Research Article DWI-Based Neural Fingerprinting...

Research ArticleDWI-Based Neural Fingerprinting Technology A PreliminaryStudy on Stroke Analysis

Chenfei Ye1 Heather Ting Ma1 Jun Wu2 Pengfei Yang1 Xuhui Chen2

Zhengyi Yang3 and Jingbo Ma1

1 Department of Electronic and Information Engineering Harbin Institute of Technology Shenzhen Graduate SchoolHIT Campus University Town Room 205C C Building Xili Nanshan Shenzhen 518055 China

2Department of Neurology Peking University Shenzhen Hospital Shenzhen 18036 China3 School of Information Technology and Electrical Engineering The University of Queensland St Lucia QLD 4072 Australia

Correspondence should be addressed to Heather Ting Ma heathertmagmailcom

Received 28 March 2014 Revised 4 June 2014 Accepted 6 June 2014 Published 12 August 2014

Academic Editor Ting Zhao

Copyright copy 2014 Chenfei Ye 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

Stroke is a common neural disorder in neurology clinics Magnetic resonance imaging (MRI) has become an important tool toassess the neural physiological changes under stroke such as diffusion weighted imaging (DWI) and diffusion tensor imaging(DTI) Quantitative analysis of MRI images would help medical doctors to localize the stroke area in the diagnosis in terms ofstructural information and physiological characterization However current quantitative approaches can only provide localizationof the disorder rather than measure physiological variation of subtypes of ischemic stroke In the current study we hypothesizethat each kind of neural disorder would have its unique physiological characteristics which could be reflected by DWI images ondifferent gradients Based on this hypothesis a DWI-based neural fingerprinting technology was proposed to classify subtypesof ischemic stroke The neural fingerprint was constructed by the signal intensity of the region of interest (ROI) on the DWIimages under different gradients The fingerprint derived from the manually drawn ROI could classify the subtypes with accuracy100 However the classification accuracy was worse when using semiautomatic and automatic method in ROI segmentationThepreliminary results showed promising potential of DWI-based neural fingerprinting technology in stroke subtype classificationFurther studies will be carried out for enhancing the fingerprinting accuracy and its application in other clinical practices

1 Introduction

Magnetic resonance imaging (MRI) has been widelyemployed in research as well as in clinical practice Forinstance diffusion weighted imaging (DWI) and diffusiontensor imaging (DTI) technologies provide remarkabledetailed information of nervous system and have becomethe important examinations for neural diseases diagnosisin neurology department of hospital Specifically DTImeasures the water diffusion situation in neural fibre so thatit is frequently used to investigate the abnormal diffusionin the brain Based on DWI principles DTI can providethe contrast of the diffusion anisotropy that was furtherdeveloped to trace the fibre tracts [1] Both DWI and DTItechnologies produce special contrast of nervous systemin terms of diffusion ability fibre integrity fibre bundle

directions and so forth In order to take advantage of theseneuroimaging approaches quantitative analysis is crucial forthe image interpretation which is also important for clinicalapplications Quantitative measures such as mean diffusivity(MD) and fractional anisotropy (FA) were proposed tomeasure the cellular diffusion state and the anisotropy offibre tract in white matter [2] based on DTI An increasingnumber of quantitative methods were introduced to DTIdata analysis such as voxel-based analysis (VBA) [3] andtract-based spatial statistics (TBSS) [4 5]Thesemethods canautomatically localize the lesion in the brain by comparingpatientsrsquo images with a normal control group [6] Howeverthe existing quantitative analysismethods ofDTI are sensitiveto the lesion location but not the physiological changes innature For instance a lesion in the brain can be localized byVBA according to the FA value changes while the inherent

Hindawi Publishing CorporationBioMed Research InternationalVolume 2014 Article ID 725052 9 pageshttpdxdoiorg1011552014725052

2 BioMed Research International

physical meaning for such changes cannot be reflectedby this analysis According to the imaging principlesimages of DTI and DWI can not only provide structuralinformation but also contain physiological meanings [7]Further development on the quantitative analysis willfacilitate interpretation of DWI data that is much helpful inboth neuroscience research and clinical practice

Stroke is a common neural disease especially for theelderly and the people with hypertension [8] In clinicalapplications DWI has shown accurate identification ofischemic tissue and the ability to discriminate betweendead and salvageable ischemic brain [9ndash11] Acute ischemiclesions in DWI can be detected with greater sensitivity thanconventional MRI such as T1 and T2 weighted imaging [12ndash15] Besides defining different stroke states by MRI imagesis important to follow up patientsrsquo response to a therapyIt is reported that decreased apparent diffusion coefficient(ADC) values indicate good sensitivity and specificity in aninfarct less than 10 days old [16] Appearances onDWI imagesfollowing stroke also vary in different states [17] Howeverthe performance of acute infarct detection or stroke statedetermination is unsatisfactory by simply identifying hyper-intensity or hypointensity on DWI images by thresholding[18] Current quantitative measures such as ADC and FA[19] are employed to provide different contrasts of lesionto identify the infarction area and location Better utilityof DWI and DTI data would make it possible to identifysubtypes of stroke which will enhance the diagnosis ofphysiological variation of the patients and thus affect furtherclinicalmanagement A quantitativemeasure which containscomprehensive features of the nerve should be developedto exploit the rich information in DWI images for detectingsubtypes of ischemic stroke

In this study we proposed a method called DWI-basedneural fingerprinting to characterize the neural physiologicalchanges that can be used for subtype classification of ischemicstroke The ldquofingerprintingrdquo concept was borrowed frommagnetic resonance fingerprinting (MRF) [20] techniquepermitting the accelerating acquisition of multiple magneticresonance parameters while in current proposal the ldquofin-gerprintrdquo refers to a feature vector constructed from theDWI images with different diffusion gradients that containscomprehensive neural information As anisotropy measure-ment shows sensitivity to degrees of fibre damage in diseaseaffecting white matter [21ndash23] we hypothesize that specificdiffusivity changewithin ischemic stroke could be consideredas a fingerprint reflecting unique neural property Thereforepathological changes within the infarction of patients afterstroke would be associated with the fingerprint extractedfrom the DWI data with different gradients By applyingclustering algorithms on the fingerprints the subjects canpossibly be classified into normal controls patientswith acutestroke and patients with stroke sequela

2 Methodology

21 Data Acquisition and Image Preprocessing The presentstudy adopted retrospective clinical data from the Neurology

Department of PekingUniversity ShenzhenHospital Clinicaldata from 19 subjects (13 men and 6 women 49 plusmn 19years old) were collected where MRI examinations withthe same protocol were conducted on the subjects The 19subjects were diagnosed as eight healthy people (4 men and4 women 31 plusmn 4 years old) eight with acute stroke lesions(6 men and 2 women 64 plusmn 15 years old) and three withstroke sequela (3 men 58 plusmn 8 years old) For the 11 patientswith acute stroke and stroke sequela 11 10 and 9 lesionswere located in the internal capsule the striatocapsularand the motor cortex respectively Extensive informationabout participantsrsquo health status has been obtained throughsymptomatic evaluation The diagnosis reports were issuedby two neurologists in the Peking University ShenzhenHospital All the participants underwent MR imaging with15 T Siemens Sigma System (Siemens Medical Systems) Thetypical MRI protocol consisted of turbo spin echo (TSE)sequence to generate T2 weighted images (TE = 89ms TR= 4000ms flip angle = 150∘ acquisition matrix = 768 times 624FOV = 230 times 187mm2) and single-shot echo-planar spin-echo (EPSE) sequence to obtainDWI images (TE= 88ms TR= 2700ms flip angle = 90∘ acquisition matrix = 128 times 128FOV = 250 times 250mm2 in-plane resolution 1 times 1mm2 b =1000 smm2 20 diffusion weighted gradient directions and1 without diffusion weighting) Nineteen axial sections in65mm slice gap with 5mm thickness were obtained

The 20 DWI images in Digital Imaging and Communi-cations in Medicine (DICOM) format of each subject wereimported into the SPM8 software (Welcome Trust CentreUCL) for preprocessing [24] involving spatial normalizationto the standard MNI space [25 26] and Gaussian smoothing(FWHM of 3mm) [27]

22 ROI Segmentation and Fingerprint Construction Obtain-ing diffusion weighted signals of the infarct is much depen-dent on the accurate localization of infarct area Three meth-ods were used in the present study named manual semiau-tomatic and automatic ROI segmentation The manual ROIwas segmented by clinicians which was also supposed to bethe reference for the semiautomatic method

First the manual ROI was defined by clinicians Forstroke patients two experienced neurologists blinded to clin-ical symptoms drew target ROI of stroke lesion independentlyon T2 weighted images while taking the T1 and ADC imagesas reference To evaluate the drawing agreement of differentoperators the error of ROIrsquos areas and center coordinatesregarding each participant were evaluated by Bland-Altmanplots and correlation coefficients The interrater reliability ofROI is shown in Figures 1 and 2 All evaluation measuresshow good agreement between the two operators (correlationcoefficients gt 095) to guarantee robust and accurate ROIsegmentation Then the intersection parts of ROIs weremapped to the corresponding DWI images to produce thefinal infarct location through multimodal registration asshown in Figure 3 For normal subjects one arbitrary cerebralhemisphere was selected as target ROI to be investigatedas there was no infarct in their brain The manual ROIwas supposed to provide the most accurate segmentation on

BioMed Research International 3

minus60

minus40

minus20

0

20

40

60

80

100

120

0 200 400 600 800 1000

Erro

r

Average area of ROI

ErrorMeanMean plusmn 2SD

Bland-Altman plot

(a) Bland-Altman plot of ROIrsquos area error between two raters

minus10

minus8

minus6

minus4

minus2

0

2

4

6

8

10

170 190 210 230 250

Erro

r

Average center X of ROI


Bland-Altman plot

(b) Bland-Altman plot of ROIrsquos center X coordinate between tworaters

minus4

minus3

minus2

minus1

0

1

2

3

105 110 115 120 125 130

Erro

r

Average center Y of ROI


Bland-Altman plot

(c) Bland-Altman plot of ROIrsquos center Y coordinate between tworaters

Figure 1 Bland-Altman plot of ROI between two raters

the stroke lesion due to the professional knowledge of theoperators

A semiautomatic method for ROI location was proposedbased on morphological segmentation Firstly DWI imagesof each participant were registered to a normal brain tem-plate Then we averaged 20 slices of registered DWI imagesinto one image for each subject Three stroke subjects werearbitrarily selected from normal subjects patients with acutestroke and patients with stroke sequela respectively TheROIs were drawn accordingly to provide necessary referencesto differentiate normal brain tissue and stroke lesions Imagesignal intensity was compared voxelwise between the ROIand the mirror area on the opposite hemisphere for each

subject The histogram of the signal intensity difference wasmapped to derive the optimal thresholds to differentiatethe lesion from the normal brain tissue Figure 4 shows theprobability density function based on the histogram wheretwo optimized thresholds were determined as the criteria forlesion identification Then the two thresholds were used forlesion ROI segmentation on other patientsrsquo DWI images

In order to segment the lesion ROI automatically we alsoproposed a TBSS based method We hypothesized that thestroke lesion in the brain would contain water diffusibilitychanges that vary the fractional anisotropy of the pixelThenthrough TBSS approach pixels with significant FA changewere detected and used to compose the lesion ROI This


30130230330430530630730830930

1030

40 240 440 640 840 1040

Rat

er 2

Area of ROI

Rater 1

R2= 09714

(a) Correlation test of ROIrsquos area between two raters

160170180190200210220230240250

170 190 210 230 250

Rat

er 2

Center X of ROI

Rater 1

R2= 09555

(b) Correlation test of ROIrsquos center X coordinate between two raters

105

110

115

120

125

130

105 110 115 120 125 130

Rat

er 2

Rater 1

Center Y of ROI

R2= 09575

(c) Correlation test of ROIrsquos center Y coordinate between two raters

Figure 2 Correlation test of ROI between two raters

(a) (b)

Figure 3 Manual drawings of ROIs of stroke lesion on DWI imagesThe stroke lesion was observed in the brain DWI image (a) T2 weightedimage facilitated infarct location on the corresponding DWI images (b) Stroke lesion was localized in DWI image (blue circle) by imageregistration with T2 image while the contralateral region generated automatically in red


minus200 minus150 minus100 minus50 0 50 100 150 2000

01

02

03

04

05

06

Differencing value

Prob

abili

ty

ROI from two acute strokesROI from the normal controlROI from one stroke sequela

reshold 1 reshold 2

Figure 4 The probability density curve of voxelwise differencingvalue Blue curve represents the normal control red curve representsthe acute stroke group green curve represents the stroke sequelagroup Two thresholds are determined to discriminate the differentdistributions

method was implemented using FMRIB Software Library(FSL 419 httpwwwfmriboxacukfsl) [28] First of all abrain template in the software was identified as a commonregistration target We aligned all subjectsrsquo FA images tothis target by nonlinear registration Then a skeletonisedmean FA image was created by a nonmaximum suppressionperpendicular to the local tract structure Each subjectrsquos FAimage (aligned) was projected onto the skeleton by fillingthe skeleton with FA values from the nearest relevant tractcenter Finally based on the voxelwise statistics across thestroke patients and the normal controls the target ROI wasdefined as the voxels with significant difference (uncorrected119875 lt 0005) as illustrated in Figure 5

After segmenting the target ROIs with the above threemethods neural fingerprints can be constructed from themrespectively to reduce intersubject variations in DWI inten-sity the mirror ROI (the contralateral region) correspondingto the target ROI was used as a reference An example ofmirror ROI from the manual segmentation is shown by thered circle generated automatically in Figure 3 The designof diffusion gradients is also critical for the constructionof neural fingerprints Considering that the arrangement ofdiffusion gradients in the three-dimensional space did notaffect classification of neural fingerprints only if applying aconstant order we applied 20 diffusion gradients distributedrandomly in a constant order in every subjectrsquos DWI dataThe neural fingerprint for each ROI segmentation method isdefined as a vector of 20 elements Each element is a ratiobetween the mean image intensities within the target ROIand the mirror ROI calculated with each diffusion gradientAn example of neural fingerprint from manual ROI in DWIimages is shown in Figure 6

23 Clustering To validate the DWI-based neural finger-printing method unsupervised learning was employed tocluster the nineteen subjects Fingerprints used are the

ratio values calculated from manual semiautomatic andautomatic ROIs respectively As the clustering metrics twotypes of distance were employed [29]The Euclidean distancewas calculated as follows

119863119864= radic(xs minus xt) (xs minus xt)

1015840

(1)

where xs and xt are feature vectors of two subjects (the princi-pal components of average image signal intensity sequences)The Cosine distance was calculated as follows

119863119862= 1 minus

xsx1015840tradic(xsx1015840s) (xtx1015840t)

(2)

The K-means algorithm [30 31] was performed for finger-print clustering By setting the number of clusters k eachobservation was assigned by minimizing the least within-cluster sum of squares (WCSS) until the assignments nolonger change WCSS was defined as follows

arg mins

119896

sum

119894=1

sum

119909119895isin119878119894

10038171003817100381710038171003817119909119895minus 120583119894

10038171003817100381710038171003817

2

(3)

where 119909119895belongs to observation (119909

1 1199092 119909

119899) S =

(1198781 1198782 119878

119896) is k clusters and 120583

119894is the mean of points in 119878

119894

We compared 6 sets of clustering results obtained using thecombinations of the three ROI methods and the two distancemetrics (Euclidean and Cosine distance) respectively

To evaluate the clustering results F score a commonmetric to estimate how close the clustering is to the prede-termined benchmark classes was calculated as follows [32]

119865 =

119904

sum

119895=1

10038161003816100381610038161003816119875119895

10038161003816100381610038161003816

sum119904

119894=1

1003816100381610038161003816119875119894

1003816100381610038161003816

sdot max1le119894le119898

2 sdot 119875 (119875119895 119862119894) sdot 119877 (119875

119895 119862119894)

119875 (119875119895 119862119894) + 119877 (119875

119895 119862119894)

(4)

where 119875 is the preclassified sample clusters and 119904 is itscorresponding number of clusters 119862 is the sample clustersand 119898 is its corresponding number of clusters Precision119875(119875119895 119862119894) is the correct results divided by the number of all

returned results and recall 119877(119875119895 119862119894) is the number of correct

results divided by the results that should have been returned[33] The F score can be interpreted as a weighted average ofthe precision and recall where F score reaches its best scoreat 1 and worst score at 0

3 Results

Based on the fingerprint the clustering results are shown inTable 1 The clinical diagnosis in the first column is takenas the standard reference where ldquo1rdquo represents the normalcontrol ldquo2rdquo represents the group with acute stroke lesionsand ldquo3rdquo represents the group with stroke sequela For themanual ROI method it can be observed that clusteringresult approached 100 accuracy with Euclidean distanceIn other words the fingerprint based on the manual ROIwith Euclidean distance gives the best clustering performancecompared to others However the accuracy is relatively poorwhen applying Cosine distance (accuracy = 68 95 CI


(a) (b)

Figure 5 Automatic ROI generation on FA map by TBSS (a) Blue regions indicate significantly decreased FA (119875 lt 0005) in patients withstroke relative to normal controls (b)White regions indicate the corresponding lesion ROIs Green regions represented themean FA skeleton

0

05

1

15

2

25

1 6 11 16 21

Neu

ral fi

nger

prin

t (a

u)

Diffusion gradient number

Normal controlAcute strokeStroke sequela

Figure 6 An example of neural fingerprints from manual ROI inDWI images These neural fingerprints are averaged from manualROI method Blue curve represents the normal control red curverepresents the acute stroke group green curve represents the strokesequela group

48ndash89) For the semiautomatic segmentation methodtwo subjects with acute stroke (Subjects 13 and 15) were falselyincluded into the normal groupwith Cosine distance and thecorresponding accuracy remained at high level (accuracy =89 95 CI 76ndash97) It indicates that the semiautomaticmethod is only sensitive to stroke sequela With Euclideandistance the accuracy for semiautomatic method is 74(95 CI 54ndash93) For automatic TBSS method mismatchoccurs muchmore frequently in all of the three groups as the

1

069

075

089

06

066

05

06

07

08

09

1

ME MC SE SC AE AC

F score

Figure 7 F scores using two distance metrics combined with man-ual semiautomatic and automatic ROIs ME Euclidean distanceusing manual ROI MC Cosine distance using manual ROI SEEuclidean distance using semiautomatic ROI SC Cosine distanceusing semiautomatic ROI AE Euclidean distance using automaticROI AC Cosine distance using automatic ROI

accuracy with Euclidean and Cosine distance is 58 (95CI36ndash80) and 63 (95 CI 41ndash85) respectively

The comparison between F scores provides an overallclustering evaluation for each method as shown in Figure 7F score of the manual ROI method with Euclidean dis-tance shows accurate clustering performance (F = 1) Thedifference of F scores between two distances for manualmethod is relatively larger than that in othermethods For thesemiautomatic segmentation method the F scores are higher(ie 075 and 089) than those in automatic TBSS method(ie 06 and 066) with both distances In other words


Table 1 Clustering result and evaluation

Clinical reference Manual ROI Semiautomatic ROI Automatic ROIEuclideandistance

Cosinedistance

Euclideandistance

Cosinedistance

Euclideandistance

Cosinedistance

Subject 1 1 1 1 1 1 1 1Subject 2 1 1 1 1 1 1 1Subject 3 1 1 1 1 1 1 1Subject 4 1 1 1 1 1 1 3Subject 5 1 1 1 1 1 1 1Subject 6 1 1 1 1 1 1 1Subject 7 1 1 1 1 1 3 1Subject 8 1 1 1 1 1 3 1Subject 9 2 2 1 3 2 1 3Subject 10 2 2 2 3 2 2 3Subject 11 2 2 1 2 2 2 2Subject 12 3 3 3 1 3 3 3Subject 13 2 2 1 2 1 1 3Subject 14 2 2 1 2 2 2 2Subject 15 2 2 2 2 1 2 2Subject 16 3 3 3 1 3 2 2Subject 17 2 2 1 3 2 3 1Subject 18 2 2 1 2 2 3 3Subject 19 3 3 3 1 3 1 3119865 value 1 069 075 089 06 066Clinical reference classification and DWI-based neural fingerprinting clustering results are shown For the clinical reference 1 represents the normal control2 represents the group with acute stroke lesion and 3 represents the group with stroke sequela For manual semiautomatic and automatic ROI methodsdifferent numbers are different clustering labels

the clustering of semiautomatic method provided a betterclassification result than that of automatic TBSS method

4 Discussions

With the development of neuroimaging technology thedetection and analysis of ischemic stroke relying on MRIhave achieved high reliability and availability T1 T2 DWIDTI and ADC maps are commonly used in clinics to detectstroke lesions based on hyperintensity or hypointensity onthe images Medical doctors usually check several typesof MRI images to determine the subtypes of the ischemicstroke lesion The DWI-based neural fingerprinting methodpresented here is a new quantitative approach which takesadvantage of diffusion weighted data to construct a featurevector representing the unique neurophysiological informa-tion of the brain tissue We further applied the fingerprint todetermine the subtypes of ischemic stroke which can be onlybased on the DWI images on different diffusion gradientsThe fingerprint produced in this study can quantitativelymeasure pathological change of neural tissue which reflectsthe specific states of stroke lesions as shown in Figure 6The preliminary results basically validated the hypothesisthat neural physiological change in ischemic stroke can bereflected by the diffusion signal variation on different gradi-ents Further the fingerprint proposed in current study can be

used for subtype determination for ischemic stroke Based onneural fingerprint technology it is possible to further developa tool to assist medical doctors in the diagnosis of strokedisease

For the fingerprint generation lesion ROI segmentationis a key step for the final clustering results It appearedthat the manual ROI method yielded the best clusteringresult among the three segmentation methods As shownin Table 1 a perfect match between clustering results andclinical reference occurs when manual ROI with Euclideandistance was used (F score = 1) It indicates that differentphases of ischemic stroke could be distinguished accuratelywhen lesions have been perfectly localized and distance hasbeen properly defined It also implies that DTI protocol issuitable to generate fingerprints for ischemic stroke Theprecise mechanism leading to diffusion changes of ischemicstroke is still not for certain Wallerian degeneration in thenervous system was found in animal model [34] whichinvolved the breakdown of the myelin sheath and disinte-gration of axonal microfilaments [35] Although disruptionof myelin and axons around acute infarct lesions might beexpected to increase the water diffusivity an accumulationof cellular debris from the breakdown of axons may hinderwater molecule motion [36] which is more likely to occurin late stage of stroke Another explanation could be theredistribution of extracellular water into the intracellular


compartment which leads to the shrinkage of extracellularspace overtime [37] These scientific findings may explainwhy the fingerprint constructed by the DTI images candistinguish the subtypes of the ischemic stroke

The semiautomatic and automatic ROI segmentationmethods were proposed to develop automatic neural finger-print construction method The F scores for both methodsare lower than those obtained in manual ROI with Euclideandistance It is probably due to the inaccuracy in the ROIlocalization of stroke lesions For the semiautomatic methodthe thresholds were determined only by three subjectsrsquo datawhich may not reflect the comprehensive features of thelesion More samples in the training set are necessary toimprove a priori knowledge of the lesion feature As for theautomatic method the inherent hypothesis FA value variedsignificantly for stroke may not be perfect for the ischemicstroke lesion In addition FA images without diffusion tensororientation information could contribute to false location ofsignificantly abnormal areas Another limitation may comefrom the TBSS itself Usually TBSS method is based onthe group comparison not for single brain lesion detectionThe weighting of one specific lesion from a single patient isweakened by the group statistical analysisThe semiautomaticsegmentation method overcomes the constraints of TBSSthus it has a better discrimination as shown in the clusteringresult But the subjects with acute stroke and the normalcontrol fail to be classified perfectly probably because thatthreshold 1 with mixed area under curve is more difficultto determine than threshold 2 (Figure 4) The clusteringresult also implies that appropriate distancemetrics should beused to achieve sufficient discriminative power ofDWI-basedneural fingerprinting The investigation of optimal distancemetric is one of the future works

The current study is the first effort trying to construct afingerprint representing neurophysiological information andwe implemented the technology on ischemic stroke classifi-cation As an ongoing research a system should be finallybuilt up to achieve the true fingerprinting function that isidentification The identity here refers to the neural specificphysiology which can be represented by a unique fingerprintIn the current study we chose to use the DWI images ondifferent diffusion gradients to construct the fingerprintOther MRI protocols are also possible for fingerprint con-struction once the protocol can provide unique pattern ofspecified neural physiology Huge efforts should be carriedout to construct a fingerprint bank which covers massiveamounts of neural fingerprints with matched information(age gender pathology etc) Then the neural fingerprint-ing technology can be finally realized by providing neuralfingerprint identification or verification through comparingan arbitrary neural fingerprint that comes from a patient tothe fingerprint bank Many applications can be rooted fromthe neural fingerprinting technology such as autodiagnosisand risk evaluation for some diseases Despite the specificbrain disease detection the quantification of general neuralproperty distributed in the whole nervous system relies onthe integrated fingerprint bank Building up thewhole systemwould involve the professional knowledge and efforts from

the medical doctors biologists engineers MRI physicistsand so forth

In conclusion this preliminary study demonstrated thatthe proposed DWI-based neural fingerprinting method hadthe potential to classify brain abnormalities such as acutestroke and stroke sequela due to its ability to exploitcomprehensive information contained in DWI data Furtherdevelopment on such technology could assist clinical practicein the future

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

This study is supported by the High-End Talent OverseasReturnees Foundation of Shenzhen (KQC201109020052A)the National Natural Science Foundation of China (8100-0647) and the Basic Research Foundation (OutstandingYoung Investigator Track) of Shenzhen (JC201005260124A)

References

[1] P J Basser J Mattiello and D LeBihan ldquoMR diffusion tensorspectroscopy and imagingrdquo Biophysical Journal vol 66 no 1pp 259ndash267 1994

[2] M SymmsH R Jager K Schmierer andTA Yousry ldquoA reviewof structural magnetic resonance neuroimagingrdquo Journal ofNeurology Neurosurgery amp Psychiatry vol 75 no 9 pp 1235ndash1244 2004

[3] J Ashburner andK J Friston ldquoVoxel-basedmorphometrymdashthemethodsrdquo NeuroImage vol 11 no 6 I pp 805ndash821 2000

[4] S M Smith M Jenkinson H Johansen-Berg et al ldquoTract-based spatial statistics voxelwise analysis of multi-subjectdiffusion datardquo NeuroImage vol 31 no 4 pp 1487ndash1505 2006

[5] M Afzali H Soltanian-Zadeh and K V Elisevich ldquoTractbased spatial statistical analysis and voxel based morphometryof diffusion indices in temporal lobe epilepsyrdquo Computers inBiology and Medicine vol 41 no 12 pp 1082ndash1091 2011

[6] S Mori K Oishi H Jiang et al ldquoStereotaxic white matteratlas based on diffusion tensor imaging in an ICBM templaterdquoNeuroImage vol 40 no 2 pp 570ndash582 2008

[7] P J Basser ldquoInferring microstructural features and the physi-ological state of tissues from diffusion-weighted imagesrdquo NMRin Biomedicine vol 8 no 7-8 pp 333ndash344 1995

[8] W Rosamond K Flegal K Furie et al ldquoHeart disease andstroke statisticsmdash2008 update a report from the AmericanHeart Association Statistics Committee and Stroke StatisticsSubcommitteerdquo Circulation vol 117 no 4 e25 2008

[9] J W Prichard and R I Grossman ldquoNew reasons for early useof MRI in strokerdquo Neurology vol 52 no 9 pp 1733ndash1736 1999

[10] W Hacke and S Warach ldquoDiffusion-weighted MRI as anevolving standard of care in acute strokerdquo Neurology vol 54no 8 pp 1548ndash1549 2000

[11] P C Sundgren Q Dong D Gomez-Hassan S K MukherjiP Maly and R Welsh ldquoDiffusion tensor imaging of the brainreview of clinical applicationsrdquo Neuroradiology vol 46 no 5pp 339ndash350 2004


[12] M G Lansberg G W Albers C Beaulieu and M P MarksldquoComparison of diffusion-weighted MRI and CT in acutestrokerdquo Neurology vol 54 no 8 pp 1557ndash1561 2000

[13] P A Barber D G Darby P M Desmond et al ldquoIdentificationof major ischemic change diffusion-weighted imaging versuscomputed tomographyrdquo Stroke vol 30 no 10 pp 2059ndash20651999

[14] R G Gonzalez PW Schaefer F S Buonanno et al ldquoDiffusion-weighted MR imaging diagnostic accuracy in patients imagedwithin 6 hours of stroke symptomonsetrdquoRadiology vol 210 no1 pp 155ndash162 1999

[15] M Fisher andGWAlbers ldquoApplications of diffusion-perfusionmagnetic resonance imaging in acute ischemic strokerdquo Neurol-ogy vol 52 no 9 pp 1750ndash1756 1999

[16] M G Lansberg V N Thijs M W OrsquoBrien et al ldquoEvolutionof apparent diffusion coefficient diffusion-weighted and T2-weighted signal intensity of acute strokerdquo American Journal ofNeuroradiology vol 22 no 4 pp 637ndash644 2001

[17] P M Desmond A C Lovell A A Rawlinson et al ldquoThevalue of apparent diffusion coefficient maps in early cerebralischemiardquoThe American Journal of Neuroradiology vol 22 no7 pp 1260ndash1267 2001

[18] M E Mullins P W Schaefer A G Sorensen et al ldquoCTand conventional and diffusion-weighted MR imaging in acutestroke study in 691 patients at presentation to the emergencydepartmentrdquo Radiology vol 224 no 2 pp 353ndash360 2002

[19] T E Conturo ldquoDifferences between gray matter and whitematter water diffusion in stroke diffusion-tensor MR imagingin 12 patientsrdquo Brain vol 215 no 1 pp 211ndash220 2000

[20] D Ma V Gulani N Seiberlich et al ldquoMagnetic resonancefingerprintingrdquo Nature vol 495 no 7440 pp 187ndash192 2013

[21] D J Werring D Brassat A G Droogan et al ldquoThe pathogen-esis of lesions and normal-appearing white matter changes inmultiple sclerosis A serial diffusion MRI studyrdquo Brain vol 123no 8 pp 1667ndash1676 2000

[22] D J Werring C A Clark G J Barker A J Thompson andD H Miller ldquoDiffusion tensor imaging of lesions and normal-appearing whitematter inmultiple sclerosisrdquoNeurology vol 52no 8 pp 1626ndash1632 1999

[23] J S Thornton R J Ordidge J Penrice et al ldquoAnisotropicwater diffusion in white and gray matter of the neonatal pigletbrain before and after transient hypoxia-ischaemiardquo MagneticResonance Imaging vol 15 no 4 pp 433ndash440 1997

[24] N Tzourio-Mazoyer B Landeau D Papathanassiou et alldquoAutomated anatomical labeling of activations in SPM using amacroscopic anatomical parcellation of the MNI MRI single-subject brainrdquo NeuroImage vol 15 no 1 pp 273ndash289 2002

[25] T White A T K Kendi S Lehericy et al ldquoDisruption ofhippocampal connectivity in children and adolescents withschizophreniamdasha voxel-based diffusion tensor imaging studyrdquoSchizophrenia Research vol 90 no 1ndash3 pp 302ndash307 2007

[26] R N Giuliani D V Calhoun and D G Pearlson ldquoVoxel-basedmorphometry versus region of interest a comparison of twomethods for analyzing gray matter differences in schizophre-niardquo Schizophrenia Research vol 74 no 2 pp 135ndash147 2005

[27] H Jiang P van Zijl J Kim et al ldquoDtiStudio resource programfor diffusion tensor computation and fiber bundle trackingrdquoComputer Methods and Programs in Biomedicine vol 81 no 2pp 106ndash116 2006

[28] M Anjari L Srinivasan J M Allsop et al ldquoDiffusion tensorimaging with tract-based spatial statistics reveals local white

matter abnormalities in preterm infantsrdquo NeuroImage vol 35no 3 pp 1021ndash1027 2007

[29] T Korenius J Laurikkala and M Juhola ldquoOn principal com-ponent analysis cosine and Euclidean measures in informationretrievalrdquo Information Sciences vol 177 no 22 pp 4893ndash49052007

[30] J Hartigan A and A Wong M ldquoAlgorithm AS 136 a k-meansclustering algorithmrdquo Applied Statistics vol 28 no 1 pp 100ndash108 1979

[31] L H Juang and M N Wu ldquoMRI brain lesion image detectionbased on color-converted K-means clustering segmentationrdquoMeasurement vol 43 no 7 pp 941ndash949 2010

[32] C Goutte and E Gaussier ldquoA probabilistic interpretation ofprecision recall and F-score with implication for evaluationrdquo inAdvances in InformationRetrieval pp 345ndash359 Springer BerlinGermany 2005

[33] D M W Powers ldquoEvaluation from precision recall and F-measure to ROC informedness markedness and correlationrdquoJournal of Machine Learning Technologies vol 2 no 1 pp 37ndash63 2011

[34] D J Werring A T Toosy C A Clark et al ldquoDiffusiontensor imaging can detect and quantify corticospinal tractdegeneration after strokerdquo Journal of Neurology Neurosurgeryamp Psychiatry vol 69 no 2 pp 269ndash272 2000

[35] C Beaulieu M D Does R E Snyder and P S Allen ldquoChangesin water diffusion due to Wallerian degeneration in peripheralnerverdquoMagnetic Resonance in Medicine vol 36 no 4 pp 627ndash631 1996

[36] S Love D Louis andW Ellison DGreenfieldrsquos Neuropathology2-Volume Set CRC Press 2008

[37] K Liu F Li T Tatlisumak et al ldquoRegional variations in theapparent diffusion coefficient and the intracellular distributionof water in rat brain during acute focal ischemiardquo Stroke vol 32no 8 pp 1897ndash1905 2001

Submit your manuscripts athttpwwwhindawicom

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


MEDIATORSINFLAMMATION

of


Behavioural Neurology

EndocrinologyInternational Journal of



Disease Markers


BioMed Research International

OncologyJournal of



Oxidative Medicine and Cellular Longevity


PPAR Research

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

ObesityJournal of



Computational and Mathematical Methods in Medicine

OphthalmologyJournal of


Diabetes ResearchJournal of



Research and TreatmentAIDS


Gastroenterology Research and Practice


Parkinsonrsquos Disease

Evidence-Based Complementary and Alternative Medicine

Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom


physical meaning for such changes cannot be reflectedby this analysis According to the imaging principlesimages of DTI and DWI can not only provide structuralinformation but also contain physiological meanings [7]Further development on the quantitative analysis willfacilitate interpretation of DWI data that is much helpful inboth neuroscience research and clinical practice

Stroke is a common neural disease especially for theelderly and the people with hypertension [8] In clinicalapplications DWI has shown accurate identification ofischemic tissue and the ability to discriminate betweendead and salvageable ischemic brain [9ndash11] Acute ischemiclesions in DWI can be detected with greater sensitivity thanconventional MRI such as T1 and T2 weighted imaging [12ndash15] Besides defining different stroke states by MRI imagesis important to follow up patientsrsquo response to a therapyIt is reported that decreased apparent diffusion coefficient(ADC) values indicate good sensitivity and specificity in aninfarct less than 10 days old [16] Appearances onDWI imagesfollowing stroke also vary in different states [17] Howeverthe performance of acute infarct detection or stroke statedetermination is unsatisfactory by simply identifying hyper-intensity or hypointensity on DWI images by thresholding[18] Current quantitative measures such as ADC and FA[19] are employed to provide different contrasts of lesionto identify the infarction area and location Better utilityof DWI and DTI data would make it possible to identifysubtypes of stroke which will enhance the diagnosis ofphysiological variation of the patients and thus affect furtherclinicalmanagement A quantitativemeasure which containscomprehensive features of the nerve should be developedto exploit the rich information in DWI images for detectingsubtypes of ischemic stroke

In this study we proposed a method called DWI-basedneural fingerprinting to characterize the neural physiologicalchanges that can be used for subtype classification of ischemicstroke The ldquofingerprintingrdquo concept was borrowed frommagnetic resonance fingerprinting (MRF) [20] techniquepermitting the accelerating acquisition of multiple magneticresonance parameters while in current proposal the ldquofin-gerprintrdquo refers to a feature vector constructed from theDWI images with different diffusion gradients that containscomprehensive neural information As anisotropy measure-ment shows sensitivity to degrees of fibre damage in diseaseaffecting white matter [21ndash23] we hypothesize that specificdiffusivity changewithin ischemic stroke could be consideredas a fingerprint reflecting unique neural property Thereforepathological changes within the infarction of patients afterstroke would be associated with the fingerprint extractedfrom the DWI data with different gradients By applyingclustering algorithms on the fingerprints the subjects canpossibly be classified into normal controls patientswith acutestroke and patients with stroke sequela

2 Methodology

21 Data Acquisition and Image Preprocessing The presentstudy adopted retrospective clinical data from the Neurology

Department of PekingUniversity ShenzhenHospital Clinicaldata from 19 subjects (13 men and 6 women 49 plusmn 19years old) were collected where MRI examinations withthe same protocol were conducted on the subjects The 19subjects were diagnosed as eight healthy people (4 men and4 women 31 plusmn 4 years old) eight with acute stroke lesions(6 men and 2 women 64 plusmn 15 years old) and three withstroke sequela (3 men 58 plusmn 8 years old) For the 11 patientswith acute stroke and stroke sequela 11 10 and 9 lesionswere located in the internal capsule the striatocapsularand the motor cortex respectively Extensive informationabout participantsrsquo health status has been obtained throughsymptomatic evaluation The diagnosis reports were issuedby two neurologists in the Peking University ShenzhenHospital All the participants underwent MR imaging with15 T Siemens Sigma System (Siemens Medical Systems) Thetypical MRI protocol consisted of turbo spin echo (TSE)sequence to generate T2 weighted images (TE = 89ms TR= 4000ms flip angle = 150∘ acquisition matrix = 768 times 624FOV = 230 times 187mm2) and single-shot echo-planar spin-echo (EPSE) sequence to obtainDWI images (TE= 88ms TR= 2700ms flip angle = 90∘ acquisition matrix = 128 times 128FOV = 250 times 250mm2 in-plane resolution 1 times 1mm2 b =1000 smm2 20 diffusion weighted gradient directions and1 without diffusion weighting) Nineteen axial sections in65mm slice gap with 5mm thickness were obtained

The 20 DWI images in Digital Imaging and Communi-cations in Medicine (DICOM) format of each subject wereimported into the SPM8 software (Welcome Trust CentreUCL) for preprocessing [24] involving spatial normalizationto the standard MNI space [25 26] and Gaussian smoothing(FWHM of 3mm) [27]

22 ROI Segmentation and Fingerprint Construction Obtain-ing diffusion weighted signals of the infarct is much depen-dent on the accurate localization of infarct area Three meth-ods were used in the present study named manual semiau-tomatic and automatic ROI segmentation The manual ROIwas segmented by clinicians which was also supposed to bethe reference for the semiautomatic method

First the manual ROI was defined by clinicians Forstroke patients two experienced neurologists blinded to clin-ical symptoms drew target ROI of stroke lesion independentlyon T2 weighted images while taking the T1 and ADC imagesas reference To evaluate the drawing agreement of differentoperators the error of ROIrsquos areas and center coordinatesregarding each participant were evaluated by Bland-Altmanplots and correlation coefficients The interrater reliability ofROI is shown in Figures 1 and 2 All evaluation measuresshow good agreement between the two operators (correlationcoefficients gt 095) to guarantee robust and accurate ROIsegmentation Then the intersection parts of ROIs weremapped to the corresponding DWI images to produce thefinal infarct location through multimodal registration asshown in Figure 3 For normal subjects one arbitrary cerebralhemisphere was selected as target ROI to be investigatedas there was no infarct in their brain The manual ROIwas supposed to provide the most accurate segmentation on


minus60

minus40

minus20

0

20

40

60

80

100

120

0 200 400 600 800 1000

Erro

r

Average area of ROI


Bland-Altman plot


minus10

minus8

minus6

minus4

minus2

0

2

4

6

8

10

170 190 210 230 250

Erro

r



Bland-Altman plot


minus4

minus3

minus2

minus1

0

1

2

3

105 110 115 120 125 130

Erro

r



Bland-Altman plot








30130230330430530630730830930

1030

40 240 440 640 840 1040

Rat

er 2

Area of ROI

Rater 1

R2= 09714


160170180190200210220230240250

170 190 210 230 250

Rat

er 2

Center X of ROI

Rater 1

R2= 09555


105

110

115

120

125

130

105 110 115 120 125 130

Rat

er 2

Rater 1

Center Y of ROI

R2= 09575



(a) (b)




01

02

03

04

05

06

Differencing value

Prob

abili

ty


reshold 1 reshold 2







1015840

(1)


119863119862= 1 minus


(2)


arg mins

119896

sum

119894=1

sum

119909119895isin119878119894

10038171003817100381710038171003817119909119895minus 120583119894

10038171003817100381710038171003817

2

(3)


1 1199092 119909

119899) S =

(1198781 1198782 119878



119894



119865 =

119904

sum

119895=1

10038161003816100381610038161003816119875119895

10038161003816100381610038161003816

sum119904

119894=1

1003816100381610038161003816119875119894

1003816100381610038161003816


2 sdot 119875 (119875119895 119862119894) sdot 119877 (119875

119895 119862119894)

119875 (119875119895 119862119894) + 119877 (119875

119895 119862119894)

(4)




3 Results



(a) (b)


0

05

1

15

2

25

1 6 11 16 21

Neu

ral fi

nger

prin

t (a

u)





1

069

075

089

06

066

05

06

07

08

09

1

ME MC SE SC AE AC

F score







Cosinedistance

Euclideandistance

Cosinedistance

Euclideandistance

Cosinedistance



4 Discussions












Acknowledgments


References













































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















minus60

minus40

minus20

0

20

40

60

80

100

120

0 200 400 600 800 1000

Erro

r

Average area of ROI


Bland-Altman plot


minus10

minus8

minus6

minus4

minus2

0

2

4

6

8

10

170 190 210 230 250

Erro

r



Bland-Altman plot


minus4

minus3

minus2

minus1

0

1

2

3

105 110 115 120 125 130

Erro

r



Bland-Altman plot








30130230330430530630730830930

1030

40 240 440 640 840 1040

Rat

er 2

Area of ROI

Rater 1

R2= 09714


160170180190200210220230240250

170 190 210 230 250

Rat

er 2

Center X of ROI

Rater 1

R2= 09555


105

110

115

120

125

130

105 110 115 120 125 130

Rat

er 2

Rater 1

Center Y of ROI

R2= 09575



(a) (b)




01

02

03

04

05

06

Differencing value

Prob

abili

ty


reshold 1 reshold 2







1015840

(1)


119863119862= 1 minus


(2)


arg mins

119896

sum

119894=1

sum

119909119895isin119878119894

10038171003817100381710038171003817119909119895minus 120583119894

10038171003817100381710038171003817

2

(3)


1 1199092 119909

119899) S =

(1198781 1198782 119878



119894



119865 =

119904

sum

119895=1

10038161003816100381610038161003816119875119895

10038161003816100381610038161003816

sum119904

119894=1

1003816100381610038161003816119875119894

1003816100381610038161003816


2 sdot 119875 (119875119895 119862119894) sdot 119877 (119875

119895 119862119894)

119875 (119875119895 119862119894) + 119877 (119875

119895 119862119894)

(4)




3 Results



(a) (b)


0

05

1

15

2

25

1 6 11 16 21

Neu

ral fi

nger

prin

t (a

u)





1

069

075

089

06

066

05

06

07

08

09

1

ME MC SE SC AE AC

F score







Cosinedistance

Euclideandistance

Cosinedistance

Euclideandistance

Cosinedistance



4 Discussions












Acknowledgments


References













































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















30130230330430530630730830930

1030

40 240 440 640 840 1040

Rat

er 2

Area of ROI

Rater 1

R2= 09714


160170180190200210220230240250

170 190 210 230 250

Rat

er 2

Center X of ROI

Rater 1

R2= 09555


105

110

115

120

125

130

105 110 115 120 125 130

Rat

er 2

Rater 1

Center Y of ROI

R2= 09575



(a) (b)




01

02

03

04

05

06

Differencing value

Prob

abili

ty


reshold 1 reshold 2







1015840

(1)


119863119862= 1 minus


(2)


arg mins

119896

sum

119894=1

sum

119909119895isin119878119894

10038171003817100381710038171003817119909119895minus 120583119894

10038171003817100381710038171003817

2

(3)


1 1199092 119909

119899) S =

(1198781 1198782 119878



119894



119865 =

119904

sum

119895=1

10038161003816100381610038161003816119875119895

10038161003816100381610038161003816

sum119904

119894=1

1003816100381610038161003816119875119894

1003816100381610038161003816


2 sdot 119875 (119875119895 119862119894) sdot 119877 (119875

119895 119862119894)

119875 (119875119895 119862119894) + 119877 (119875

119895 119862119894)

(4)




3 Results



(a) (b)


0

05

1

15

2

25

1 6 11 16 21

Neu

ral fi

nger

prin

t (a

u)





1

069

075

089

06

066

05

06

07

08

09

1

ME MC SE SC AE AC

F score







Cosinedistance

Euclideandistance

Cosinedistance

Euclideandistance

Cosinedistance



4 Discussions












Acknowledgments


References













































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


















01

02

03

04

05

06

Differencing value

Prob

abili

ty


reshold 1 reshold 2







1015840

(1)


119863119862= 1 minus


(2)


arg mins

119896

sum

119894=1

sum

119909119895isin119878119894

10038171003817100381710038171003817119909119895minus 120583119894

10038171003817100381710038171003817

2

(3)


1 1199092 119909

119899) S =

(1198781 1198782 119878



119894



119865 =

119904

sum

119895=1

10038161003816100381610038161003816119875119895

10038161003816100381610038161003816

sum119904

119894=1

1003816100381610038161003816119875119894

1003816100381610038161003816


2 sdot 119875 (119875119895 119862119894) sdot 119877 (119875

119895 119862119894)

119875 (119875119895 119862119894) + 119877 (119875

119895 119862119894)

(4)




3 Results



(a) (b)


0

05

1

15

2

25

1 6 11 16 21

Neu

ral fi

nger

prin

t (a

u)





1

069

075

089

06

066

05

06

07

08

09

1

ME MC SE SC AE AC

F score







Cosinedistance

Euclideandistance

Cosinedistance

Euclideandistance

Cosinedistance



4 Discussions












Acknowledgments


References













































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)


0

05

1

15

2

25

1 6 11 16 21

Neu

ral fi

nger

prin

t (a

u)





1

069

075

089

06

066

05

06

07

08

09

1

ME MC SE SC AE AC

F score







Cosinedistance

Euclideandistance

Cosinedistance

Euclideandistance

Cosinedistance



4 Discussions












Acknowledgments


References













































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of



















Cosinedistance

Euclideandistance

Cosinedistance

Euclideandistance

Cosinedistance



4 Discussions












Acknowledgments


References













































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
























Acknowledgments


References













































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
















Research Article DWI-Based Neural Fingerprinting...

Documents

Transcript of Research Article DWI-Based Neural Fingerprinting...