventional value (Plt 005)however KNN had a P value greater than the conventionalvalue SVM had a P value significantly better than the otherthree methods and classified mAD vs aAD with a highaccuracy of 9777 sensitivity of 100 specificity of 9523precision of 96 and F1 score of 9795

332 Multiclass Classification For multiclass classificationaccuracy may be a misleading indication of the performancethus in this case the F1 score metric is particularly useful

AD vs HC vs mAD 0e classification results for AD vsHC vs mAD are summarized in Table 9 and the plot isshown in Figure 6(d) In this case SVM classified AD vs HCvs mAD with an accuracy of 9942 sensitivity of 9918specificity of 995 precision of 9999 and F1 score of 99430e F1 score indicates a high correct percentage it can beclaimed that SVM classified AD vs HC vs mAD moreaccurately

AD vs HC vs aAD0e classification results for AD vs HCvs aAD are summarized in Table 10 and are shown inFigure 6(d) It can be seen that NB classified the multiclasscase significantly more accurately with an accuracy of 9653sensitivity of 9588 specificity of 9764 precision of 9588and F1 score of 9588

333 Comparing with Othersrsquo State-of-the-Art MethodsAs we know the NRCD dataset is not available publiclythat is why we apply our proposed method on the OASISdataset which can be downloaded from (httpswwwoasis-brainsorg) 0e OASIS database was launched in2007 as a public-private partnership OASIS providesbrain-imaging dataset which is freely available for sharingand data study 0is dataset consists of a cross-sectionalgroup of 416 patients which covers the adult lifespan agedfrom 18 to 96 including individuals with early-phaseAlzheimerrsquos disease (AD) 0e subjects are right-handedand they include both men and women From 416 subjects100 included subjects above the age of 60 have beenidentified with the very mild-to-mild AD and theremaining 316 are diagnosed as normal All structuralsMRI scans used in this experiment were acquired from15T scanners

At first we extracted 110 (cortical and subcortical)features from all 416 subjects using the same Freesurferversion which is used for the NRCD dataset After that wefollowed the same procedure that we had applied for theNRCD dataset As can be seen from Table 11 all methodsperformed significantly better and their P value was smaller

Table 5 Contingency table for the McNemarrsquos test

Group 2 correctlyclassified

Group 2misclassified

Group 1 correctly classified A BGroup 1 misclassified C D

Table 6 Classification results for AD vs HC

AD vs HC ACC SEN SPEC PRE F1score

McNemarrsquostest

Softmaxclassifier 9934 9814 100 100 9906 Plt 0000001

SVM 9802 9787 9805 9583 9684 Plt 0000001KNN 9868 9811 9898 9811 9811 Plt 0000001Naıve Bayes 9840 100 9904 9791 9894 Plt 0000001

Table 7 Classification results for HC vs mAD

HC vs mAD ACC SEN SPEC PRE F1score

McNemarrsquostest

Softmaxclassifier 9603 100 8437 9494 974 P 0000113

SVM 992 9902 100 100 9951 Plt 0000001KNN 9841 100 931 9797 9897 P 0000002Naıve Bayes 9761 100 90 9696 9846 P 0000008

Table 8 Classification results for mAD and aAD

mAD vs aAD ACC SEN SPEC PRE F1score

McNemarrsquostest



Table 9 Classification results for AD vs HC vs mAD

AD vs HC vs mAD ACC SEN SPEC PRE F1 scoreSoftmax classifier 8971 9267 9576 8638 8942SVM 9942 9918 995 9999 9943KNN 9852 9746 987 9923 9833Naıve Bayes 9828 9907 9932 9655 9779

Table 10 Classification results of AD vs HC vs aAD



than the conventional 005 0e assumption was that therewere significant dissimilarities between these two parts Ourproposed method achieved a better result with 9840 ofaccuracy 9375 of sensitivity 100 of specificity 100 ofprecision and 9677 of F1 score with softmax classifier Itcan also be seen that SVM follows the softmax classifier interms of performance

To further demonstrate the usefulness of our proposedmethod we compared it with 9 state-of-the-art approachesas shown in Table 12 0e result in Table 12 shows that Jacket al [11] used the kernel-LICA-DC method to classify theAD vs HC and achieved a classification accuracy of 7425a sensitivity of 96 and a specificity of 525 Jha et al [19]used (deep learning method) sparse autoencoder techniqueto classify the classification problem and achieved a clas-sification accuracy of 916 a sensitivity of 9809 and aspecificity of 8409 In both of these cases they haveachieved decent accuracy but their specificity is very lowcompared to our proposed method Likewise Islam andZhang [20] used an ensemble of deep convolutional neuralnetworks technique and achieved a classification accuracyof 9318 and a specificity of 93 Farhan et al [21] used anensemble of a classifier for the classification of AD vs HCand achieved a classification accuracy of 9375 a

93949596979899

100

Accuracy Sensitivity Specificity Precision F1 score

Softmax classifierSVM

KNNNaiumlve Bayes

(a)

75

80

85

90

95

100



KNNNaiumlve Bayes

(b)

80828486889092949698

100



KNNNaiumlve Bayes

(c)

75

80

85

90

95

100



KNNNaiumlve Bayes

(d)

75

80

85

90

95

100



KNNNaiumlve Bayes

(e)

Figure 6 Classification report of each group with the performance measure of accuracy sensitivity specificity precision and F1 score (a)AD vs HC (b) HC vs mAD (c) mAD vs aAD (d) AD vs NC vs mAD and (e) AD vs NC vs aAD



McNemarrsquostest




sensitivity of 100 and a specificity of 875 Despite usingdeep learning and ensemble classifier methods their ac-curacy sensitivity and specificity are low compared to ourproposed method Khajehnejad et al [18] used the semi-supervised technique for classification and achieved anaccuracy of 9386 a sensitivity of 9465 and a specificityof 9322 Jha et al [16] used the extreme learning machinemethod for classification of AD vs HC with dual-treewavelet principal coefficients and achieved an accuracyof 9527 a sensitivity of 9659 and a specificity of9303 And Lama et al [22] used cortical features forearly detection of AD and RELM for classification andachieved an accuracy of 7788 a sensitivity of 6885 anda specificity of 8354 Despite using new methods forclassification their classification performance is very lowcompared to our results as shown in Table 12 As can beseen from Table 12 our proposed method using softmaxclassifier has outperformed all state-of-the-art methods andachieved an accuracy of 9934 a sensitivity of 9814 anda specificity of 100 for the NRCD dataset and for theOASIS dataset our proposed method achieved an accuracyof 9840 a sensitivity of 9375 and a specificity of 100Additionally to enhance the performance results we havecompared with the performance results of the proposedmethod and the image analysis system [27] as shown inFigure 7

4 Discussion

In this study different techniques for the classification ofsubjects with AD and mAD based on anatomical T1-weighted MRI were compared To assess and comparetheir performance five experiments were conducted AD vsHC HC vs mAD mAD vs aAD (binary classification) ADvs HC vs mAD and AD vs HC vs aAD (multiclass clas-sification) 0e set of subjects was randomly split into twogroups in the ratio of 7030 for training and testing 0enPCA was applied for dimensionality reduction Sub-sequently four classifiers were used to evaluate perfor-mance For the SVM classifier the optimal parameter valuewas determined by using a grid search and SKF CV 0enthose values were used to train the SVM classifier using thetraining dataset and the testing dataset was evaluated on

the training model to evaluate the performance Overallthe radical basic function (RBF) SVM classifier yieldedbetter result compared with the other methods in severalcases Regarding the softmax classifier Adam optimizationwas used with learning rate 1eminus2 1000 epochs and batchsize 50 and it yielded highly satisfactory results for binaryclassification particularly for AD vs HC with an accuracyof 9934

5 Conclusions

A new method for automatic classification of an AD fromMCI (asymptomatic or converted to the AD) and a healthygroup is proposed using combined features extracted froman automated toolbox where sMRI scans of NRCD datasetis used as an input image Here four different types ofclassifiers (softmax classifier SVM KNN and naıve Bayes)were used to classify five different types of the classificationproblem 0ese experimental results confirm that theproposed method had effectively predicted future clinicalchanges between (mAD vs aAD) patients as can be seen inTable 8 Our proposed model has achieved 9799 of ac-curacy 100 of sensitivity 9523 of specificity 96 ofprecision and F1 score as 9795 using the RBF-SVMclassifier for mAD vs aAD classification Efficient and re-liable outcomes were obtained which prove the effec-tiveness of the proposed technique Comparison of theresults showed that the RBF-SVM classifier achieved highlysatisfactory performance for three cases (HC vs mADmAD vs aAD and AD vs HC vs mAD) where the NBclassifier was suitable for predicting AD vs HC vs aAD Aswe can say that our proposed technique to combinecortical thickness and subcortical volume features of thesame subjects for classification of a binary and tertiarygroup performed very well in our dataset as well as in theOASIS dataset

In the present study only cortical thickness and sub-cortical volume features were considered for the classifica-tion process However in the future other different types offeatures will be used such as the hippocampus and amygdalafor the classification of the AD Moreover longitudinaldatasets will be used for studying the changes over timebetween AD mAD and aAD In this study only NRCD

Table 12 Algorithm performance comparison over OASIS and ADNI MRI data

Approach Year Dataset Classifier ModalitiesAD vs HC

ACC SEN SPECCuingnet et al [23] 2011 ADNI SVM MRI NA 81 95Cho et al [25] 2012 ADNI LDA MRI NA 82 93Chyzhyk et al [17] 2012 OASIS Kernel-LICA-DC MRI 7425 96 525Lama et al [22] 2017 ADNI RELM MRI 7788 6885 8354Jha and Kwon [19] 2017 OASIS Sparse autoencoder MRI 916 9809 8409Islam and Zhang [20] 2017 OASIS Ensemble of deep convolutional neural networks MRI 9318 NA 93Farhan et al [21] 2014 OASIS Ensemble of classifier MRI 9375 100 875Khajehnejad et al [18] 2017 OASIS Semisupervised classifier MRI 9386 9465 9322

Jha et al [16] 2018 ADNI ELM MRI 9026 9027 9020OASIS 9527 9659 9303

Proposed method 2018 NRCD Softmax classifier MRI 9934 9814 1002018 OASIS Softmax classifier MRI 9840 9375 100


dataset was used for research so in future we are going tocompare our model with differently available datasets forearly prediction of the AD

Data Availability

0eNational Research Center for Dementia (NRCD) datasetwas used to support the findings of this study At first wewould like to say that NRCD is a private dataset which wasgenerated in Chosun University hospitals and it originallybelongs to Chosun University We cannot share it or open itonline for others due to privacy reasons Later to comparewith other recent state-of-the-art methods we have usedOASIS dataset which was downloaded from httpswwwoasisbrainsorg

Conflicts of Interest

0e authors declare that there are no conflicts of interestregarding the publication of this paper

Authorsrsquo Contributions

Yubraj Gupta and Kun Ho Lee contributed equally to thiswork

Acknowledgments

0is research was supported by the Brain Research Programthrough the National Research Foundation of Korea fundedby the Ministry of Science ICT amp Future Planning

(NRF-2014M3C7A1046041 and NRF-2014M3C7A1046050)Additionally this work was supported by the National Re-search Foundation of Korea grant funded by the Koreangovernment (NRF-2017R1A2B4006533)

References

[1] C P Ferri M Prince C Brayne et al ldquoGlobal prevalence ofdementia a Delphi consensus studyrdquo =e Lancet vol 366no 9503 pp 2112ndash2117 2005

[2] P B Rosenberg and C Lyketsos ldquoMild cognitive impairmentsearching for the prodrome of Alzheimerrsquos diseaserdquo WorldPsychiatry vol 7 no 2 pp 72ndash78 2008

[3] M S Albert ldquoCognitive and neurobiologic markers of earlyAlzheimer diseaserdquo Proceedings of the National Academy ofSciences vol 93 no 24 pp 13547ndash13551 1996

[4] C Marra M Ferraccioli M G Vita D Quaranta andG Gainotti ldquoPatterns of cognitive decline and rates of con-version to dementia in patients with degenerative and vascularforms of MCIrdquo Current Alzheimer Research vol 8 no 1pp 24ndash31 2011

[5] K R Gray R Wolz R A Heckemann P AljabarA Hammers and D Rueckert ldquoMulti-region analysis oflongitudinal FDG-PET for the classification of Alzheimerrsquosdiseaserdquo NeuroImage vol 60 no 1 pp 221ndash229 2012

[6] H Hanyu T Sato K Hirao H Kanetaka T Iwamoto andK Koizumi ldquo0e progression of cognitive deterioration andregional cerebral blood flow patterns in Alzheimerrsquos disease alongitudinal SPECT studyrdquo Journal of the Neurological Sci-ences vol 290 no 1-2 pp 96ndash101 2010

[7] Y J Chen G Deutsch R Satya H G Liu and J M MountzldquoA semi-quantitative method for correlating brain disease

(a) (b)

(c) (d)

Figure 7 Comparison between the proposed results and conventional image analysis tool [27] (a) list of subjects (b) measuring the wholebrain gray matter white matter and lateral ventricle by the standard brain map (c) setting the ROI and (d) volumetric and ICV regardingROIs


groups with normal controls using SPECT Alzheimerrsquosdisease versus vascular dementiardquo Computerized MedicalImaging and Graphics vol 37 no 1 pp 40ndash47 2013

[8] J Barnes J W Bartlett L A van de Pol et al ldquoA meta-analysis of hippocampal atrophy rates in Alzheimerrsquos diseaserdquoNeurobiology of Aging vol 30 no 11 pp 1711ndash1723 2008

[9] J R Petrella ldquoNeuroimaging and the search for a cure forAlzheimer diseaserdquo Radiology vol 269 no 3 pp 671ndash6912013

[10] G B Frisoni N C Fox C R Jack Jr P Scheltens andP M 0ompson ldquo0e clinical use of structural MRI inAlzheimer diseaserdquo Nature Reviews Neurology vol 6 no 2pp 67ndash77 2010

[11] C R Jack D A Bennett K Blennow et al ldquoNIA-AA researchframework toward a biological definition of Alzheimerrsquos dis-easerdquoAlzheimerrsquos ampDementia vol 14 no 4 pp 535ndash562 2018

[12] F Barkhof T M Polvikoski E C W van Straaten et al ldquo0esignificance of medial temporal lobe atrophy a postmortemMRI study in the very oldrdquo Neurology vol 69 no 15pp 1521ndash1527 2007

[13] C Zarow H V Vinters W G Ellis et al ldquoCorrelates ofhippocampal neuron number in Alzheimerrsquos disease and is-chemic vascular dementiardquo Annals of Neurology vol 57no 6 pp 896ndash903 2005

[14] M Bobinski M J de Leon J Wegiel et al ldquo0e histologicalvalidation of post mortem magnetic resonance imaging-determined hippocampal volume in Alzheimerrsquos diseaserdquoNeuroscience vol 95 no 3 pp 721ndash725 2000

[15] E Braak and H Braak ldquoAlzheimerrsquos disease transientlydeveloping dendritic changes in pyramidal cells of sector CA1of the Ammonrsquos hornrdquo Acta Neuropathologica vol 93 no 4pp 323ndash325 1997

[16] D Jha S Alam J Y Pyun K H Lee and G R KwonldquoAlzheimerrsquos disease detection using extreme learning ma-chine complex dual tree wavelet principal coefficients andlinear discriminant analysisrdquo Journal of Medical Imaging andHealth Informatics vol 8 no 5 pp 881ndash890 2018

[17] D Chyzhyk M Grantildea A Savio and J Maiora ldquoHybriddendritic computing with kernel-LICA applied to Alzheimerrsquosdisease detection in MRIrdquo Neurocomputing vol 75 no 1pp 72ndash77 2012

[18] M Khajehnejad F Saatlou and H Mohammadzade ldquoAlz-heimerrsquos disease early diagnosis using manifold-based semi-supervised learningrdquo Brain Sciences vol 7 no 12 p 1092017

[19] D Jha Department of Information and CommunicationEngineering Chosun University and G R Kwon ldquoAlz-heimerrsquos disease detection using sparse autoencoder scaleconjugate gradient and softmax output layer with fine tun-ingrdquo International Journal of Machine Learning and Com-puting vol 7 no 1 pp 13ndash17 2017

[20] J Islam and Y Zhang ldquoAn ensemble of deep convolutionalneural networks for Alzheimerrsquos disease detection and clas-sificationrdquo 2017 httparxivorgabs171201675

[21] S Farhan M A Fahiem and H Tauseef ldquoAn ensemble-of-classifiers based approach for early diagnosis of Alzheimerrsquosdisease classification using structural features of brain im-agesrdquoComputational andMathematical Methods inMedicinevol 2014 Article ID 862307 11 pages 2014

[22] R K Lama J Gwak J S Park and S W Lee ldquoDiagnosis ofAlzheimerrsquos disease based on structural MRI images using aregularized extreme learning machine and PCA featuresrdquoJournal of Healthcare Engineering vol 2017 Article ID5485080 11 pages 2017

[23] R Cuingnet E Gerardin J Tessieras et al ldquoAutomaticclassification of patients with Alzheimerrsquos disease fromstructural MRI a comparison of ten methods using the ADNIdatabaserdquo NeuroImage vol 56 no 2 pp 766ndash781 2011

[24] D Zhang Y Wang L Zhou H Yuan and D Shen ldquoMul-timodal classification of Alzheimerrsquos disease and mild cog-nitive impairmentrdquo NeuroImage vol 55 no 3 pp 856ndash8672011

[25] Y Cho J K Seong Y Jeong and S Y Shin ldquoIndividualsubject classification for Alzheimerrsquos disease based on in-cremental learning using a spatial frequency representation ofcortical thickness datardquo Neuroimage vol 59 no 3pp 2217ndash2230 2012

[26] R Wolz V Julkunen J Koikkalainen et al ldquoMulti-methodanalysis of MRI images in early diagnostics of Alzheimerrsquosdiseaserdquo PLoS One vol 6 no 10 Article ID e25446 2011

[27] NeuroI 2019 httpscdssneuroaicokr[28] P M 0ompson M S Mega R P Woods et al ldquoCortical

change in Alzheimerrsquos disease detected with a disease-specificpopulation-based brain atlasrdquo Cerebral Cortex vol 11 no 1pp 1ndash16 2011

[29] B Fischl D H Salat E Busa et al ldquoWhole brain segmen-tationrdquo Neuron vol 33 no 3 pp 341ndash355 2002

[30] R S Desikan F Segonne B Fischl et al ldquoAn automatedlabeling system for subdividing the human cerebral cortex onMRI scans into gyral based regions of interestrdquo NeuroImagevol 31 no 3 pp 968ndash980 2006

[31] A H Andersen D M Gash and M J Avison ldquoPrincipalcomponent analysis of the dynamic response measured byfMRI a generalized linear systems frameworkrdquo MagneticResonance Imaging vol 17 no 6 pp 795ndash815 1999

[32] V N Vapnik =e Nature of Statistical Learning =eorySpringer New York NY USA 1995

[33] C Y Wee P T Yap and D Shen ldquoPrediction of Alzheimerrsquosdisease and mild cognitive impairment using cortical mor-phological patternsrdquo Human Brain Mapping vol 34 no 12pp 3411ndash3425 2013

[34] A B Tufail A Abidi A M Siddiqui and M S YounisldquoAutomatic classification of initial categories of Alzheimerrsquosdisease from structural MRI phase images a comparison ofPSVM KNN and ANN methodsrdquo Age vol 6 pp 713ndash7172012

[35] H Warner A Toronto L Veasy et al ldquoA mathematicalapproach to medical diagnosisrdquo JAMA vol 177 no 3pp 177ndash183 1961

[36] S R Bhagya Shree and H S Sheshadri ldquoDiagnosis of Alz-heimerrsquos disease using naive bayesian classifierrdquo NeuralComputing and Applications vol 29 no 1 pp 123ndash132 2016

[37] J Murdoch and J A Barnes ldquoLinear regression theoryrdquo inStatistics Problems and Solutions Palgrave MacmillanBasingstoke UK 1973


International Journal of

AerospaceEngineeringHindawiwwwhindawicom Volume 2018

RoboticsJournal of

Hindawiwwwhindawicom Volume 2018


Active and Passive Electronic Components

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawiwwwhindawicom

Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Control Scienceand Engineering

Journal of



Journal ofEngineeringVolume 2018

SensorsJournal of



RotatingMachinery


Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation


Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia

Submit your manuscripts atwwwhindawicom

inability to correctly differentiate patients at the mild cog-nitive impairment (MCI) stage who will progress on todevelop AD from those that have MCI symptoms that aredue to other causes A potentially promising treatment forADmay not be beneficial when administered to those whoseMCI symptoms are not due to AD causing an apparentlypromising treatment to be considered a failure Anotherreason for distinguishing those patients with prodromal ADat the MCI stage is that interventions early in the course ofthe disease may help delay the onset andor mitigate the riskof full-blown AD [2] whereas interventions later in thecourse of the disease may limit the disease but would not beable to reverse the pathology-induced neuronal loss after ithas already occurred Hence diagnosis of presymptomaticAD in the MCI stage is a highly important task which willbecome even more crucial and urgent when curativetreatment becomes available Such diagnosis can also bevaluable for those whose MCI is due to causes other thanAD as these causes may be easier to determine and manageHence neuroimaging improves the positive predictive valueof the diagnosis and involves measurements by structuralMRI to test medial temporal lobe atrophy (MTL) andpositron emission tomography using fluorodeoxyglucose(FDG) or amyloid markers MTL structures are fundamentalfor the creation of new memories and are centrally related tothe growth of AD [3] Specifically MTA atrophy and alliedepisodic memory impairment are trademark features of ADand they both progressively lessen over the course of thedisease [4] Normally MTA is determined by using vertex-based [5] voxel-based [6] and ROI-based [7] approaches

In this study the focus is on binary and tertiary clas-sification between AD HC mAD (MCI due to AD) andaAD (asymptomatic AD) using structural MRI By using theatrophy measure from sMRI scans the intensity and stage ofthe neurodegeneration can be determined 0ese studiesinclude morphometric approaches such as the region ofinterest (ROI) and volume of interest (VOI) of gray mattervoxels for the automatic segmentation of sMRI images andthe sMRI volume measurement of the hippocampus and themedial progressive lobe [8] Various machine-learningmethods have been used to differentiate binary and ter-tiary classification between AD HC mAD and aAD Assuggested in [9 10] only using amyloid imaging biomarkermay be less sensitive in tracing AD progression mainly inthe symptomatic stage recent consensus statements haveunderlined the importance of a biomarker of neuro-degeneration which is a critical component of AD patho-physiology in the prodromal and early dementia stages andalso from [11] atrophy on sMRI reflects on cumulative lossand shrinkage of the neuropil [12ndash14] 0ese indicate that avolumetric measure of cortical thickness and subcorticalvolume is one essential biomarker for early detection of theAD Cortical thickness analysis establishes as another widelyaccepted approach to measure gray matter atrophy in theAD and cortical thinning has been found in MCI and ADDespite evidence for subcortical amyloid and neurofibrillarytangle formation in AD [15] MRI research has drawn itsattention to AD-related subcortical structure changes onlyrecently Advanced segmentation techniques now permit the

quantification of subcortical volumes and provide the basisfor subcortical shape analysis Measurement of structuralchanges based on brain MRI scans has previously been usedto classify AD patients versus cognitively normal (CN)subjects and to predict the risk of progression from MCI tothe AD Jha et al [16] proposed a method for diagnosis of theAD using complex dual-tree wavelet principal coefficientstransform (DTCWT) 0ey have used two datasets OASISand ADNI for the validation of their results An extremelearning machine was used as a classifier and an accuracy of9026 sensitivity of 9027 and specificity of 9027 wereachieved for the 2D ADNI dataset and an accuracy of9575 a sensitivity of 9659 and a specificity of 9303were achieved for the OASIS dataset Chyzhyk et al [17] usesa lattice independent component analysis (LICA) techniquefor the feature section stage with combine kernel trans-formation of the data 0eir approach improved the gen-eralization of the dendritic computing classifiers 0en theyapply their proposed model on the OASIS dataset forclassification of AD patients with normal subjects and theirmodel achieve 7475 accuracy 96 sensitivity and 525specificity Khajehnejad et al [18] uses the manifold-basedsemisupervised learning method for early diagnosis of theAD 0ey have used the OASIS dataset for experiment andtheir proposed method yield accuracy of 9386 sensitivityof 9465 and specificity of 9322 Jha and Kwon [19]proposed an AD detection method using sparse autoen-coder scale conjugate gradient and softmax output layerwith fine-tuning 0ey have used the OASIS dataset andachieved an accuracy of 916 9809 of sensitivity and8409 of specificity Islam and Zhan [20] proposed anensemble of a deep convolutional neural network for earlydetection of AD Here they have used DenseNet-121DenseNet-161 and DenseNet-169 deep learning modelsfor classification of the OASIS dataset and later they haveused their proposed ensemble model which was fused withdeep learning for classification of the same dataset 0eirproposed method yields 9318 of accuracy 94 of pre-cision 93 of sensitivity and 92 of F1 score Farhan et al[21] proposed a model which uses two extracted features ofthe brain first features are the volume of GMWM and CSFand second features are the (l hrh) area of the hippocampusIn addition later they have used four different classifiers(SVM MLP j48 and ensemble of classifiers) to evaluate theclassification accuracy An ensemble of classifiers has got thehigh accuracy compared to other classifiers with 9375accuracy for combined features Lama et al [22] proposed abinary and a tertiary classification between AD MCI andNC using sMRI data from the ADNI dataset A regularizedextreme learning machine (RELM) was used as a classifierand an accuracy of 7730 for binary classification (AD vsNC) and 7661 for tertiary classification (AD vs NC vsMCI) were obtained cortical thickness features were used inthe experiment Based on a large subgroup of ADNI dataCuingnet et al [23] compared ten sMRI-based feature ex-traction techniques and their ability to distinguish betweenclinically relevant subject groups 0ese methods comprisedfive voxel-based techniques three methods based on corticalthickness and two techniques based on the hippocampus

















std xj1113872 1113873 (1)








EducationM F











thickness

(a) (b) (c)


3D input image



Normalization

PCA



Feature selection

Diagnosis output

AD

HCKNN

Naiumlve Bayes

Softmax classifier

SVMaAD

mAD

(e)(a)

(b)

(h) (g)

(d)(c)

(f)

+








dj

1113944

n


11139741113972

(3)







pc

x1113874 1113875

p(xc)p(c)

p(x) (4)



12E + 12

1E + 12

8E + 11

6E + 11

No

of c

ompo

nent

s

4E + 11

2E + 11

0

1 5 9 13 17 21 25 29 33 37 41 45 49 53 57

No of variances

61 65 69 73 77 81 85 89 93 97 101

105

109




pi expyi

1113936cc1expyc

(5)







f1

f2

Maxmarg

in

Maxmarg


(a)

Not max

margin

Not max

margin

f1

f2

(b)






acc TP + TN

TP + TN + FP + FN (7)


sen TP

TP + FN (8)

spe TN

TN + FP (9)

ppv TP

TP + FP (10)

F1 score 2TP

2TP + FP + FN (11)




























McNemarrsquostest





McNemarrsquostest





McNemarrsquostest










93949596979899

100



KNNNaiumlve Bayes

(a)

75

80

85

90

95

100



KNNNaiumlve Bayes

(b)

80828486889092949698

100



KNNNaiumlve Bayes

(c)

75

80

85

90

95

100



KNNNaiumlve Bayes

(d)

75

80

85

90

95

100



KNNNaiumlve Bayes

(e)




McNemarrsquostest





4 Discussion



5 Conclusions










Data Availability






Acknowledgments



References








(a) (b)

(c) (d)





































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia

















std xj1113872 1113873 (1)








EducationM F











thickness

(a) (b) (c)


3D input image



Normalization

PCA



Feature selection

Diagnosis output

AD

HCKNN

Naiumlve Bayes

Softmax classifier

SVMaAD

mAD

(e)(a)

(b)

(h) (g)

(d)(c)

(f)

+








dj

1113944

n


11139741113972

(3)







pc

x1113874 1113875

p(xc)p(c)

p(x) (4)



12E + 12

1E + 12

8E + 11

6E + 11

No

of c

ompo

nent

s

4E + 11

2E + 11

0

1 5 9 13 17 21 25 29 33 37 41 45 49 53 57

No of variances

61 65 69 73 77 81 85 89 93 97 101

105

109




pi expyi

1113936cc1expyc

(5)







f1

f2

Maxmarg

in

Maxmarg


(a)

Not max

margin

Not max

margin

f1

f2

(b)






acc TP + TN

TP + TN + FP + FN (7)


sen TP

TP + FN (8)

spe TN

TN + FP (9)

ppv TP

TP + FP (10)

F1 score 2TP

2TP + FP + FN (11)




























McNemarrsquostest





McNemarrsquostest





McNemarrsquostest










93949596979899

100



KNNNaiumlve Bayes

(a)

75

80

85

90

95

100



KNNNaiumlve Bayes

(b)

80828486889092949698

100



KNNNaiumlve Bayes

(c)

75

80

85

90

95

100



KNNNaiumlve Bayes

(d)

75

80

85

90

95

100



KNNNaiumlve Bayes

(e)




McNemarrsquostest





4 Discussion



5 Conclusions










Data Availability






Acknowledgments



References








(a) (b)

(c) (d)





































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia









thickness

(a) (b) (c)


3D input image



Normalization

PCA



Feature selection

Diagnosis output

AD

HCKNN

Naiumlve Bayes

Softmax classifier

SVMaAD

mAD

(e)(a)

(b)

(h) (g)

(d)(c)

(f)

+








dj

1113944

n


11139741113972

(3)







pc

x1113874 1113875

p(xc)p(c)

p(x) (4)



12E + 12

1E + 12

8E + 11

6E + 11

No

of c

ompo

nent

s

4E + 11

2E + 11

0

1 5 9 13 17 21 25 29 33 37 41 45 49 53 57

No of variances

61 65 69 73 77 81 85 89 93 97 101

105

109




pi expyi

1113936cc1expyc

(5)







f1

f2

Maxmarg

in

Maxmarg


(a)

Not max

margin

Not max

margin

f1

f2

(b)






acc TP + TN

TP + TN + FP + FN (7)


sen TP

TP + FN (8)

spe TN

TN + FP (9)

ppv TP

TP + FP (10)

F1 score 2TP

2TP + FP + FN (11)




























McNemarrsquostest





McNemarrsquostest





McNemarrsquostest










93949596979899

100



KNNNaiumlve Bayes

(a)

75

80

85

90

95

100



KNNNaiumlve Bayes

(b)

80828486889092949698

100



KNNNaiumlve Bayes

(c)

75

80

85

90

95

100



KNNNaiumlve Bayes

(d)

75

80

85

90

95

100



KNNNaiumlve Bayes

(e)




McNemarrsquostest





4 Discussion



5 Conclusions










Data Availability






Acknowledgments



References








(a) (b)

(c) (d)





































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia







dj

1113944

n


11139741113972

(3)







pc

x1113874 1113875

p(xc)p(c)

p(x) (4)



12E + 12

1E + 12

8E + 11

6E + 11

No

of c

ompo

nent

s

4E + 11

2E + 11

0

1 5 9 13 17 21 25 29 33 37 41 45 49 53 57

No of variances

61 65 69 73 77 81 85 89 93 97 101

105

109




pi expyi

1113936cc1expyc

(5)







f1

f2

Maxmarg

in

Maxmarg


(a)

Not max

margin

Not max

margin

f1

f2

(b)






acc TP + TN

TP + TN + FP + FN (7)


sen TP

TP + FN (8)

spe TN

TN + FP (9)

ppv TP

TP + FP (10)

F1 score 2TP

2TP + FP + FN (11)




























McNemarrsquostest





McNemarrsquostest





McNemarrsquostest










93949596979899

100



KNNNaiumlve Bayes

(a)

75

80

85

90

95

100



KNNNaiumlve Bayes

(b)

80828486889092949698

100



KNNNaiumlve Bayes

(c)

75

80

85

90

95

100



KNNNaiumlve Bayes

(d)

75

80

85

90

95

100



KNNNaiumlve Bayes

(e)




McNemarrsquostest





4 Discussion



5 Conclusions










Data Availability






Acknowledgments



References








(a) (b)

(c) (d)





































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia



pi expyi

1113936cc1expyc

(5)







f1

f2

Maxmarg

in

Maxmarg


(a)

Not max

margin

Not max

margin

f1

f2

(b)






acc TP + TN

TP + TN + FP + FN (7)


sen TP

TP + FN (8)

spe TN

TN + FP (9)

ppv TP

TP + FP (10)

F1 score 2TP

2TP + FP + FN (11)




























McNemarrsquostest





McNemarrsquostest





McNemarrsquostest










93949596979899

100



KNNNaiumlve Bayes

(a)

75

80

85

90

95

100



KNNNaiumlve Bayes

(b)

80828486889092949698

100



KNNNaiumlve Bayes

(c)

75

80

85

90

95

100



KNNNaiumlve Bayes

(d)

75

80

85

90

95

100



KNNNaiumlve Bayes

(e)




McNemarrsquostest





4 Discussion



5 Conclusions










Data Availability






Acknowledgments



References








(a) (b)

(c) (d)





































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia





acc TP + TN

TP + TN + FP + FN (7)


sen TP

TP + FN (8)

spe TN

TN + FP (9)

ppv TP

TP + FP (10)

F1 score 2TP

2TP + FP + FN (11)




























McNemarrsquostest





McNemarrsquostest





McNemarrsquostest










93949596979899

100



KNNNaiumlve Bayes

(a)

75

80

85

90

95

100



KNNNaiumlve Bayes

(b)

80828486889092949698

100



KNNNaiumlve Bayes

(c)

75

80

85

90

95

100



KNNNaiumlve Bayes

(d)

75

80

85

90

95

100



KNNNaiumlve Bayes

(e)




McNemarrsquostest





4 Discussion



5 Conclusions










Data Availability






Acknowledgments



References








(a) (b)

(c) (d)





































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia
















McNemarrsquostest





McNemarrsquostest





McNemarrsquostest










93949596979899

100



KNNNaiumlve Bayes

(a)

75

80

85

90

95

100



KNNNaiumlve Bayes

(b)

80828486889092949698

100



KNNNaiumlve Bayes

(c)

75

80

85

90

95

100



KNNNaiumlve Bayes

(d)

75

80

85

90

95

100



KNNNaiumlve Bayes

(e)




McNemarrsquostest





4 Discussion



5 Conclusions










Data Availability






Acknowledgments



References








(a) (b)

(c) (d)





































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




93949596979899

100



KNNNaiumlve Bayes

(a)

75

80

85

90

95

100



KNNNaiumlve Bayes

(b)

80828486889092949698

100



KNNNaiumlve Bayes

(c)

75

80

85

90

95

100



KNNNaiumlve Bayes

(d)

75

80

85

90

95

100



KNNNaiumlve Bayes

(e)




McNemarrsquostest





4 Discussion



5 Conclusions










Data Availability






Acknowledgments



References








(a) (b)

(c) (d)





































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia



4 Discussion



5 Conclusions










Data Availability






Acknowledgments



References








(a) (b)

(c) (d)





































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia



Data Availability






Acknowledgments



References








(a) (b)

(c) (d)





































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia


alzheimer’sdiseasediagnosisbasedoncorticaland...

Documents