a computer-aided analysis method of spect brain images for...

Research ArticleA Computer-Aided Analysis Method of SPECT BrainImages for Quantitative Treatment Monitoring PerformanceEvaluations and Clinical Applications

Xiujuan Zheng12 Wentao Wei1 Qiu Huang3 Shaoli Song2

Jieqing Wan4 and Gang Huang2

1Department of Automation The School of Electrical Engineering Information Sichuan University Chengdu China2Department of Nuclear Medicine Renji Hospital School of Medicine Shanghai Jiaotong University Shanghai China3Med-X Institute Shanghai Jiaotong University Shanghai China4Department of Neurosurgery Renji Hospital School of Medicine Shanghai Jiaotong University Shanghai China

Correspondence should be addressed to Xiujuan Zheng xiujuanzhengscueducn

Received 30 September 2016 Revised 4 December 2016 Accepted 28 December 2016 Published 31 January 2017

Academic Editor Zexuan Ji

Copyright copy 2017 Xiujuan Zheng et alThis 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

The objective and quantitative analysis of longitudinal single photon emission computed tomography (SPECT) images aresignificant for the treatment monitoring of brain disorders Therefore a computer aided analysis (CAA) method is introducedto extract a change-rate map (CRM) as a parametric image for quantifying the changes of regional cerebral blood flow (rCBF)in longitudinal SPECT brain images The performances of the CAA-CRM approach in treatment monitoring are evaluated bythe computer simulations and clinical applications The results of computer simulations show that the derived CRMs have highsimilarities with their ground truths when the lesion size is larger than system spatial resolution and the change rate is higher than20 In clinical applications the CAA-CRM approach is used to assess the treatment of 50 patients with brain ischemiaThe resultsdemonstrate that CAA-CRM approach has a 934 accuracy of recovered regionrsquos localization Moreover the quantitative indexesof recovered regions derived from CRM are all significantly different among the groups and highly correlated with the experiencedclinical diagnosis In conclusion the proposed CAA-CRM approach provides a convenient solution to generate a parametric imageand derive the quantitative indexes from the longitudinal SPECT brain images for treatment monitoring

1 Introduction

Single photon emission computed tomography (SPECT)with99mTc-ethyl cysteine dimer (99mTc-ECD) has been widelyused to evaluate many types of cerebrovascular diseases andbrain disorders by measuring regional cerebral blood flow(rCBF) [1ndash6] Moreover longitudinal 99mTc-ECD SPECTbrain imaging can be adopted to monitor the changes ofrCBF to support the treatment plan [7 8]Themost commonapproach for interpreting SPECT brain images is visualinspection in daily clinical practice Usually the relevantstructural images such as CT or MRI images are preferredfor the visual interpretation together with SPECT images Fortreatment monitoring the baseline and follow-up SPECTbrain images should be parallelly interpreted under the same

condition to figure out the differences Due to the lack of thequantitative standards the accuracy and reliability of visualinspection mainly rely on the experience of the physicianssuch that only the qualitative results are presented in thereports The aforementioned drawbacks of vissual insspec-tion prevent the applications of SPECT imaging in distantdiagnosis and multicenter studies The image processingtechnology can bring solutions to the visual inspection prob-lems and explore the hidden information in the images

Beside the region of interest (ROI)volume of inter-est (VOI) analysis methods statistical parametric mapping(SPM) has been used in group-wise comparisons of func-tional brain images to evaluate the responses to the treat-ments [9 10] On the other hand subtraction analysis isalso a useful technology to extract differences in a series

HindawiBioMed Research InternationalVolume 2017 Article ID 1962181 11 pageshttpsdoiorg10115520171962181

2 BioMed Research International

Baseline ampfollow-up

images

Imagescoregistration

Valuenormalization

Parametricimaging

Change-ratemap

Figure 1 The workflow of the computer-aided analysis method to extract a change-rate map

of images for individual treatment monitoring In ictal-interictal SPECT imaging the subtraction analysis has beenproved to benefit the treatment plan of epilepsy patients [11ndash14] Recent studies have reported that the quantitative SPECTanalysis would be playing an ever-growing role in treatmentplan and response monitoring of several disorders relatedwith the central nervous system [15 16]These studies encour-age us to generate the parametric imaging and extract quan-titative indexes from the SPECT images to support the treat-ment plan In this study we introduce a computer-aided anal-ysis (CAA) method inherited from subtraction analysis toquantify the changes of rCBF in longitudinal SPECT imagesfor individual treatment monitoringThe performance of theproposed method would be objectively and systematicallyevaluated by the computer simulations and the clinical appli-cations

2 Materials and Methods

21 Computer-Aided Analysis Method When using SPECTimaging in treatment monitoring the pre- and postscansare usually performed to acquire the baseline and follow-upSPECT images before and after delivering treatments Theobtained baseline and follow-up images from an individualsubject are a set of longitudinal SPECT images requiringindependent analysis For interpreting the individual subjectrsquosdata a computer-aided analysis (CAA)method is establishedto process the longitudinal SPECT images via three mainsteps coregistration value normalization and parametricimaging The workflow chart is shown in Figure 1

In the first step the follow-up SPECT brain images arealigned with the baseline images by the rigid registrationalgorithm provided by SPM 80 software package (httpwwwfilionuclacukspm) If the SPECT brain imaging isperformed using SPECT-CT integrated system the CT andSPECT images can be obtained in the same position andconsidered as well aligned In this condition the baselineand follow-up CT images could be used as the reference andsource image respectively in the step of coregistration foraligning longitudinal SPECT images

In the second step the value normalization is appliedon the SPECT images Before the numerical calculation theextracerebral voxels are removed by a predefinedwhole-brainmask If only the SPECT image is available then the whole-brain mask can be obtained by segmenting the enhancedSPECT image with Otsursquos algorithm If the correspondingaligned CT image is available then the whole-brainmask canbe more accurately defined by separating the brain tissuesfromnonbrain tissue inCT images using fuzzy119862-means clus-tering algorithm [17] The whole-brain mask extracted basedon the CT images is then applied on the SPECT images todelineate the brain area

After deriving the whole-brain area in the SPECT imagesthe value of each cerebral voxel is normalized by the averagevoxel value of the reference area that is automatically selectedby 119885-map approach [18] In 119885-map approach the 119885 value ofthe 119894th voxel was calculated as in (1) Two119885-maps are respec-tively estimated for baseline and follow-up SPECT brainimages Then the reference region is the intersection of the119885 lt 1 areas of these two 119885-maps

119885119894 =1003816100381610038161003816119862119894 minusmean1003816100381610038161003816

SD (1)

where 119862119894 is the value of the 119894th voxel in one SPECT brainimage mean and SD respectively denote the average andstandard deviation of voxel values of brain area in the SPECTimage

In the third step the changes in the longitudinal SPECTimages are expressed in a parametric image to reflect thedisease progress or responses to the treatment As in sub-traction analysis the difference can be directly obtained bysubtracting two aligned normalized images as

119863119894 = 119862119891119894 minus 119862119887119894 (2)

where 119862119891119894 and 119862119887119894 denote the normalized values of the 119894th

voxel in the follow-up and baseline SPECT images respec-tively

Next the change-rate map can be calculated voxel-by-voxel to reflect the extent of the changes between the baselineand follow-up images The value of the 119894th voxel in theestimated change-rate map is denoted by 119877119894 which can bederived from

119877119894 =119863119894119862119887119894 (3)

The change-rate map (CRM) is a parametric image whichcan be fused with aligned SPECTCT images for visualinspections The positive voxel value in CRM demonstratesthe recovery of hypoperfusion while the negative value indi-cates the recovery of hyperperfusion For the visualizationGaussian smoothing filter can be applied in the CRM toreduce the impact of noise Based on the CRM the changedregions are automatically obtained by thresholding and clus-tering Firstly the voxels with lower change-rate (lt20) wereset to 0 in CRM Then the 119870-means clustering algorithm isadopted to recognize the changed regions The morphologi-cal processing is applied to refine and distinguish each regionConsidering the SPECT image resolution and partial volumeeffects the regions with larger volumes (gt120 voxels) areselected as the recovered regions

For the localization of recovered regions an atlas of brainlobes which consists of 12 brain anatomical structures (listed

BioMed Research International 3

Table 1 The list of brain anatomical structures in the atlas of brainlobes

Brain anatomical structures Values in atlasCerebellum anterior lobe 80Cerebellum posterior lobe 10Frontal lobe Left 50 right 55Frontal-temporal space Left 110 right 115Limbic lobe Left 40 right 45Medulla 20Midbrain 100Occipital lobe Left 90 right 95Parietal lobe Left 120 right 125Pons 70Sublobar 60Temporal lobe Left 30 right 35

in Table 1) is created from the Talairach Daemon atlas [1920] and then translated into MNI (Montreal NeurologicalInstitute) space [21] with dimensions of 91 times 109 times 91 sampledat 2mm intervals corresponding to the SPM templates [22]The SPECTCT images as well as the obtained CRM aremapped to SPM template by the nonrigid registration algorithmprovided by SPM 80 software package and then aligned withthe atlas of brain lobes Based on the atlas of brain lobesthe recovered regions could be located in the different brainlobesThequantitative indexes for the recovered regions suchas themean andmaximum change-rate and the proportion ofthe recovered regions to the corresponding brain lobes couldbe calculated for each detected recovered regions

In order to facilitate the expression the proposed approachused to evaluate longitudinal SPECT images through a CRMderived by the CAAmethod is noted as CAA-CRM approachin the subsequent parts

22 Computer Simulations

221 Simulated Data In this study the performance of theCAA-CRM approach in treatment monitoring was objec-tively and systematically evaluated by the computer simula-tions The longitudinal SPECT images are simulated usingpredesigned digital brain phantomsThe normal digital brainphantom is a 100 times 100 times 82 matrix representation of thehardwareHoffmanphantom [23] whose voxel size is 213mmtimes 213mm times 213mm In the digital brain phantom the valueof each voxel presents the radioactivity in the correspondingposition Generally this digital brain phantom is used tosimulate the normal brain perfusion images acquired by99mTc-ECD SPECT imaging Comparing with other regulargeometrical objects the sphere ismore suitable for simulatingthe ischemic lesions in perfusion images For convenience asphere is created in the normal digital brain phantom locatedin the right frontal lobe as a lesion analogueThediameter andradioactivity of the sphere can be changed for several scalesto simulate the varied degrees of hypoperfusion for brainischemia The diameter of the sphere was designed in three

scales 8mm 16mm and 24mm In addition the radioactiv-ity in the sphere was set based on the predefined change-ratescaled in 9 different levels uniformly distributed from 10 to90 The simulated brain SPECT images are generated withan injected dose of 25mCi of 99mTc-ECD The abnormal andnormal brain images are respectively regarded as the base-line and follow-up images obtained in treatment monitoringfor brain ischemia

The system parameters used in computer simulations areset according to the geometry of the dual-head Philips Prece-dence 6 SPECTCT scanner Each detector head is mountedwith a low-energy and high-resolution (LEHR) collimatorThe two heads rotate in H-mode to obtain 128 projections intotal over 360∘ around the predesigned phantom Projectiondata are acquired for 10 minutes with a total count numberof 108 accompanied with measurement noise that is modeledas an additive Poisson noiseThe radioactivity distribution inthe brain phantom is reconstructed with a maximum a pos-teriori (MAP) algorithm with total variation regularization

222 Performance Evaluation In the performance evalua-tion the ground truth of CRM is directly defined based onthe phantoms for every simulated lesion size at each change-rate The estimated CRM is compared to the correspondingground truth for evaluating its quality In this study theindexes of image quality are adopted to objectively and sys-tematically quantify the performance Denote the value of the119894th voxel in the estimated CRM as 119877119894 and denote the value ofcorresponding voxel in ground truth as119866119894Thus the normal-ized absolute error (NAE) which is the simplest metric formeasuring the difference between two images can be calcu-lated by

NAE =sum119873119894=11003816100381610038161003816119877119894 minus 1198661198941003816100381610038161003816

sum119873119894=1 119866119894 (4)

As shown in (5) the peak signal to noise ratio (PSNR)is the index to reflect the image quality of obtained CRMcomparing with its ground truth

PSNR = 20 log10119866maxradicMSE

MSE = 1119873

119873

sum119894=1

(119877119894 minus 119866119894)2

(5)

where 119866max is the maximum voxel value of ground truth ofCRM MSE is for mean square error

The normalized cross-correlation (NCC) is used to quan-tify the similarity between the obtained CRM and its groundtruth The NCC can be calculated by

NCC = 1119873 minus 1

119873

sum119894=1

(119877119894 minus 119877) (119866119894 minus 119866)120590119888120590119866

(6)

where 119877 and 119866 represent the mean values of the obtainedCRM and the corresponding ground truth respectively 120590119888and 120590119866 denote their standard deviations


After the recovered region is detected based on CRMthe Dice similarity coefficient (DSC) which is calculated asin (7) is used to measure the accuracy of the recoveredregions detection comparing with the ground truth that is thepredefined sphere in the phantom

DSC = 2 times TPTP + FN + TP + FP

(7)

where TP is true positive that is the set of voxels commonto the derived recovered region and ground truth TN is truenegative that is the set of voxels not labelled as the derivedrecovered region and ground truth FN is false negative andFP is false positive

Moreover the change-rates derived from the recoveredregions are also used to evaluate the accuracy of recoveredregion detection Because of the homogeneity of voxel valuesin the predefined digital phantom only themean change-rateof the recovered region is calculated and compared with thepredefined real value by linear regressions

23 Clinical Applications

231 Clinical Data Acquisition In this study 99mTc-ECDSPECT brain imaging is used in the treatment monitoring ofthe internal carotid artery (ICA) stenting which is a commontreatment technique for brain ischemia This study has beenapproved by the Ethics Committee of Renji Hospital Schoolof Medicine Shanghai Jiaotong University All the SPECTscans are performed in accordance with the guidelines forbrain perfusion SPECT using 99mTc-labelled radiopharma-ceuticals [1] 50 patients in total (7 women 43 men andaverage age 629 plusmn 105 years) prescribed ICA stenting arerecruited 27 of them have cerebral infarction while the restsuffer different degrees of cerebral ischemia For each patientthe baseline 99mTc-ECD SPECT imaging is performed within7 days before surgery and then the follow-up scan is generallyconducted in sim7 days (ranged from 2 to 12 days) after thetreatment of ICA stenting The SPECT imaging was startedwithin 20sim30 minutes after the radiotracer injection (around25mCi 99mTc-ECD) using dual-head Philips Precedence 6SPECTCT scanner with low-energy and high-resolutioncollimators For 41 patients the CT scans are performedtogether with SPECT imaging For the rest of 9 patients thecorresponding CT images are not available The system reso-lution is 74mm full width half maximum (FWHM) at 10 cmfor SPECT imaging Three-dimensional SPECT images werereconstructed using Astonish technology which adopts aniterative ordered-subset expectation-maximization (OSEM)algorithm with built-in scatter correction and attenuationcorrection [24] For each patient a pair of baseline andfollow-up SPECT images was used to evaluate the therapy ofICA stenting

232 Clinical Data Analysis For the data analysis thetraditional visual inspection and theCAA-CRMapproach areboth used to assess the recovery levels For the visual inspec-tions the baseline and follow-up SPECT brain images werecompared by two independent experienced physicians in the

same image workstation When CT images are available thephysicians inspected the SPECT images with the support ofthe aligned CT images The hypoperfusion lesions caused bycerebral infarction or ischemia are carefully studied and thenthe recovered regions are delineated manually Furthermorethe overall recovery level for each patient is formally reportedin four scales (none mild recovery moderate recovery andsevere recovery) based on the physiciansrsquo experience On theother hand these longitudinal SPECT images are quantita-tively analyzed by the CAA-CRM approach After all auto-matic processes a CRM is derived for each patient Then theestimated CRM is fused with the original SPECTCT imagesThe recovered regions are automatically derived based on theCRM by thresholding and clustering The threshold is setas 20 according to the results of computer simulations (asmentioned in Section 3) Small regions (less than 120 voxels)are excluded to eliminate the influence ofmeasurement noiseThen the automatically recovered regions are located by theatlas of brain lobes and the quantitative indexes related torecovery level are estimated In the further evaluation of theperformance of CAA-CRM approach in treatment monitor-ing the clinical diagnosis derived by visual inspection isconsidered as the standard of the classification of recoverygroups to validate the results of the CAA-CRM approach

In the performance evaluation the automatically recov-ered regions are firstly compared with manually definedones If an automatic recovered region hits the correspondingmanual one it would be considered as a successful detectionof the real recovered region Moreover McNemarrsquos test [25]for paired automatic and manual recovered regions is usedto investigate the concordance of the CAA-CRM approachand traditional visual inspection in the detection of recoveredregionsMeanwhile quantitative indexes including themeanchange-rate maximum change-rate and proportion of therecovered regions to the corresponding lobes are estimatedfrom the CRM The statistics of these quantitative indexesrelated to the recovery levels are analyzed in three differentrecovery groups classified according to clinical reports Non-parametric one-way ANOVA is also applied in the analysis ofvariance of different recovery groups

3 Results

31 Results of Computer Simulation Using CAA-CRMapproach the CRMs are derived based on the simulated dataThese estimated CRMs are compared with the correspondingground truths The properties of CRM in treatment moni-toring are quantified by image quality indexes including theNAE PSNR and NCCThe comparative results of the imagequality indexes for different lesion sizes at each change-rateare shown in Figure 2 In Figure 2(a) NAE declines withthe increase of change-rate for each lesion size Moreover forsmall-size lesion (1206018mm) the calculatedNAE ismuchhigherthan larger-size lesions (12060116mm and 12060124mm) especiallyin the condition of low change-rates (10 and 20) InFigure 2(b) PSNR progressively rises along the increase ofchange-rate and it climbs quicker in the case of small-sizelesion Figure 2(c) shows that NCC goes up with the increaseof change-rate when the lesion size is comparatively large


00 10 20 30 40 50 60 70 80 90 100

50

100

150

200

250

300

350

Change-rate ()

NA

E

empty24

empty16

empty8

(a)

0 10 20 30 40 50 60 70 80 90 100Change-rate ()

5

10

15

20

25

30

PSN

R

empty24

empty16

empty8

(b)

0 10 20 30 40 50 60 70 80 90 100Change-rate ()

00

02

04

06

08

10

NCC

empty24

empty16

empty8

(c)

Figure 2 Comparisons of the image quality indexes for three-size lesions (1206018mm diameter of 8mm 12060116mm diameter of 16mm and12060124mm diameter of 24mm) at nine different change-rates (10 to 90) (a)Thenormalized average error (NAE) formeasuring the differencebetween the change-rate map (CRM) and its ground truth (b) the peak signal-noise ratio (PSNR) for reflecting the image quality of the CRM(c) the normalized cross-correlation (NCC) for quantifying the similarity between the CRM and its ground truth

(12060116mm and 12060124mm) However NCC increases slightly forsmall-size lesion

DSC andmean change-rate estimates of recovered regionare both used to evaluate the performance of the derivedCRM in recovered region detection From the curve chartin Figure 3(a) the DSC of large lesion (12060124mm) maintainsa higher value (gt07) For the medium lesion (12060116mm) theDSC increases quickly with rising of change-rate when thechange-rate is less than 40 Meanwhile the DSC changesslightly and stays in a high level when the recovery level ishigher than 40 However the DSC fluctuates with a lowvalue along the increase of recovery level for the small lesion(1206018mm) In Figure 3(b) the linear regressions are plotted for

three different lesion sizes The highly linear relations (1199032 =099119901 lt 00001) between the change-rates estimates and realvalues are observed in the condition of larger lesions (12060116mmand 12060124mm) It is also found that the estimated meanchange-rates are underestimated since the liner regressionlines of the estimates and real values are under the line 119910 = 119909(the black solid line)The underestimation ismore serious forthe small lesions It seems that the CAA-CRM approach failsin deriving the acceptable change-rate for small-size lesionwith diameter of 8mm

Based on the results of compuater simulations it canbe concluded that the CAA-CRM approach has a betterperformance in the detection of recovered regions and


DSC

0 10 20 30 40 50 60 70 80 90 100Change-rate ()

empty24

empty16

empty8

00

02

04

06

08

10

(a)

0102030405060708090

100

Standard CR ()

Estim

ated

CR

()

0 10 20 30 40 50 60 70 80 90 100

empty24 y = 096x minus 236 r2 = 099

empty16 y = 066x + 371 r2 = 099

empty8 y = 001x + 118 r2 = 001

(b)

Figure 3 Comparisons of detected recovered regions with the ground truth for three-size lesions (1206018mm diameter of 8mm 12060116mmdiameter of 16mm and 12060124mm diameter of 24mm) at nine different change-rates (10 to 90) (a) The Dice similarity coefficient (DSC)for measuring the accuracy of the recovered regions detection comparing with the ground truth (b) linear regressions of estimated change-rates (CR) of the recovered regions with the real values

the quantification of change-rates when the lesion size issufficiently large (larger than a sphere with diameter of 8mmthat is closed to the proposed spatial resolution of SPECTimages) and the change-rate is high enough (at least not lowerthan 20)

32 Results of Clinical Application For the visual inspection17 of 50 patients are reported as severe recovery in brainperfusion after ICA stenting while 22 patients are scaled asmoderate recovery In the remaining 11 patients there are 8patients with mild recovery and 3 patients are diagnosed asno improvement along with the treatment Considering thepopulation distributions of these 4 groups themild and nonerecovery patients are combined as one group to compare withthe other groups In the visual inspection 106 manual VOIsare totally delineated to locate the recovered regions for these50 patients

Beside the visual inspection the CAA-CRM approach isapplied to monitor the changes based on the longitudinalSPECT images for each patient Figure 4 illuminates a typicalcase of severe recovery In Figure 4(a) there are three trans-verse slices (the 36th 40th and 44th slices) of baseline 99mTc-ECD SPECT images The lesion of cerebral infarction can beclearly found in the left parietal lobe which is pointed by awhite arrow The cerebral ischemia is easily detected for thehypoperfusion regions around the lesionThe correspondingslices of aligned follow-up SPECT images are shown inFigure 4(b) In the clinical report based on visual inspectionthe lesion of cerebral infarction had no improvement after thetreatment of ICA stenting However the cerebral blood flowrecovers significantly in the hypoperfusion regions aroundthe lesion of cerebral infarction The overall recovery levelgiven by physicians is severe recovery In Figure 4(c) the

Table 2 The concordance of automatic VOIs with manual ones inthe localization of recovered regions

FN TP FPSever 3 41 5Moderate 3 42 3Mildnone 1 16 4Total 7 99 12FN false negative the number ofmanually recovered regions whichwere nothit by automatic ones TP true positive the number of manually recoveredregions which were hit by automatic ones FP false positive the numberof automatically recovered regions which could not find the paired manualones

estimated CRM presents as a parametric image fused withthe corresponding baseline SPECT image In this case thederived CRM could be used to enhance the visual inspectionfor physicians in treatment monitoring In Figure 4(d) theCRM as well as the corresponding SPECT images is mappedto the standard atlas of brain lobes for the convenientlocalization of recovered regions

The CAA-CRM approach also has the advantages inautomatic and quantitative analysis of recovered regions Forall the patientsrsquo SPECT images the recovered regions areautomatically derived located and compared tomanual oneswhich are used as the standard for validation The detailedresults are listed in Table 2 to reflect the concordance betweenthe automatic detection and manual detection of recoveredregions From Table 2 there are in total 99 concordat pairs(automatically recovered regions that hit manual ones) andthe total concordance rate of localization is 934 Forthe rest of 19 discordant pairs there are 12 (1219 632)


36 40 44

(a)

36 40 44

(b)

36 40 44

Change-rate ()10 15 20 25 30 35 40 45

(c)

Change-rate ()10 15 20 25 30 35 40 45

36 40 44

(d)

Figure 4 A typical case of treatment monitoring using CAA-CRM approach where the estimated change-rate map can directly reflectresponse to treatmentThe patient (male 69-year-old) suffered cerebral infarction in left parietal lobe While the baseline 99mTc-ECD SPECTscan was performed 3 days before the treatment of ICA stenting the follow-up scan was obtained 7 days after the treatment (a) Threetransverse slices (the 36th 40th and 44th slices) of baseline SPECT images The lesion of cerebral infarction is pointed out by white arrowAround this lesion the hypoperfusion could be observed (b) Three corresponding transvers slices of the follow-up SPECT image There isno improvement in the lesion of cerebral infarction while the severe recovery is observed for the cerebral ischemia in hypoperfusion regionsaround the lesion The recovery level was scaled as severe by the traditional visual inspection in clinical report (c) Three transverse slicesof the derived change-rate map (CRM) fused with baseline SPECT image The scales of change-rate are presented by rainbow color bar Thewarmer color denotes higher change-rate In this case the recovered regions can be detected easily and clearly in the change-rate map (d)The selected transverse slices of the CRM fused with the atlas of brain lobes The contours of the brain lobes are delineated The recoveredregions can be conveniently localized in the brain area

pairs where the automatic method recommended recoveredregions while the physicians did not and 7 (719 368)pairs are in the contrary condition By the conventionalcriteria of McNemarrsquos test (119901 lt 005) this difference isconsidered to be not statistically significant There is alsono significant difference between the automatic and manualapproaches in localization of recovered regions whatevergroup (severe moderate and mildnone groups) is chosenThe results indicate that the CAA-CRM approach and visualinspections of experienced physicians have the concordancein the localization of recovered regions

After localizing the recovered regions the mean andmaximum change-rates for each patient could be calculatedbased on the delineated recovered regions and then comparedin groups The group-wise results of statistical comparisonsare illuminated by bar graphs in Figure 5 Figure 5(a)illuminates that the mean change-rates of the severe recoverygroup are mainly higher than those for the moderate groupand mildnone group The similar results can be observedin Figure 5(b) for the maximum change-rate estimates forthree groupsThe statistical results of nonparametric one-wayANOVA indicate that the significant differences (119901 lt 00001)exist among the meanmaximum change-rates of three dif-ferent recovery groups Beside the indexes of change-rate

the proportion of recovered regions to the correspondingbrain lobes which is calculated as a volume ratio betweenrecovered regions and the corresponding brain lobes is also aspecific index for quantifying the recovery level for treatmentmonitoring The comparison of proportion of recoveredregions for three recovery groups is shown in Figure 5(c)The tendency of the proportions in groups is similar to thatof the change-rate It is obvious that the better recoverygroups have the higher proportions According to the resultsof nonparametric one-way ANOVA three recovery groupshad significant effect (119901 lt 0001) on proportion of recoveredregions to the brain lobes

To sum up the results of clinical applications the higherchange-rates and larger recovered regions could correspondto better recovery levels given in the clinical reports Thesequantitative indexes derived by the CAA-CRM approach canbe used to quantify the response to the treatment

4 Discussion

In this study the performance of the CAA-CRM approach isobjectively and systematically evaluated by the computer sim-ulations as well as the clinical applications From the results


S M MN20

22

24

26

28

30

Mea

n CR

()

(a)S M MN

20

30

40

50

60

70

80

Max

CR

()

(b)S M MN

0

5

10

15

20

25

Prop

ortio

n of

RRs

()

(c)

Figure 5 The bar graphs (mean + SD) for comparisons of quantitative indexes for three recovery groups (S severe M moderate and MNmildnone) (a) The mean change-rate (CR) estimates of the recovered regions (b) the maximum CR estimates of the recovered regions (c)the proportion of recovered regions (RRs) to the corresponding brain lobes

of computer simulations the lesion size and change-rate areconsidered as two major factors for extracting reliable CRMAccording to the changing tendency of image quality indexesof CRM it can be concluded that the lesions with larger sizeand higher change-rate could be easier to detect moreoverthe estimates of change-rate could be more accurate as thereal values This conclusion confirmed our experiences fromclinical practice Additionally the results also clarified thatthe spatial resolution of SPECT image could be a majorlimitation of the CAA-CRM approach to get accurate quanti-tative indexes for quantifying the recovery levels It has beenreported that the quantitative accuracy of the radioactiveconcentration in SPECT image would deteriorate with thedecreasing of target size especially when the targets are belowthree times of spatial resolution [26] For the CAA-CRMapproach the poor quantitative accuracy of original SPECTimages would directly result in estimated bias of change-rateand even in the missing recovered regions As the resultsshown in Figure 2(c) for the small lesion (1206018mm)whose sizeis closer to spatial resolution (74mm FWHM at 10 cm) thevalues of NCC aremuch lower than those of the larger lesions(12060116mm and 12060124mm)The uptrend with increasing change-rate is also very slightThe low value of NCC reflects the poorsimilarity between the obtained CRM and its ground truthIt means that the obtained CRM could not accurately reflectthe small recovered regions when the region size is closed toor even smaller than the spatial resolution of SPECT images

In the computer simulations the underestimation ofchange-rate is observed for all lesions with varied sizesHowever the underestimation is more significant for thesmall lesionThebias probably comes from the partial volumeeffects (PVEs) that are usually related to the spatial resolution

The PVEs could induce the underestimation for quantitativeSPECT images especially when the target is smaller thanthree times of spatial resolution [26] This impact coulddirectly propagate into the CRM that is derived based onSPECT images As illuminated in Figure 3(b) the estimatedchange-rate is much lower than the predefined change-ratefor the small-size lesion (1206018mm)The estimated change-ratesfor middle-size lesion (12060116mm) have the high linear relationwith the predefined values although the values are onlynearly 70 of the real values It is concluded that the change-rate could be significantly underestimated comparing withthe real value when the sizes of recovered regions are belowthree times of spatial resolutionTherefore this underestima-tion should be kept in mind for clinical applications

For the clinical applications the CAA-CRM approachcould be used to define the recovered regions and derivedquantitative indexes to measure recovery levels In this studythe thresholding and clustering method is applied to auto-matically derive the recovered regions The chosen thresholdof change-rate could be themajor impact factor in delineatingthe recovered regions Furthermore it could influence theestimations of quantitative indexes from recovered regions[27 28] The experimental threshold is set as 20 in the clin-ical applications It is chosen based on the results of the com-puter simulations As shown in Figure 2 the image qualityindexes for the change-rates of 10 and 20 are both muchpoorer than the other change-rates regardless of the lesionsize The CRM would not accurately reflect the change-ratelower than 20This indicates that the significant distortionsmay exist in the CRM for the voxels with lower change-ratesIn this case the lowest limit for available estimated rangeof change-rate is required to eliminate the turbulences from


S_20 M_20 M_105

10

15

20

25

30

35

Mea

n CR

()

(a)

S_20 M_20 M_100

5

10

15

20

25

Prop

ortio

n of

RRs

()

(b)

Figure 6 The bar graphs (mean + SD) for comparisons of quantitative indexes for severe and moderate recovery groups with differentthresholds (S 20 severe recovery group with the threshold of 20 M 20 moderate recovery group with the threshold of 20 and M 10moderate recovery group with the threshold of 10) (a) The mean change-rate (CR) estimates of the recovered regions (b) the proportionof recovered regions (RRs) to the corresponding lobes

lower change-rate estimates Meanwhile the threshold set-ting should try to retain as much information as possible inthe CRM for further quantitative analysis Hence the experi-mental threshold set in this study is chosen as 20 rather than10 The thresholds could also be able to reset for differentconditions of varied clinical applications

In the clinical applications considering the concordanceof the CAA-CRM approach in the detection of recoveryregions the false positive (1299) and false negative (799)could be found The false positive might be caused byusing the uniform 20 threshold which might result in thedetection of not only themain recovered regions (validated byvisual inspections) but also the otherminor recovered regionsHowever the minor recovered regions might be ignored bythe physicians On the other hand the uniform thresholdcould easily introduce the small changed regions (lt120voxels) for a certain case The small changed regions couldbe removed or even miss the main recovered regions Thismay lead to the false negative In this condition the thresholdshould be adjusted carefully to balance the false positive andfalse negative for detecting the recovered regions Moreoverthe chosen threshold could further impact on the estimationsof the mean change-rate and the proportion of recoveredregions to the brain lobes As shown in Figure 6 when athreshold of 10 instead of 20 is applied in the moderategroup the values of mean change-rate are decreased sharplyMeanwhile the values of proportion of recovered regions tothe brain lobes have increased From the comparisons themean change-rate is negatively related to the proportion ofrecovered region Therefore these two quantitative indexesshould be used together to scale the recovery level in

treatment monitoring To an extent the maximum change-rate could not be affected by the chosen threshold in theCAA-CRM approach so that it becomes an important quantitativeindex in treatment monitoring

Because all the algorithms used in the proposed CAA-CRM approach totally rely on image contents this approachcould be extended to analyze other types of SPECT brainimages such as 99mTc-HMPAO SPECT brain images Theobtained change-rate map could also illuminate the globalchanges for reflecting the response to treatment The derivedquantitative indexes have the potential to quantify the recov-ery levels

5 Conclusion

In this study a CAA-CRM approach has been introducedto evaluate the longitudinal 99mTc-ECD SPECT images intreatment monitoring This approach can provide change-rate map as a parametric image to reflect the changes of rCBFComputer simulations show the efficacy of the proposedapproach in detecting the recovered regions and in quantify-ing the change-rates for the lesions larger than spatial resolu-tion In clinical applications this method is used to assess thetreatment of ICA stenting The results demonstrate that thequantitative indexes derived from CRM are all significantlydifferent among the groups and highly correlated with theexperienced clinical diagnosis

In conclusion the CAA-CRM approach has the advan-tages of directly illuminating the global recovery and conve-niently quantifying the recovery levels It could be helpful in


improving the efficiency and accuracy of therapy evaluationsusing SPECT brain images in clinical routines

Competing Interests

The authors declare no conflict of interests

Acknowledgments

This work was supported by the National Natural ScienceFoundation of China (no 81201146 to Xiujuan Zheng no51627807 to Qiu Huang and no 81471708 to Shaoli Song)The authors would like to thank the clinicians and the tech-nicians of Renji Hospital for their help with data collectionand the experimental proceduresThe authors would also liketo thank the reviewers of this paper for their constructivesuggestions

References

[1] O Kapucu F Nobili A Varrone et al ldquoEANM procedureguideline for brain perfusion SPECT using 99mTc-labelledradiopharmaceuticals version 2rdquo European Journal of NuclearMedicine and Molecular Imaging vol 36 no 12 pp 2093ndash21022093

[2] Y-A Chung O J Hyun J-Y Kim K-J Kim and K-J AhnldquoHypoperfusion and ischemia in cerebral amyloid angiopathydocumented by 99119898Tc-ECD brain perfusion SPECTrdquo Journal ofNuclear Medicine vol 50 no 12 pp 1969ndash1974 2009

[3] M Farhadi S Mahmoudian F Saddadi et al ldquoFunctional brainabnormalities localized in 55 chronic tinnitus patients fusionof SPECT coincidence imaging and MRIrdquo Journal of CerebralBlood Flow and Metabolism vol 30 no 4 pp 864ndash870 2010

[4] G Uruma K Hashimoto and M Abo ldquoA new method forevaluation ofmild traumatic brain injury with neuropsycholog-ical impairment using statistical imaging analysis for Tc-ECDSPECTrdquo Annals of Nuclear Medicine vol 27 no 3 pp 187ndash2022013

[5] S T Donta R B Noto and J A Vento ldquoSPECT brain imagingin chronic lyme diseaserdquo Clinical Nuclear Medicine vol 37 no9 pp e219ndashe222 2012

[6] L S Zuckier and O O Sogbein ldquoBrain perfusion studies inthe evaluation of acute neurologic abnormalitiesrdquo Seminars inNuclear Medicine vol 43 no 2 pp 129ndash138 2013

[7] J M Mountz H-G Liu and G Deutsch ldquoNeuroimaging incerebrovascular disorders measurement of cerebral physiologyafter stroke and assessment of stroke recoveryrdquo Seminars inNuclear Medicine vol 33 no 1 pp 56ndash76 2003

[8] D G Amen D Highum R Licata et al ldquoSpecific ways brainspect imaging enhances clinical psychiatric practicerdquo Journal ofPsychoactive Drugs vol 44 no 2 pp 96ndash106 2012

[9] S Sestini A Pupi F Ammannati et al ldquoPredictive potential ofpre-operative functional neuroimaging in patients treated withsubthalamic stimulationrdquo European Journal of NuclearMedicineand Molecular Imaging vol 37 no 1 pp 12ndash22 2010

[10] K Chen J B S Langbaum A S Fleisher et al ldquoTwelve-month metabolic declines in probable Alzheimerrsquos disease andamnestic mild cognitive impairment assessed using an empiri-cally pre-defined statistical region-of-interest findings from the

Alzheimerrsquos Disease Neuroimaging InitiativerdquoNeuroImage vol51 no 2 pp 654ndash664 2010

[11] N J Kazemi G A Worrell S M Stead et al ldquoIctal SPECT sta-tistical parametric mapping in temporal lobe epilepsy surgeryrdquoNeurology vol 74 no 1 pp 70ndash76 2010

[12] L Wichert-Ana P M De Azevedo-Marques L F Oliveira etal ldquoIctal technetium-99methyl cysteinate dimer single-photonemission tomographic findings in epileptic patients withpolymicrogyria syndromes a Subtraction of ictal-interictalSPECT coregistered toMRI studyrdquo European Journal of NuclearMedicine and Molecular Imaging vol 35 no 6 pp 1159ndash11702008

[13] H W Lee S B Hong and W S Tae ldquoOpposite ictal perfusionpatterns of subtracted SPECT hyperperfusion and hypoperfu-sionrdquo Brain vol 123 no 10 pp 2150ndash2159 2000

[14] G I VargheseM J Purcaro J EMotelow et al ldquoClinical use ofictal SPECT in secondarily generalized tonicndashclonic seizuresrdquoBrain vol 132 no 8 pp 2102ndash2113 2009

[15] D H Lewis J P Bluestone M Savina W H Zoller E BMeshberg and S Minoshima ldquoImaging cerebral activity inrecovery from chronic traumatic brain injury a preliminaryreportrdquo Journal of Neuroimaging vol 16 no 3 pp 272ndash2772006

[16] Y-W Chuang C-C Hsu Y-F Huang et al ldquoBrain perfusionSPECT in patients with Behcetrsquos diseaserdquo Journal of Neuroradi-ology vol 40 no 4 pp 288ndash293 2013

[17] S Miyamoto H Ichihashi and K Honda Algorithms for FuzzyClustering Methods in C-Means Clustering with Applicationsvol 229 of Studies in Fuzziness and Soft Computing SpringerBerlin Germany 2008

[18] N Boussion CHouzard KOstrowsky P Ryvlin FMauguiereand L Cinotti ldquoAutomated detection of local normalizationareas for ictal-interictal subtraction brain SPECTrdquo Journal ofNuclear Medicine vol 43 no 11 pp 1419ndash1425 2002

[19] J L Lancaster J L Summerlin L Rainey C S Freitas and PT Fox ldquoThe Talairach Daemon a database server for talairachatlas labelsrdquo NeuroImage vol 5 no 4 article S633 1997

[20] J L Lancaster M GWoldorff L M Parsons et al ldquoAutomatedTalairach Atlas labels for functional brain mappingrdquo HumanBrain Mapping vol 10 no 3 pp 120ndash131 2000

[21] A C Evans D L Collins and B Milner ldquoAn MRI-basedstereotactic atlas from 250 young normal subjectsrdquo Society forNeuroscience Abstracts vol 18 no 179 p 408 1992

[22] J A Maldjian P J Laurienti R A Kraft and J H Burdette ldquoAnautomated method for neuroanatomic and cytoarchitectonicatlas-based interrogation of fMRI data setsrdquo NeuroImage vol19 no 3 pp 1233ndash1239 2003

[23] E J Hoffman P D Cutler T M Guerrero W M Digbyand J C Mazziotta ldquoAssessment of accuracy of PET utiliz-ing a 3-D phantom to simulate the activity distribution of[18F]fluorodeoxyglucose uptake in the human brainrdquo Journalof Cerebral Blood Flow amp Metabolism vol 11 supplement 1 ppA17ndashA25 1991

[24] J Ye X Song Z Zhao A J Da Silva J S Wiener and L ShaoldquoIterative SPECT reconstruction using matched filtering forimproved image qualityrdquo in Proceedings of the Nuclear ScienceSymposium Conference Record (NSSMIC rsquo10) pp 2285ndash2287IEEE Knoxville Tenn USA November 2006

[25] Q McNemar ldquoNote on the sampling error of the differencebetween correlated proportions or percentagesrdquo Psychometrikavol 12 no 2 pp 153ndash157 1947


[26] D L Bailey and K P Willowson ldquoAn evidence-based reviewof quantitative SPECT imaging and potential clinical applica-tionsrdquo Journal of NuclearMedicine vol 54 no 1 pp 83ndash89 2013

[27] M Vanderhoek S B Perlman and R Jeraj ldquoImpact of thedefinition of peak standardized uptake value on quantificationof treatment responserdquo Journal of Nuclear Medicine vol 53 no1 pp 4ndash11 2012

[28] P Tylski S Stute N Grotus et al ldquoComparative assessment ofmethods for estimating tumor volume and standardized uptakevalue in 18F-FDG PETrdquo Journal of Nuclear Medicine vol 51 no2 pp 268ndash276 2010

Submit your manuscripts athttpswwwhindawicom

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


MEDIATORSINFLAMMATION

of


Behavioural Neurology

EndocrinologyInternational Journal of



Disease Markers


BioMed Research International

OncologyJournal of



Oxidative Medicine and Cellular Longevity


PPAR Research

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

ObesityJournal of



Computational and Mathematical Methods in Medicine

OphthalmologyJournal of


Diabetes ResearchJournal of



Research and TreatmentAIDS


Gastroenterology Research and Practice


Parkinsonrsquos Disease

Evidence-Based Complementary and Alternative Medicine

Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom


Baseline ampfollow-up

images

Imagescoregistration

Valuenormalization

Parametricimaging

Change-ratemap

Figure 1 The workflow of the computer-aided analysis method to extract a change-rate map

of images for individual treatment monitoring In ictal-interictal SPECT imaging the subtraction analysis has beenproved to benefit the treatment plan of epilepsy patients [11ndash14] Recent studies have reported that the quantitative SPECTanalysis would be playing an ever-growing role in treatmentplan and response monitoring of several disorders relatedwith the central nervous system [15 16]These studies encour-age us to generate the parametric imaging and extract quan-titative indexes from the SPECT images to support the treat-ment plan In this study we introduce a computer-aided anal-ysis (CAA) method inherited from subtraction analysis toquantify the changes of rCBF in longitudinal SPECT imagesfor individual treatment monitoringThe performance of theproposed method would be objectively and systematicallyevaluated by the computer simulations and the clinical appli-cations

2 Materials and Methods

21 Computer-Aided Analysis Method When using SPECTimaging in treatment monitoring the pre- and postscansare usually performed to acquire the baseline and follow-upSPECT images before and after delivering treatments Theobtained baseline and follow-up images from an individualsubject are a set of longitudinal SPECT images requiringindependent analysis For interpreting the individual subjectrsquosdata a computer-aided analysis (CAA)method is establishedto process the longitudinal SPECT images via three mainsteps coregistration value normalization and parametricimaging The workflow chart is shown in Figure 1

In the first step the follow-up SPECT brain images arealigned with the baseline images by the rigid registrationalgorithm provided by SPM 80 software package (httpwwwfilionuclacukspm) If the SPECT brain imaging isperformed using SPECT-CT integrated system the CT andSPECT images can be obtained in the same position andconsidered as well aligned In this condition the baselineand follow-up CT images could be used as the reference andsource image respectively in the step of coregistration foraligning longitudinal SPECT images

In the second step the value normalization is appliedon the SPECT images Before the numerical calculation theextracerebral voxels are removed by a predefinedwhole-brainmask If only the SPECT image is available then the whole-brain mask can be obtained by segmenting the enhancedSPECT image with Otsursquos algorithm If the correspondingaligned CT image is available then the whole-brainmask canbe more accurately defined by separating the brain tissuesfromnonbrain tissue inCT images using fuzzy119862-means clus-tering algorithm [17] The whole-brain mask extracted basedon the CT images is then applied on the SPECT images todelineate the brain area

After deriving the whole-brain area in the SPECT imagesthe value of each cerebral voxel is normalized by the averagevoxel value of the reference area that is automatically selectedby 119885-map approach [18] In 119885-map approach the 119885 value ofthe 119894th voxel was calculated as in (1) Two119885-maps are respec-tively estimated for baseline and follow-up SPECT brainimages Then the reference region is the intersection of the119885 lt 1 areas of these two 119885-maps

119885119894 =1003816100381610038161003816119862119894 minusmean1003816100381610038161003816

SD (1)

where 119862119894 is the value of the 119894th voxel in one SPECT brainimage mean and SD respectively denote the average andstandard deviation of voxel values of brain area in the SPECTimage

In the third step the changes in the longitudinal SPECTimages are expressed in a parametric image to reflect thedisease progress or responses to the treatment As in sub-traction analysis the difference can be directly obtained bysubtracting two aligned normalized images as

119863119894 = 119862119891119894 minus 119862119887119894 (2)

where 119862119891119894 and 119862119887119894 denote the normalized values of the 119894th

voxel in the follow-up and baseline SPECT images respec-tively

Next the change-rate map can be calculated voxel-by-voxel to reflect the extent of the changes between the baselineand follow-up images The value of the 119894th voxel in theestimated change-rate map is denoted by 119877119894 which can bederived from

119877119894 =119863119894119862119887119894 (3)

The change-rate map (CRM) is a parametric image whichcan be fused with aligned SPECTCT images for visualinspections The positive voxel value in CRM demonstratesthe recovery of hypoperfusion while the negative value indi-cates the recovery of hyperperfusion For the visualizationGaussian smoothing filter can be applied in the CRM toreduce the impact of noise Based on the CRM the changedregions are automatically obtained by thresholding and clus-tering Firstly the voxels with lower change-rate (lt20) wereset to 0 in CRM Then the 119870-means clustering algorithm isadopted to recognize the changed regions The morphologi-cal processing is applied to refine and distinguish each regionConsidering the SPECT image resolution and partial volumeeffects the regions with larger volumes (gt120 voxels) areselected as the recovered regions

For the localization of recovered regions an atlas of brainlobes which consists of 12 brain anatomical structures (listed











NAE =sum119873119894=11003816100381610038161003816119877119894 minus 1198661198941003816100381610038161003816

sum119873119894=1 119866119894 (4)



MSE = 1119873

119873

sum119894=1

(119877119894 minus 119866119894)2

(5)



NCC = 1119873 minus 1

119873

sum119894=1

(119877119894 minus 119877) (119866119894 minus 119866)120590119888120590119866

(6)





(7)








3 Results



00 10 20 30 40 50 60 70 80 90 100

50

100

150

200

250

300

350

Change-rate ()

NA

E

empty24

empty16

empty8

(a)

0 10 20 30 40 50 60 70 80 90 100Change-rate ()

5

10

15

20

25

30

PSN

R

empty24

empty16

empty8

(b)

0 10 20 30 40 50 60 70 80 90 100Change-rate ()

00

02

04

06

08

10

NCC

empty24

empty16

empty8

(c)







DSC

0 10 20 30 40 50 60 70 80 90 100Change-rate ()

empty24

empty16

empty8

00

02

04

06

08

10

(a)

0102030405060708090

100

Standard CR ()

Estim

ated

CR

()

0 10 20 30 40 50 60 70 80 90 100

empty24 y = 096x minus 236 r2 = 099

empty16 y = 066x + 371 r2 = 099

empty8 y = 001x + 118 r2 = 001

(b)










36 40 44

(a)

36 40 44

(b)

36 40 44

Change-rate ()10 15 20 25 30 35 40 45

(c)

Change-rate ()10 15 20 25 30 35 40 45

36 40 44

(d)






4 Discussion



S M MN20

22

24

26

28

30

Mea

n CR

()

(a)S M MN

20

30

40

50

60

70

80

Max

CR

()

(b)S M MN

0

5

10

15

20

25

Prop

ortio

n of

RRs

()

(c)







S_20 M_20 M_105

10

15

20

25

30

35

Mea

n CR

()

(a)

S_20 M_20 M_100

5

10

15

20

25

Prop

ortio

n of

RRs

()

(b)






5 Conclusion





Competing Interests


Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


























NAE =sum119873119894=11003816100381610038161003816119877119894 minus 1198661198941003816100381610038161003816

sum119873119894=1 119866119894 (4)



MSE = 1119873

119873

sum119894=1

(119877119894 minus 119866119894)2

(5)



NCC = 1119873 minus 1

119873

sum119894=1

(119877119894 minus 119877) (119866119894 minus 119866)120590119888120590119866

(6)





(7)








3 Results



00 10 20 30 40 50 60 70 80 90 100

50

100

150

200

250

300

350

Change-rate ()

NA

E

empty24

empty16

empty8

(a)

0 10 20 30 40 50 60 70 80 90 100Change-rate ()

5

10

15

20

25

30

PSN

R

empty24

empty16

empty8

(b)

0 10 20 30 40 50 60 70 80 90 100Change-rate ()

00

02

04

06

08

10

NCC

empty24

empty16

empty8

(c)







DSC

0 10 20 30 40 50 60 70 80 90 100Change-rate ()

empty24

empty16

empty8

00

02

04

06

08

10

(a)

0102030405060708090

100

Standard CR ()

Estim

ated

CR

()

0 10 20 30 40 50 60 70 80 90 100

empty24 y = 096x minus 236 r2 = 099

empty16 y = 066x + 371 r2 = 099

empty8 y = 001x + 118 r2 = 001

(b)










36 40 44

(a)

36 40 44

(b)

36 40 44

Change-rate ()10 15 20 25 30 35 40 45

(c)

Change-rate ()10 15 20 25 30 35 40 45

36 40 44

(d)






4 Discussion



S M MN20

22

24

26

28

30

Mea

n CR

()

(a)S M MN

20

30

40

50

60

70

80

Max

CR

()

(b)S M MN

0

5

10

15

20

25

Prop

ortio

n of

RRs

()

(c)







S_20 M_20 M_105

10

15

20

25

30

35

Mea

n CR

()

(a)

S_20 M_20 M_100

5

10

15

20

25

Prop

ortio

n of

RRs

()

(b)






5 Conclusion





Competing Interests


Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of



















(7)








3 Results



00 10 20 30 40 50 60 70 80 90 100

50

100

150

200

250

300

350

Change-rate ()

NA

E

empty24

empty16

empty8

(a)

0 10 20 30 40 50 60 70 80 90 100Change-rate ()

5

10

15

20

25

30

PSN

R

empty24

empty16

empty8

(b)

0 10 20 30 40 50 60 70 80 90 100Change-rate ()

00

02

04

06

08

10

NCC

empty24

empty16

empty8

(c)







DSC

0 10 20 30 40 50 60 70 80 90 100Change-rate ()

empty24

empty16

empty8

00

02

04

06

08

10

(a)

0102030405060708090

100

Standard CR ()

Estim

ated

CR

()

0 10 20 30 40 50 60 70 80 90 100

empty24 y = 096x minus 236 r2 = 099

empty16 y = 066x + 371 r2 = 099

empty8 y = 001x + 118 r2 = 001

(b)










36 40 44

(a)

36 40 44

(b)

36 40 44

Change-rate ()10 15 20 25 30 35 40 45

(c)

Change-rate ()10 15 20 25 30 35 40 45

36 40 44

(d)






4 Discussion



S M MN20

22

24

26

28

30

Mea

n CR

()

(a)S M MN

20

30

40

50

60

70

80

Max

CR

()

(b)S M MN

0

5

10

15

20

25

Prop

ortio

n of

RRs

()

(c)







S_20 M_20 M_105

10

15

20

25

30

35

Mea

n CR

()

(a)

S_20 M_20 M_100

5

10

15

20

25

Prop

ortio

n of

RRs

()

(b)






5 Conclusion





Competing Interests


Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















00 10 20 30 40 50 60 70 80 90 100

50

100

150

200

250

300

350

Change-rate ()

NA

E

empty24

empty16

empty8

(a)

0 10 20 30 40 50 60 70 80 90 100Change-rate ()

5

10

15

20

25

30

PSN

R

empty24

empty16

empty8

(b)

0 10 20 30 40 50 60 70 80 90 100Change-rate ()

00

02

04

06

08

10

NCC

empty24

empty16

empty8

(c)







DSC

0 10 20 30 40 50 60 70 80 90 100Change-rate ()

empty24

empty16

empty8

00

02

04

06

08

10

(a)

0102030405060708090

100

Standard CR ()

Estim

ated

CR

()

0 10 20 30 40 50 60 70 80 90 100

empty24 y = 096x minus 236 r2 = 099

empty16 y = 066x + 371 r2 = 099

empty8 y = 001x + 118 r2 = 001

(b)










36 40 44

(a)

36 40 44

(b)

36 40 44

Change-rate ()10 15 20 25 30 35 40 45

(c)

Change-rate ()10 15 20 25 30 35 40 45

36 40 44

(d)






4 Discussion



S M MN20

22

24

26

28

30

Mea

n CR

()

(a)S M MN

20

30

40

50

60

70

80

Max

CR

()

(b)S M MN

0

5

10

15

20

25

Prop

ortio

n of

RRs

()

(c)







S_20 M_20 M_105

10

15

20

25

30

35

Mea

n CR

()

(a)

S_20 M_20 M_100

5

10

15

20

25

Prop

ortio

n of

RRs

()

(b)






5 Conclusion





Competing Interests


Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















DSC

0 10 20 30 40 50 60 70 80 90 100Change-rate ()

empty24

empty16

empty8

00

02

04

06

08

10

(a)

0102030405060708090

100

Standard CR ()

Estim

ated

CR

()

0 10 20 30 40 50 60 70 80 90 100

empty24 y = 096x minus 236 r2 = 099

empty16 y = 066x + 371 r2 = 099

empty8 y = 001x + 118 r2 = 001

(b)










36 40 44

(a)

36 40 44

(b)

36 40 44

Change-rate ()10 15 20 25 30 35 40 45

(c)

Change-rate ()10 15 20 25 30 35 40 45

36 40 44

(d)






4 Discussion



S M MN20

22

24

26

28

30

Mea

n CR

()

(a)S M MN

20

30

40

50

60

70

80

Max

CR

()

(b)S M MN

0

5

10

15

20

25

Prop

ortio

n of

RRs

()

(c)







S_20 M_20 M_105

10

15

20

25

30

35

Mea

n CR

()

(a)

S_20 M_20 M_100

5

10

15

20

25

Prop

ortio

n of

RRs

()

(b)






5 Conclusion





Competing Interests


Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















36 40 44

(a)

36 40 44

(b)

36 40 44

Change-rate ()10 15 20 25 30 35 40 45

(c)

Change-rate ()10 15 20 25 30 35 40 45

36 40 44

(d)






4 Discussion



S M MN20

22

24

26

28

30

Mea

n CR

()

(a)S M MN

20

30

40

50

60

70

80

Max

CR

()

(b)S M MN

0

5

10

15

20

25

Prop

ortio

n of

RRs

()

(c)







S_20 M_20 M_105

10

15

20

25

30

35

Mea

n CR

()

(a)

S_20 M_20 M_100

5

10

15

20

25

Prop

ortio

n of

RRs

()

(b)






5 Conclusion





Competing Interests


Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















S M MN20

22

24

26

28

30

Mea

n CR

()

(a)S M MN

20

30

40

50

60

70

80

Max

CR

()

(b)S M MN

0

5

10

15

20

25

Prop

ortio

n of

RRs

()

(c)







S_20 M_20 M_105

10

15

20

25

30

35

Mea

n CR

()

(a)

S_20 M_20 M_100

5

10

15

20

25

Prop

ortio

n of

RRs

()

(b)






5 Conclusion





Competing Interests


Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















S_20 M_20 M_105

10

15

20

25

30

35

Mea

n CR

()

(a)

S_20 M_20 M_100

5

10

15

20

25

Prop

ortio

n of

RRs

()

(b)






5 Conclusion





Competing Interests


Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


















Competing Interests


Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

























of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
















a computer-aided analysis method of spect brain images for...

Documents