B R A I N R E S E A R C H 1 4 5 8 ( 2 0 1 2 ) 8 6 – 9 2

Ava i l ab l e on l i ne a t www.sc i enced i r ec t . com

www.e l sev i e r . com/ loca te /b ra i n res

Research Report

Incorporating ADC temporal profiles to predict ischemic tissuefate in acute stroke

Virendra Desai, Qiang Shen, Timothy Q. Duong⁎

Research Imaging Institute, Department of Ophthalmology and Radiology, University of Texas Health Science Center, San Antonio, TX, USA

A R T I C L E I N F O

⁎ Corresponding author at: University of TexaAntonio, TX 78229, USA. Fax: +1 210 567 8152

E-mail address: duongt@uthscsa.edu (T.Q

0006-8993/$ – see front matter © 2012 Elseviedoi:10.1016/j.brainres.2012.04.004

A B S T R A C T

Article history:Accepted 5 April 2012Available online 20 April 2012

Algorithms to predict ischemic tissue fate based on acute strokeMRI typically utilized data at asingle time point. The goal of this study was to investigate the potential improvement inprediction accuracy when incorporating MRI diffusion data from multiple time points duringacute phase to improve prediction accuracy. This study was carried out using MRI data fromrats subjected to permanent, 60-min and 30-min of middle cerebral artery occlusion (MCAO).The sensitivity and specificity of prediction accuracywere calculated. In the permanentMCAOgroup, prediction withmultiple time-point diffusion data improved sensitivity and specificitycomparedwith prediction using a single timepoint. In the 60-minMCAOgroup,multiple time-point analysis improved specificity but decreased sensitivity compared to the single time-point analysis. In the 30-min MCAO group, multiple time-point analysis showed nostatistically significant improvement in specificity and sensitivity compared with the singletime point analysis. This is because reperfusion transiently or permanently reversed thedecline in ADC values, resulting in increased uncertainty and thus decreased predictionaccuracy. Incorporating this a priori information could further improve prediction accuracy inthe reperfusion group. These findings suggest that incorporatingMRI data frommultiple timepoints could improve prediction accuracy under certain ischemic conditions.

© 2012 Elsevier B.V. All rights reserved.

Keywords:DiffusionPerfusion–diffusion mismatchMCAOFocal ischemiaDWIPWIADCCBF

1. Introduction

Stroke is the fourth leading cause of mortality and theleading cause of long-term disability in the United States(Roger et al., 2012). The only FDA-approved drug to treatischemic stroke is intravenous administration of recombi-nant tissue plasminogen activator (rtPA) within 4.5 h ofstroke onset (Hacke et al., 2008). Unfortunately, only 1.8–2.1% of ischemic stroke patients receive treatment with rtPA(Kleindorfer et al., 2008). Imaging modalities have thepotential to identify injured but salvageable tissue, knownas the ischemic penumbra. In some patients, salvageable

s Health Science Center a.. Duong).

r B.V. All rights reserved

tissue exists well beyond the 4.5 hour time window (i.e., upto 24 h after symptom onset (Darby et al., 1999)). Thus, thereis value to accurately predict which group of stroke patientswill benefit from therapeutic interventions.

When cerebral blood flow (CBF) drops below a criticalthreshold, energetic failure results and the apparent diffusioncoefficient (ADC) of water in the tissue starts to decrease(Moseley et al., 1990), although the precise biophysicalmechanisms of ADC reduction remain incompletely under-stood (Duong et al., 1998). Diffusion-weighted magnetic reso-nance imaging (MRI) in which image contrast is based onwater ADC can detect ischemic injury within minutes after

t San Antonio, Research Imaging Institute, 8403 Floyd Curl Dr, San

.

87B R A I N R E S E A R C H 1 4 5 8 ( 2 0 1 2 ) 8 6 – 9 2

onset, whereas computed tomography and other imagingmodalities fail to detect stroke injury for at least a few hours(Moseley et al., 1990). The critical ADC threshold below whichtissue usually destines to infarct has been reported to be0.53×10−3 mm2/s (Shen et al., 2003). However, the evolution ofthe initial ADC lesion depends on many conditions (such asduration of ischemia and extent of occlusion or reperfusion).Some tissue with initial ADC reduction is salvageable whileother is not (Kidwell et al., 2003; Li et al., 1999). Despite itsuncertainty differentiating salvageable from non-salvageabletissue, diffusion-weighted MRI remains commonly used toguide clinical decision making in acute stroke management(Kidwell et al., 2003).

Various sophisticated algorithms have been developed toquantitatively predict ischemic tissue fate, including thegeneralized linear model (Wu et al., 2001; Wu et al., 2007),probability-of-infarct (Shen and Duong, 2008; Shen et al.,2005b), artificial neural network (Huang et al., 2010) andsupport vector machine (Huang et al., 2011). These predictionalgorithms incorporated imaging data from a single timepoint. Tissue ADC changes with time after ischemic injury.Incorporating ADC data from multiple time points couldimprove prediction accuracy. The goal of this study was thusto investigate the potential improvement in predictionaccuracy by incorporating ADC measurements at multipletime points during acute stroke phase. We investigated datafrom rats subjected to permanent, 60-min and 30-min of

Table 1 – Single time-point and multiple time-point analyses. Fortissue comprising that cluster which is destined to infarct at thtissue made up by each cluster. This is shown for (A) the perm

(A)

Permanen

Single time-point

Cluster % infarcting % of ti

Blue 25 ± 18 15 ±

Green 45 ± 21 32 ±

Yellow 86 ± 9 29 ±

Red 96 ± 7 24 ±

(B)

60-min M

Single time-point

Cluster % infarcting % of ti

Blue 15 ± 9 12 ±

Green 21 ± 10 40 ±

Yellow 77 ± 9 19 ±

Red 91 ± 5 29 ±

(C)

30-min M

Single time-point

Cluster % infarcting % of ti

Blue 7 ± 8 19 ±

Green 11 ± 16 36 ±

Yellow 52 ± 29 20 ±

Red 73 ± 25 25 ±

middle cerebral artery occlusion (MCAO). The sensitivity andspecificity of the prediction accuracy were calculated, andcomparisons were made with the prediction accuracy whenusing only a single acute time point for each MCAO group.

2. Results

With k-means clustering, the ADC data for each rat wasseparated into four apparent temporal clusters, with each oneshowing a different pattern across time. We investigated eachcluster's proportion of total tissue, infarction rate, variabilityin infarction rate and ADC trend across time. Sensitivity andspecificity calculations formed the prediction analysis. Eachof these clusters was mapped onto image space.

2.1. Permanent group

Table 1A shows that both the single and multiple time-pointmethods divided the tissue into similar proportions percluster, with the blue clusters making up about 15% of thetotal tissue, the green clusters about 30%, and the yellow andred clusters between 24% and 29%. Both methods had similarrates of infarction for each cluster with near 100% infarctionfor the red clusters, a high percentage of infarction for theyellow clusters, a moderate percentage for the green clustersand a low percentage for the blue clusters. Moreover, the

each cluster, “% infarcting” represents the percentage ofe end point. “% of tissue” represents the percentage of totalanent, (B) the 60-min, and (C) the 30-min MCAO groups.

t MCAO

Multiple time-point

ssue % infarcting % of tissue

7 17 ± 13 15 ± 10

13 38 ± 25 28 ± 11

9 89 ± 10 28 ± 13

8 98 ± 2 29 ± 8

CAO

Multiple time-point

ssue % infarcting % of tissue

4 17 ± 7 25 ± 14

11 31 ± 28 39 ± 15

5 86 ± 8 18 ± 8

7 98 ± 2 19 ± 7

CAO

Multiple time-point

ssue % infarcting % of tissue

7 8 ± 10 19 ± 8

9 18 ± 22 36 ± 5

5 47 ± 38 24 ± 10

6 69 ± 27 21 ± 6

Fig. 1 – Permanent MCAO group. (A) Time-series ADCmaps from a representative animal, chosen based on its superior depictionof lesion evolution. The image labeled “final infarct” indicates the endpoint tissue outcome – survival or infarct – determinedvia ISODATA analysis. The image labeled “cluster map” depicts an example from the same rat of the tissue clustered intofour groups, with the blue pixels representing the cluster with the highest ADC values, the green pixels representing the nexthighest and so forth. As this map represents the clusters from only one rat, it may not exactly correlate with the numbersshown in Table 1, which are group averages. (B) Averaged ADC values of different clusters versus time for the single time-pointgroup, and (C) averaged ADC values of different clusters versus time for the multiple time-point group. The color of the pixelsin the cluster maps correlates with the color of the ADC curves shown in the graphs.

88 B R A I N R E S E A R C H 1 4 5 8 ( 2 0 1 2 ) 8 6 – 9 2

green and blue clusters showed considerable variability in thepercent of tissue destined to infarct at the endpoint. In Fig. 1,while both methods had monotonically decreasing ADCcurves for the blue, green and yellow clusters, the singletime-point method had a stable red cluster across timewhereas the multiple time-point method showed an initialdecrease followed by a plateau in the red cluster.

The prediction analysis (Table 2) showed that the multi-ple time-point method had higher sensitivity (80% versus73%, p=0.009) and higher specificity (89% versus 85%,p=0.045). Thus, incorporating the temporal behavior of ADCinto tissue fate analysis improved prediction accuracy inpermanent MCAO.

2.2. 60-min MCAO group

Table 1B shows that both clustering methods had the greenclusters making up about 40% of total tissue and the yellow

Table 2 – Prediction analysis. For each group, the sensitivity an

Permanent MCAO

Singletime-point

Multipletime-point

Singletime-po

Sensitivity 73±16% 80±19% 80±8%Specificity 85±10% 89±5% 86±9%

clusters almost 20%. For the remaining tissue, the single time-point method clustered the majority of it into the red clusterwhile the multiple time-point method clustered the majorityof it into the blue cluster. In other words, the single time-pointmethod predicted more tissue destined to infarct at theendpoint than the multiple time-point method. Each cluster'sfinal outcome varied between the two analytical methods.While both methods had similar rates of infarction for theblue clusters, the multiple time-point method had higherpercentages of infarction for the green, yellow and redclusters. The green cluster in the multiple time-point analysishad significant variability in the percent of tissue destined toinfarct at the endpoint. These findings suggest increasedcertainty when predicting infarction but decreased certaintywhen predicting survival with the multiple time-point meth-od. In Fig. 2, the ADC curves for each method were slightlydifferent. The single time-point method showed all theclusters decreasing before reperfusion. The blue cluster was

d specificity calculations are shown.

60-min MCAO 30-min MCAO

intMultiple

time-pointSingle

time-pointMultiple

time-point

67±15% 84±12% 76±14%94±3% 75±11% 71±18%

Fig. 2 – 60-min MCAO group. (A) Time-series ADCmaps from a representative animal, chosen based on its superior depiction oflesion evolution with occlusion and resolution with reperfusion. The image labeled “final infarct” shows the endpoint tissueoutcome – survival or infarct – determined via ISODATA analysis. The image labeled “cluster map” depicts an example fromthe same rat of the tissue clustered into four color-coded groups. As this map represents the clusters from only one rat, it maynot exactly correlate with the numbers in Table 1, which are group averages. (B) Averaged ADC values of different clustersversus time for the single time-point group, and (C) averaged ADC values of different clusters versus time for the multipletime-point group. The color of the pixels in the cluster maps correlates with the color of the ADC curves shown in the graphs.“REP” and arrow on x-axis indicate the time of reperfusion.

89B R A I N R E S E A R C H 1 4 5 8 ( 2 0 1 2 ) 8 6 – 9 2

relatively stable after reperfusion while the remaining clus-ters showed recovery after reperfusion. The multiple time-point method showed a relatively stable blue cluster bothbefore and after reperfusion, a monotonically decreasing redcluster, and a transiently decreasing with post-reperfusionrecovery of the green and yellow clusters.

The prediction analysis (Table 2) showed that the multipletime-point method had decreased sensitivity (67% versus 80%,p=0.016) higher specificity (94% versus 86%, p=0.002). Thus,incorporating the temporal behavior of ADC into tissue fateanalysis decreased sensitivity but improved specificity in ratssubjected to 60-min MCAO.

2.3. 30-min MCAO group

Table 1C shows that both clustering methods divided thetissue into similar proportions per cluster, with the greenclusters making up about 36% of the total tissue while theremaining clusters made up between 19% and 25%. The finaloutcomes of each cluster were similar between the twoanalytical methods, with a high percentage of infarction forthe red clusters, around 50% infarction for the yellow clusters,and a low percentage for the green and blue clusters. In bothgroups, the yellow and red clusters had considerable variabil-ity in the percent destined to infarct at the endpoint. Theblue and green clusters sustained relatively stable ADCvalues across time while the yellow and red clusters showedADC recovery with reperfusion for both methods (Fig. 3).

The prediction analysis (Table 2) showed that the twomethods were not statistically significantly different in sen-sitivity (76% versus 84%, p=0.09) and specificity (71% versus75%, p=0.29).

3. Discussion

This study reports the use of acute ADC data at multiple timepoints to improve prediction accuracy in a rat stroke model ofpermanent, 60-min and 30-min MCAO. Compared with predic-tion using a single time point, multiple time-point analysisimproved prediction accuracy in the permanent MCAO group.Multiple time-point analysis improved specificity but decreasedsensitivity in the 60-min MCAO group compared to single timepoint analysis. Multiple time-point analysis showed no statis-tically significant improvement in specificity and sensitivitycompared with single time point analysis in the 30-min MCAOgroup. This is because reperfusion transiently or permanentlyreversed the decline in ADC values, resulting in increaseduncertainty and thus decreased prediction accuracy.

In the permanent and the 30-min groups, the proportions,infarction rates and ADC trends were similar across bothsingle and multiple time-point analyses. In the 60-min group,all of these differed between the two methods. In the singletime-point method, the red cluster composed a higher pro-portion of total tissue, had a lower infarction rate and showedADC recovery with reperfusion. This suggests the single time-

Fig. 3 – 30-min MCAO group. (A) Time-series ADC maps from a representative animal, chosen based on its superior depictionof lesion evolution with occlusion and resolution with reperfusion. The image labeled “final infarct” shows the endpointtissue outcome – survival or infarct – determined via ISODATA analysis. The image labeled “cluster map” depicts an examplefrom the same rat of the tissue clustered into four color-coded groups. As this map represents the clusters from only one rat,it may not exactly correlate with the numbers in Table 1, which are group averages. (B) Averaged ADC values of differentclusters versus time for the single time-point group, and (C) averaged ADC values of different clusters versus time for themultiple time-point group. The color of the pixels in the cluster maps correlates with the color of the ADC curves shown inthe graphs. “REP” and arrow on x-axis indicate the time of reperfusion.

90 B R A I N R E S E A R C H 1 4 5 8 ( 2 0 1 2 ) 8 6 – 9 2

point method predicted a higher rate of infarction than themultiple time-point method by including in the red clustersome pixels with initially low ADC values that recovered withreperfusion, resulting in its increased sensitivity but de-creased specificity in predicting infarct tissue.

The single time-point method was not able to differentiatepixels with initially low ADC values that continued to declinefrom those that recovered with reperfusion. By contrast, themultiple time-point method was able to make this differen-tiation, and it removed from the red cluster all the pixels thatshowed ADC recovery with reperfusion, utilizing this tem-poral information to increase the certainty in its prediction oftissue destined to infarct. Thus, it is evident that ADC re-covery with reperfusion did indeed signify potential forsurvival, though it was not absolute. These findings demon-strate the value of temporal information.

This study has a few shortcomings. First, the clusteringmethod is suboptimal as the k-means clustering is based onthe sum of squares of differences in ADC values at each timepoint. Future studies will improve clustering method to betteraccount the temporal information from multiple time points.Second, the endpoint imaging was 24 h after MCAO. Althoughthe infarct volume had largely stopped evolving at this time,there could be additional infarct growth, especially for the 30-min and 60-min MCAO groups (Li et al., 1999). Similar analysisof data at a later endpoint (such as 48 or 72 h) will need to beinvestigated to determine how the choice of endpoint MRIaffects single and multiple time-point analyses. Third, our

analysis used only ADC data. Incorporating additional imag-ing data (such as CBF) could further improve predictionaccuracy. Finally, although performance was evaluated bysensitivity and specificity calculations, future studies willutilize more sophisticated algorithms, such as support vectormachine with separate training and experimental groups, toquantitatively predict tissue fate (Huang et al., 2011).

Prediction based on data at multiple time points has thepotential to provide quantitative and objective frameworks toextend the treatment window for stroke patients and to aidclinical decision-making in the treatment of acute stroke anddrug testing. This approach may also be applied to patientswith transient ischemic attack who often return to theemergency room with a large stroke within 48 h (Rothwelland Warlow, 2005). Objective and accurate prediction modelsof tissue fate may help drug trials by accelerating theidentification of promising potential therapies and patientselection. It may also help to individualize the treatmentwindow for stroke patients.

4. Experimental procedures

4.1. Animal preparations

The data analyzed in this study were those previouslypublished (Shen and Duong, 2008). Briefly, a total of 35Sprague–Dawley rats (300–350 g) were subjected to 30-min

91B R A I N R E S E A R C H 1 4 5 8 ( 2 0 1 2 ) 8 6 – 9 2

(n=13), 60-min (n=12) and permanent (n=10) MCAO, withreperfusion accomplished remotely without taking the ani-mal out of the scanner.

4.2. MR experiments

MRI scanner specifications and settings as well as imageacquisition methods were as described elsewhere (Shen andDuong, 2008). Briefly, ADC and CBF data were collected at 30,60, 90, 120, and 180 min, and again at 24 h post-occlusion forall three study groups. For the 30-min and 60-min MCAOgroups, the 30-min and 60-min data, respectively, wereacquired before reperfusion. Additionally, for the 30-min and60-min MCAO groups, an additional imaging time point wasperformed 10 min post-reperfusion, at 40-min and at 70-minrespectively. Endpoint T2-weighted MRI was also performedat 24 h post-occlusion. Histological infarct volume wasdetermined using TTC (2,3,5-triphenyltetrazolium chloride)staining and with edema correction (Meng et al., 2004).

4.3. Data analysis

Five anterior slices were analyzed to avoid susceptibilitydistortion around the ear canals. Images were co-registeredusing custom-designed semi-automatic co-registration soft-ware between acute phase and 24-hour data within thesame animals and between animals as described previously(Liu et al., 2004; Schmidt et al., 2006; Shen et al., 2005a). ADCmaps with intensity in unit of mm2/s (Meng et al., 2004;Shen et al., 2005a) and CBF maps with intensity in units ofmL/g/min were calculated (Duong et al., 2000). Image dis-plays and overlays were performed on the STIMULATEsoftware (University of Minnesota). All data were reported asmean±SEM.

4.4. K-means clustering

Codes written in Matlab (MathWorks, Natick, MA) were usedto perform k-means clustering analysis. First, data wereanalyzed via a single time point — the ADC values at the30-min time point for each rat were used to form four clusters.This was done by first clustering the tissue into two groups,then clustering each of these groups further into two sub-groups. For each cluster, the ADC values of the pixels wereaveraged together to form the curves. The endpoint tissueoutcome for each group was determined by comparison withISODATA at 24 h. The percentage of tissue made up by eachcluster was calculated as well as the percent of each clusterto infarct at the endpoint.

The 30-min time point data were chosen for single time-point analysis because in our rat stroke model, there was asubstantial mismatch at 30-min after stroke but it disap-peared within 2–3 h after stroke onset (Shen et al., 2003),although in humans the mismatch exists considerably longer(Hacke et al., 2008). The 30-min time point was also chosenbecause only a pre-reperfusion time point can be reasonablyused for single time-point analysis, as prediction using a post-reperfusion time point when ADC decline has reversed wasunreliable. To be consistent, the 30-min time point data wereused for all three groups.

Data were analyzed via multiple time points – all the timepoints for the reperfusion groups and the first four time pointsfor the permanent group – forming four ADC clusters. Again,the tissue was clustered into two groups which were eachsplit further to form a total of four clusters. Only the first fourtime points were used for the permanent group as the fifthand last time point was used to determine endpoint tissueoutcome. And the percentage of tissue made up by eachcluster was calculated as well as the percent of each clusterto infarct at the endpoint.

4.5. Prediction analysis

The single time-point analysis was compared with themultiple time-point analysis via sensitivity and specificitycalculations. The four curves obtained via k-means werecombined into two groups — one which included the lowertwo clusters, red and yellow, and the other which included thehigher two clusters, green and blue. The grouping wasperformed in this manner as the lower ADC group includedthe clusters having any ADC value near or below the thresholdvalue of 0.53×10−3 mm2/s. This threshold value was deter-mined as the value at which the ADC-defined lesion volumebest approximated the TTC-identified final infarct volume(Shen et al., 2003). The higher ADC group, consisting of ADCvalues well above threshold, was assumed to predict 100%survival, while the lower group, consisting of any ADC valuebelow or near threshold, was assumed to predict 100%infarction. The sensitivity and specificity for each groupwere then determined by comparing the predictions withthe actual outcome determined by ISODATA. These werecalculated for each rat, and subsequently, the means andstandard deviations were calculated.

4.6. Statistical analysis

For each group, the sensitivity and specificity values of thesingle time-point analysis were compared to that of themultiple time-point analysis via paired one-tailed t-test, andp<0.05 was taken as significant.

Acknowledgment

This work was supported by the NIH (R01-NS45879) and theAmerican Heart Association (EIA 0940104N, SDG-0430020N,SDG-0830293N and 12BGIA9300047).

R E F E R E N C E S

Darby, D.G., Barber, P.A., Gerraty, R.P., Desmond, P.M., Yang, Q.,Parsons, M., Li, T., Tress, B.M., Davis, S.M., 1999.Pathophysiological topography of acute in ischemia bycombined diffusion-weighted and perfusion MRI. Stroke 30,2043–2052.

Duong, T.Q., Ackerman, J.J.H., Ying, H.S., Neil, J.J., 1998. Evaluationof extra- and intracellular apparent diffusion in normal andglobally ischemic rat brain via 19F NMR. Magn. Reson. Med. 40,1–13.

92 B R A I N R E S E A R C H 1 4 5 8 ( 2 0 1 2 ) 8 6 – 9 2

Duong, T.Q., Silva, A.C., Lee, S.-P., Kim, S.-G., 2000. Functional MRIof calcium-dependent synaptic activity: cross correlation withCBF and BOLD measurements. Magn. Reson. Med. 43, 383–392.

Hacke, W., Kaste, M., Bluhmki, E., Brozman, M., Davalos, A.,Guidetti, D., Larrue, V., Lees, K.R., Medeghri, Z., Machnig, T.,Schneider, D., von Kummer, R., Wahlgren, N., Toni, D., 2008.Thrombolysis with alteplase 3 to 4.5 hours after acute ischemicstroke. N. Engl. J. Med. 359, 1317–1329.

Huang, S., Shen, Q., Duong, T.Q., 2010. Artificial neural-networkprediction of ischemic tissue fate in acute stroke imaging.J. Cereb. Blood Flow Metab. 30, 1661–1670.

Huang, S., Shen, Q., Duong, T.Q., 2011. Quantitative prediction ofacute ischemic tissue fate using support vector machine. BrainRes. 1405, 77–84.

Kidwell, C.S., Alger, J.R., Saver, J.L., 2003. Beyond mismatch:evolving paradigms in imaging the ischemic penumbra withmultimodal magnetic resonance imaging. Stroke 34,2729–2735.

Kleindorfer, D., Lindsell, C.J., Brass, L., Koroshetz, W., Broderick,J.P., 2008. National US estimates of recombinant tissueplasminogen activator use: ICD-9 codes substantiallyunderestimate. Stroke 39, 924–928.

Li, F., Han, S.S., Tatlisumak, T., Liu, K.F., Garcia, J.H., Sotak, C.H.,Fisher, M., 1999. Reversal of acute apparent diffusioncoefficient abnormalities and delayed neuronal deathfollowing transient focal cerebral ischemia in rats. Ann.Neurol. 46, 333–342.

Liu, Z.M., Schmidt, K.F., Sicard, K.M., Duong, T.Q., 2004. Imagingoxygen consumption in forepaw somatosensory stimulationin rats under isoflurane anesthesia. Magn. Reson. Med. 52,277–285.

Meng, X., Fisher, M., Shen, Q., Sotak, C.H., Duong, T.Q., 2004.Characterizing the diffusion/perfusion mismatch inexperimental focal cerebral ischemia. Ann. Neurol. 55,207–212.

Moseley, M.E., Cohen, Y., Mintorovitch, J., Chileuitt, L., Shimizu, H.,Kucharczyk, J., Wendland, M.F., Weinstein, P.R., 1990. Earlydetection of regional cerebral ischemia in cats: comparison ofdiffusion- and T2-weighted MRI and spectroscopy. Magn.Reson. Med. 14, 330–346.

Roger, V.L., Go, A.S., Lloyd-Jones, D.M., Benjamin, E.J., Berry, J.D.,Borden, W.B., Bravata, D.M., Dai, S., Ford, E.S., Fox, C.S.,Fullerton, H.J., Gillespie, C., Hailpern, S.M., Heit, J.A., Howard,V.J., Kissela, B.M., Kittner, S.J., Lackland, D.T., Lichtman, J.H.,Lisabeth, L.D., Makuc, D.M., Marcus, G.M., Marelli, A., Matchar,D.B., Moy, C.S., Mozaffarian, D., Mussolino, M.E., Nichol, G.,Paynter, N.P., Soliman, E.Z., Sorlie, P.D., Sotoodehnia, N.,Turan, T.N., Virani, S.S., Wong, N.D., Woo, D., Turner, M.B.,2012. Heart disease and stroke statistics—2012 update: a reportfrom the American Heart Association. Circulation 125, e2–e220.

Rothwell, P.M., Warlow, C.P., 2005. Timing of transient ischemicattacks preceding ischemic stroke. Neurology 64, 817–820.

Schmidt, K.F., Febo, M., Shen, Q., Ferris, C.F., Stein, E., Duong, T.Q.,2006. Hemodynamic and metabolic changes induced bycocaine in anesthetized rat observed with multimodalfunctional. Psychopharmacology 185, 479–486.

Shen, Q., Meng, X., Fisher, M., Sotak, C.H., Duong, T.Q., 2003.Pixel-by-pixel spatiotemporal progression of focal ischemiaderived using quantitative perfusion and diffusion imaging.J. Cereb. Blood Flow Metab. 23, 1479–1488.

Shen, Q., Ren, H., Cheng, H., Fisher, M., Duong, T.Q., 2005a.Functional, perfusion and diffusion MRI of acute focalischemic brain injury. J. Cereb. Blood Flow Metab. 25,1265–1279.

Shen, Q., Ren, H., Fisher, M., Duong, T.Q., 2005b. Statisticalprediction of tissue fates in acute ischemic brain injury.J. Cereb. Blood Flow Metab. 25, 1336–1345.

Shen, Q., Duong, T.Q., 2008. Quantitative prediction of ischemicstroke tissue fate. NMR Biomed. 21, 839–848.

Wu, O., Koroshetz, W.J., Ostergard, L., Buonanno, F.S., Copen, W.,Gonzales, G., Rordorf, G., Rosen, B.R., Schwamm, L.H.,Weisskoff, R.M., Sorensen, A.G., 2001. Predicting tissueoutcome in acute human cerebral ischemia using combineddiffusion-and perfusion-weighted MR imaging. Stroke 32,933–942.

Wu, O., Sumii, T., Asahi, M., Sasamata, M., Ostergaard, L., Rosen,B.R., Lo, E.H., Dijkhuizen, R.M., 2007. Infarct prediction andtreatment assessment with MRI-based algorithms inexperimental stroke models. J. Cereb. Blood Flow Metab. 27,196–204.