Original Research

Predicting Final Infarct Size Using Acute andSubacute Multiparametric MRI Measurements inPatients With Ischemic Stroke

Mei Lu, PhD,1* Panayiotis D. Mitsias, MD,2 James R. Ewing, PhD,2

Hamid Soltanian-Zadeh, PhD,3 Hassan Bagher-Ebadian, MS,2,3 Qingming Zhao, MS,2,3

Nancy Oja-Tebbe, BS,1 Suresh C. Patel, MD,3 and Michael Chopp, PhD2

Purpose: To identify early MRI characteristics of ischemicstroke that predict final infarct size three months poststroke.

Materials and Methods: Multiparametric MRI (multispinecho T2-weighted [T2W] imaging, T1-weighted [T1W] imag-ing, and diffusion-weighted imaging [DWI]) was performedacutely (�24 hours), subacutely (three to five days), and atthree months. MRI was processed using maps of apparentdiffusion coefficient (ADC), T2, and a self-organizing dataanalysis (ISODATA) technique. Analyses began with testingfor individual MRI parameter effects, followed by multiva-riable modeling with assessment of predictive ability (R2) onfinal infarct size.

Results: A total of 45 patients were studied, 15 of whomwere treated with tissue plasminogen activator (tPA) beforeacute MRI. The acute DWI and DWI-ISODATA mismatchlesion size, and the interactions of ADC, T2, and T2W im-aging lesion with tPA remained in the final multivariablemodel (R2 � 70%). A large acute DWI lesion or DWI �ISODATA lesion independently predicted increase in thefinal infract size, with predictive ability 68%. Predictiveability increased (R2 � 83%) when subacute MRI parame-ters were included along with acute DWI, DWI-ISODATAmismatch, and acute T2W image lesion size by tPA treat-ment interaction. Subacute DWI � acute DWI lesion sizepredicted an increased final infarct size (P � 0.01).

Conclusion: Acute-phase DWI and DWI-ISODATA mis-match strongly predict the final infarct size. An acute-to-subacute DWI lesion size change further increases the pre-dictive ability of the model.

Key Words: ischemic stroke; prediction; magnetic reso-nance imaging; generalized estimating equations (GEE);thrombolysisJ. Magn. Reson. Imaging 2005;21:495–502.© 2005 Wiley-Liss, Inc.

TREATMENT OF acute ischemic stroke can be effectiveas long as the ischemic tissue is still viable. If, earlyafter stroke onset, the ischemic tissue destined for in-farction could be distinguished from tissue capable ofrecovery, therapeutic decision-making in acute isch-emic stroke might be improved.

MRI has higher sensitivity than computed tomogra-phy (CT) for identifying ischemic damage in cerebraltissue (1–3). In stroke, diffusion-weighted imaging(DWI) obtained within the first several hours after ictusis capable of detecting ischemic injury of brain tissuewith a sensitivity over 90%, compared to less than 70%for CT (4). The acute ischemic lesion size determined byDWI is not only correlated with acute clinical neurolog-ical scores, but is also strongly correlated with the finalinfarct size measured by T2-weighted (T2W) imaging(4). The final infarct size and depth of apparent diffu-sion coefficient (ADC) reduction are also highly associ-ated when combined with the early DWI lesion size (5);the depth of ADC decline improves the ability of MRI topredict the infarct size (6–9). Unfortunately, ADCstrongly depends on the time elapsed since the onset ofsymptoms, thus complicating its predictive potential(10). For its part, the T2W image has long been recog-nized as a measurement of increase in total tissue watercontent and edema (11). In the acute and subacuteperiod after the onset of a stroke, the T2W image maynot provide a sensitive measurement of lesion size.However, an early T2W image lesion presence may bean important indicator of edema, which is strongly as-sociated with a larger infarct size at three months afterstroke (12).

As for treatment, thrombolysis using tissue plasmin-ogen activator (tPA) within three hours of symptom on-set is an effective treatment for acute stroke (13). Re-cent studies (14,15) have demonstrated that candidates

1Department of Biostatistics and Research Epidemiology, Henry FordHealth System, Detroit, Michigan, USA.2Department of Neurology, Henry Ford Health System, Detroit, Michi-gan, USA.3Department of Radiology, Henry Ford Health System, Detroit, Michi-gan, USA.Contract grant sponsor: National Institute of Neurological Disordersand Stroke (NINDS); Contract grant number: PO1 NS23393.*Address reprint requests to: M.L., Department of Biostatistics andResearch Epidemiology, Henry Ford Health Sciences Center, One FordPlace, 3E, Detroit MI 48202. E-mail: mlu1@hfhs.orgReceived June 25, 2004; Accepted January 31, 2005.DOI 10.1002/jmri.20313Published online in Wiley InterScience (www.interscience.wiley.com).

JOURNAL OF MAGNETIC RESONANCE IMAGING 21:495–502 (2005)

© 2005 Wiley-Liss, Inc. 495

for intravenous thrombolysis in the three- to six-hourtime frame can be identified based on the presence ofperfusion-weighted imaging (PWI)-DWI mismatch.However, DWI and PWI studies are often limited by thefeasibility of obtaining both DWI and PWI at the acutetime of stroke, and by the technical complications ofimage coregistration. In our experience, only 30% ofstroke patients have high quality quantitative perfusionimages at the acute stroke time (�12 hours of symptomonset); in the other study (16) only 14/94 (15%) ofpatients had a good quality PWI. Nevertheless, it wouldbe significant progress in the treatment of stroke if amethod could be found to identify a subgroup of strokepatients more than three hours postictus who wouldbenefit from thrombolysis.

The self-organizing data analysis (ISODATA) tech-nique is a method of processing and analysis of animage basis set utilizing multiple imaging parameters.This algorithm, developed and implemented at our lab-oratory in an experimental model of ischemic stroke, iscapable of characterizing the status of the ischemictissue vis-a-vis tissue viability (17,18), and has beenvalidated in human stroke (19). We have demonstratedthat the ISODATA lesion is highly correlated with thePWI lesion at the acute time of stroke and that thethree-month ISODATA infarct size is highly correlatedwith the T2W image infarct and the severity of the neu-rological deficit (20). Moreover, ISODATA can furthergrade the acute ischemic tissue status for ischemicdamage, which enhances the ability to predict the ulti-mate ischemic tissue recovery (21).

In this work, we sought to identify a set of MRI pa-rameters that would predict the final infarct size threemonths after stroke. Data were collected from ischemicstroke patients in the acute and subacute phases ofischemic stroke; the patients include both those treatedwith tPA and not treated with tPA, within three hours ofictus.

MATERIALS AND METHODS

Subjects

We recruited patients with acute neurological deficitcompatible with ischemic stroke, whether treated or notwith intravenous tPA, who could be studied with mul-tiparametric MRI within 24 hours of symptom onset.These patients were enrolled in the MRI Stroke Regis-try. We excluded patients with cerebral hemorrhageand/or history of prior significant stroke or other neu-rological deficit that would hamper accurate follow-upassessments. Informed consent was obtained beforeenrollment. The study was approved by the HumanRights Committee of the Henry Ford Health System.

Stroke onset was defined as the last time the patientwas known to be without neurological deficit. Patientswere evaluated by the study neurologist. Clinical neu-rological deficit was graded using the National Insti-tutes of Health Stroke Scale (NIHSS) at the time eachMRI study was obtained. MRI studies were performed atacute (�24 hours after stroke onset), subacute (three tofive days), and at outcome (three months) phases ofstroke. The ischemic stroke subtype was defined ac-

cording to the trial of Org 10172 (danaparoid) in acutestroke treatment (TOAST) classification (22) based onthe entire diagnostic evaluation the patient underwentduring hospitalization and follow-up.

For the present analysis, we included all patients whomet the following criteria: 1) acute-phase MRI com-pleted within 24 hours of stroke onset, 2) outcome-phase MRI completed at the three-month follow-up as-sessment, 3) diagnosis of supratentorial ischemicstroke, and 4) MRI data of sufficient quality to allowcomputer processing and analysis.

MRI Methodology

MRI studies were performed on a 1.5-T GE Signa MRscanner with echo-planar (EPI) capability (GE, Milwau-kee, WI, USA). Each MRI study consisted of axial mul-tispin echo T2W imaging, T1-weighted (T1W) imagingpre- and postcontrast gadolinium-diethylenetriaminepentaacetic acid (Gd-DTPA), and DWI with slice thick-ness of 6 mm. For T1W images and T2W images, thefield of view (FOV) was 200 � 200 mm, matrix size256 � 192; for DWI, FOV was 240 � 240 mm, matrixsize 128 � 128. Additional parameters for each studywere: 1) T1W images: TR/TE � 600/14 msec; 2) T2Wimages: TR/TE � 2800/30, 60, 90, 120 msec; 3) axialDWI was performed using an echo-planar sequence:TR/TE � 10,000/101 msec, b-values � 1000, 600,300, 0 seconds/mm2, number of excitations (NEX) � 1.Throughout the acquisition of the images, the patientwas monitored. The MRI protocol contains a broaderrange of images for exploratory purposes, with totalacquisition time of about 45 minutes.

MRI Selection and ISODATA Analysis

The study neuroradiologist and the neurologist selectedat most three slices (images) that would cover most, orall, of the ischemic lesion evident on the acute DWI;usually these were adjacent slices. Given the slicethickness (6 mm), one or two slices might suffice. Formultiple lesions at different locations, slices were se-lected to cover most of, or all of, the largest infarction.Figures 1, 2 and 3 illustrate multiparametric imagingof three patients with ischemic stroke of various sizes,one treated with tPA, and with the measurements oflesion on acute phase DWI (left), acute phase ISODATAcomposite (middle), and three-month T2W image(right).

MR image processing included image coregistrationand warping (23), and ISODATA multiparametric seg-mentation (17–20,24). ISODATA included one proton-density-weighted image (TE: 30 msec), one T2W image(TE: 90 msec), one T1W image, and two DWI images(b-values: 600, 1000 seconds/mm2). ISODATA is anunsupervised segmentation method related to the K-means algorithm, with additional splitting and mergingsteps that allow for the adjustment of cluster centers.The ISODATA process has been described in a previouspublication (21).

After convergence, ISODATA clusters were standard-ized to the characteristics of two known clusters: nor-mal white matter and cerebrospinal fluid (CSF). Signa-

496 Lu et al.

ture 1 (represented as blue color on the color coded bar;Fig. 1) was assigned to normal white matter and signa-ture 12 (represented as white color; Fig. 1) to CSF.Thus, ischemic tissue clusters were identified with asignature number, in a range of 2–11 (see the colorcoded bar, between green and red codes; Fig. 1). Theacute ISODATA signature value per slice was calculatedadjusting for the lesion size among the regions based onthe following formula:

1Total ISODATA area �

m � 1

M

ISODATA Signature valuem � ISODATA aream , (1)

where M is the number of regions (signatures) per slice,and m � 1,2, . . ., M.

To determine the value of T2 and ADC in the lesionarea, the ISODATA region corresponding to the lesionwas overlaid on the ADC and T2 map, and the meanvalue of the parameter in the infarct region of interest(ROI) was determined. The ROI was also reflected onto

the contralateral normal area in order to obtain meanratios (with contralateral values) of ADC and T2 (rADCand rT2, respectively). For each of the different MRIweightings, we calculated the lesion area per MRI slice,as the sum of each lesion pixel area, multiplied by thenumber of lesion pixels. We report lesion size as mean(�SD) of the lesion area.

Data Collection

MRI data were collected from stroke patients in theacute, subacute (when possible), and final (threemonths) phases after stroke. Our focus was to use theacute or the subacute MRI to predict the final infarctsize. The final infarct size was measured per MRI sliceusing the three-month T2W image.

We included MRI measurements of ISODATA signa-ture value, ADC/rADC, T2 value/rT2, and the ISO-DATA, DWI, and T2W image lesion sizes at the acutetime. Given the impact of tPA on MRI parameters (21)obtained posttreatment, and its benefit on functionaland lesion recovery (12,13), we included the variable oftPA treatment in the analysis. We also included the time

Figure 2. Acute-phase DWI (left), ISODATA (middle), and outcome-phase T2W (right) images, demonstrating a moderate size leftmesial temporal lobe infarct. The bar indicates the signature numbers and color-coding of the ISODATA-defined ischemic lesionsignatures.

Figure 1. Acute-phase DWI (Left), ISODATA (Middle) and outcome-phase T2W (Right) images, demonstrating a large righthemispheric infarct and an old left hemispheric infarct. Please note the acute phase ISODATA-DWI mismatch and the size of theoutcome-phase T2W ischemic lesion as well as the inhomogeneity of the acute-phase ISODATA lesion. The bar indicates thesignature numbers and color-coding of the ISODATA-defined ischemic lesion signatures.

Early MRI Predicts Final Infarct Size 497

of stroke onset to time of acute MRI as a potentialpredictor for the final infarct, given that the MRI param-eters may be dependent on time.

The ISODATA-DWI mismatch was calculated at theacute phase, defined as the lesion size difference be-tween ISODATA lesion and DWI lesion. In addition, wecalculated the changes and relative changes in MRIparameters at subacute from the acute (e.g., subacuteMRI – acute MRI) on a subgroup of patients with MRIobtained at both time points, yielding a total of 23possible MRI parameters of interest. Of the 23 MRIparameters, 10, including the time of stroke to MRI,were collected at the acute phase; 13 MRI changes werecollected at the subacute phase.

Statistical Analysis Approach

Since infarct size is not normally distributed, a similardata transformation, proposed in the National Instituteof Neurological Disorders and Stroke (NINDS) rt-PAStroke Study Group for CT infarct (12) was performed,thus producing a distribution more approximately nor-mal. To avoid a possible variable confounding (25), orthe treatment effect on acute MRI parameters, we testedthe MRI parameter difference between the tPA-treatedgroup and non-tPA-treated groups at the 0.05 level. Wefirst tested the individual MRI parameter effect (e.g.,ADC variable alone) or the effect of the MRI parameterby tPA treatment interaction (e.g., ADC with the treat-ment of tPA interaction) on the final infarct size, fol-lowed by multivariable modeling for independent pre-dictive abilities. An MRI parameter with the tPAtreatment interaction indicates that the predictive abil-ity of the MRI parameter on the infarct size is dependenton whether or not patients received tPA treatment. Weused the generalized estimation equation (GEE) (26,27),assuming independent subject (patients) and commoncorrelation among slices per subject. All analyses wereadjusted for the acute DWI lesion size by including theacute DWI lesion size in the regression model as acovariate.

Variables with an individual effect, or an interactionbetween the variable and the treatment of tPA at the0.15 level, were included in the first multivariable

model. To avoid redundancy, the highly correlated vari-ables (r � 0.70), such as ADC and rADC, had to beevaluated separately. Two-way or three-way effects(MRI parameter interactions or MRI parameters andtPA treatment interactions) were considered only ifthere was a one-way effect at the 0.05 level or a two-wayMRI variable interaction at the 0.10 level, respectively.

A final multivariable model included individual MRIparameters with P-value � 0.05 or MRI parameter in-teractions with P-value � 0.10. A positive coefficient ofthe MRI parameter indicated an increase in final infarctsize when the value of the MRI parameter is increased.A negative coefficient, in contrast, indicates a reductionin the final infarct size when the MRI parameter value isincreased. We assessed the predictive ability by mea-suring the squared multiple correlation coefficient (R2),which reflects the degree of variation in the outcomeexplained by the predictor variables, and is related tothe correlation coefficient �r � �R2�. R2 ranges from 0 to1.0, with a value close to 1 indicating that the modelexplains most of the variation in the outcome.

Due to the paucity of MRI data collection at the sub-acute phase, we conducted two different multivariablemodels. The first model considered acute MRI parame-ters only (the acute multivariable model), and the sec-ond model considered both MRI at the acute and sub-acute time.

Note that acute or subacute MRI parameters maypredict a large reduction in infarct size in a patient witha single or multiple MRI slices, but it may not reflect thewhole picture of this patient’s recovery, because of theMRI slice selection.

RESULTS

A total of 45 ischemic stroke patients, mean age 63(�14) years, were studied. A total of 21 patients (46%)were males and 34 (67%) were African-Americans. Themedian acute NIHSS was 7 (range: 2–26). A total of 15patients (31%) had received treatment with intravenoustPA within three hours of stroke onset, and prior to theacute-phase MRI. The mean time interval from strokeonset to MRI study was 10.7 (�5.6) hours. Populations

Figure 3. Acute-phase DWI (left), ISODATA (middle), and outcome-phase T2W (right) images, demonstrating a small rightinternal capsule lacunar infarct in a patient who received treatment with tPA prior to MRI. The bar indicates the signaturenumbers and color-coding of the ISODATA-defined ischemic lesion signatures.

498 Lu et al.

of the ischemic stroke subtypes were: cardiogenic em-bolism (N � 17; 38%), large vessel atherosclerosis (N �3; 7%), lacunar stroke (N � 12; 27%), undeterminedetiology (N � 10; 22%), other determined etiology (N � 3;7%). A total of 123 acute MRI slices were included fromthese 45 patients. A total of 32 of the 45 patients, 11treated with tPA, had both acute- and subacute-phaseMRI data on 88 slices.

MRI parameters were first compared between the tPAtreated group and the untreated group. The only MRImeasures which differed between tPA treated and un-treated groups were both the change and relativechange in T2 value from the acute to the subacute withP-values of 0.02 and 0.04, respectively.

Effect of Individual MRI Parameters on FinalInfarct Size

As noted, MRI parameters were first tested for theirindividual predictive effect with regard to the final in-farct size, followed by multivariable modeling for inde-pendent predictive abilities. At the acute phase, theMRI variables, DWI and ISODATA lesion size, and thestroke onset–MRI time interval had individual effectsfor prediction of the final infarct size, with P-values �0.15. In addition, ADC and rADC values, T2 and rT2values, T2W image lesion size and ISODATA-DWI lesionmismatch also had predictive effects for the final infarctsize (P � 0.15); however, these effects were dependenton the treatment of tPA presence and absence (an MRIby tPA interaction). There was no ISODATA weightedsignature effect predictive of the final infarct size (Table1a).

At the subacute phase, only two MRI-related param-eters, the relative change of DWI lesion size (lesion size

difference at subacute from the acute divided by acutelesion size), and the change in T2 value (the differenceat subacute from the acute phase) had effects in pre-diction of final infarct size with P-values � 0.15 (Table1b). No MRI by treatment tPA interaction was observedat the P � 0.15 level.

Multivariable Modeling Results

Using the backward model selection, and utilizing onlythe acute-phase MRI data (123 slices from 45 patients),the DWI lesion size (P � 0.01), the DWI-ISODATA mis-match (P � 0.01), as well as the interactions of rADC bytPA treatment (P � 0.04), rT2 by tPA treatment (P �0.03), and T2W image lesion size by tPA treatment (P �0.01) remained in the final acute-phase multivariablemodel. These findings mean that a larger lesion size onacute-phase DWI predicted a large final infarct [with acoefficient and (SD) of 0.90 (0.10)]. A DWI lesion sizelarger than the ISODATA lesion size predicted a de-crease in the size of the final infarct, while a DWI lesionsize smaller than the ISODATA lesion size predicted anincrease in the size of the final infarct [coefficient –0.08(0.03)]. The predictive ability of the acute-phase DWIand ISODATA-DWI lesion size mismatch on the finalinfarct size were independent of tPA treatment. How-ever, the predictive abilities of the acute-phase rADC,

Table 1aTest of Individual Association Between Acute-phase MRIParameters and Final Infarct Size

Variables of interest

P-valuea forindividual MRI

parametereffect

P-valuea for MRIparameter by tPA

Interaction

Tissue signature 0.70 0.19ADC value 0.83 0.13b

rADC 0.72 0.05b

T2 value 0.92 0.02b

rT2 0.86 0.03b

ISODATA lesionsizec

�0.01b 0.37

DWI lesion sizee �0.01b 0.99T2W image lesion

sizec0.88 0.07b

DWI-ISODATAmismatch sizec

0.01 0.05b

Stroke onset-MRItime interval

0.12b 0.44

Treatment with tPA 0.71 NAd

aAnalysis was based on transformed data using GEE adjusting forthe acute DWI lesion and correlated MRI slices.bvariables, or variable by tPA interaction, included in the first stepmultivariable model.cmeasured by area per slice.dNA � not applicable.

Table 1bTest of Individual Association Between Subacute-phase MRIParameters and Final Infarct Size*

Variables of interest

P-valuea forindividual MRI

parametereffect

P-valuea forMRI parameter

by tPAInteraction

Change of ISODATA lesionsizeb

0.64 0.84

Relative change of ISODATAlesion sizeb

0.33 0.15

Change of DWI lesion sizeb 0.35 0.76Relative change of DWI

lesion sizeb�0.01c 0.89

Change of T2W image lesionsizeb

0.20 0.26

Relative change of T2Wimage lesion sizeb

Change of ADC value 0.89 0.86Relative change of ADC

value0.94 0.68

Change of rADC value 0.84 0.43Relative change of rADC

value0.59 0.66

Change of T2 value 0.12c 0.85Relative change of T2 value 0.18 0.77Change of rT2 value 0.59 0.66Relative change of rT2 value 0.65 0.54

*Change means the change in lesion size from acute to subacutephase. Relative change means the relative change in lesion sizefrom acute to subacute phase which is defined as: (acute-subacute)phase lesion size/acute phase lesion size.aAnalysis was based on transformed data using GEE adjusting forthe acute DWI lesion and correlated MRI slices.bMeasured by area per slicecVariables included in the first step multivariable model.

Early MRI Predicts Final Infarct Size 499

rT2, or the T2W image lesion size on the final infarctsize were dependent on whether treatment with tPA wasgiven to the patient prior to the acute-phase MRI study.In patients who received tPA prior to the acute-phaseMRI studies, a higher rT2 value [coefficient10.43(6.73)], or a smaller T2W Image lesion size [coef-ficient –0.21 (0.07)] predicted a larger final infarct size(P � 0.02, based on subgroup analysis). On the otherhand, for patients who had not received tPA, a lower rT2value [coefficient –4.40 (1.99)] predicted a larger finalinfarct size (P � 0.02), while no effect of the acute-phaseT2W image lesion size was detected (P � 0.35). Similareffects were observed with the rADC value, but theeffects were marginal (P � 0.11, based on subgroupanalysis). No other two-way or three-way interactionswere detected. The predictive ability (R2) of the finalmodel was 70% (Table 2), and 68% considering both theacute DWI and the acute DWI-ISODATA mismatch in amultivariable model. As a result, including rADC, rT2,T2W image lesion and the MRI parameters-by-treat-ment interactions added only 2% more predictive abilityas to the size of the final infarct.

Using the same model analysis approach, we con-structed the final multivariable model based on allacute and subacute phase data (88 slices from 32 pa-tients with both MRI in the acute and subacute phases).The acute DWI lesion size (P � 0.01), the DWI-ISODATAmismatch (P � 0.01), the T2W image lesion size by tPAinteraction (P � 0.08), as well as the relative change inthe DWI lesion size from the acute to subacute phase(P � 0.01) remained in the final model (Table 3). Thepredictive ability of the final model (R2) was 83%. Thepredictive abilities of the acute-phase parameters (DWI

lesion size, DWI-ISODATA mismatch, and T2W imagelesion size) on the final infarct size were similar com-pared to their predictive abilities in the acute-phasemultivariable model. In addition to these acute-phaseMRI predictors, the combined acute-phase and sub-acute-phase MRI parameters predicted the final infarctsize with a higher predictive ability of 83% (increasedfrom 70%). No other interactions were detected at the0.10 level.

DISCUSSION

This study identifies those early MRI characteristicsthat can predict the final infarct size (e.g., at threemonths after stroke). A total of 23 parameters takenfrom MRI images obtained at the acute and subacutephases of stroke were assessed as possible predictorsfor the infarct size. With a rigorous statistical analysisapproach, we developed two predictive models, one (theacute model) including only parameters from acutephase of stroke and one (the final model) incorporatingparameters from both the acute and subacute phasesof stroke. With these analyses we replicated previousstudies’ findings that the final infarct size depends onthe acute DWI lesion, and also demonstrated that itadditionally depends on the acute DWI-ISODATA lesionmismatch and the change in the DWI lesion size fromthe acute to the subacute phase. We also demonstratedthat the final infarct size also depends on the acutephase T2W image lesion size, as well as the acute phaserADC and rT2 values. However, the predictive abilitiesof the latter parameters are influenced by whethertreatment with tPA was applied prior to acute-phaseimage acquisition. Patients who received tPA treatmentand had a lower rADC value developed a larger infarctsize at three months. Also, patients who received tPAtreatment and had a larger T2W image lesion wouldhave a smaller infarct, compared to the tPA-treatedpatients with a smaller T2W image lesion. We also ob-

Table 2Acute Multivariable Model for Prediction of Final Infarct Size

Variables in the final modelCoefficientsa

(SD)P valueb

DWI lesion size 0.90 (0.10) �0.01tPA treatment NA NADWI-ISODATA mismatch

size�0.08 (0.03) 0.01

Interaction of T2w imagelesion size by tPAtreatment

�0.01

With tPA presence �0.21 (0.07)With tPA absence 0.07 (0.07)

Interaction of rADC by tPAtreatment

0.04

With t-PA presence �13.70 (7.71)With t-PA absence 2.14 (1.51)

Interaction of rT2 with tPAtreatment

0.03

With t-PA presence 10.43 (6.73)With t-PA absence �4.40 (1.99)

Goodness-of-fit (R2) �0.70

aA positive coefficient indicates an increase in the final infarct sizewhen the value of the MRI parameter is increased. In contrast, anegative coefficient indicates a reduction in the final infarct size whenthe value of the MRI parameter is increased.bAnalysis was based on transformed data using GEE adjusting forthe acute DWI lesion and correlated MRI slices.NA � not applicable.

Table 3Final Multivariable Model, Utilizing Acute and Subacute-phaseParameters, for the Prediction of Final Infarct Size

Variables remaining in the final modelCoefficientsa

(SD)P valueb

Acute-phase DWI lesion size 1.24 (0.08) �0.01tPA treatment NA NAAcute-phase DWI-ISODATA

mismatch size�.10 (0.03) �0.01

Acute-phase T2w image lesion sizeby tPA interaction

0.08

With tPA presence �0.19 (0.07)With tPA absence �0.03 (0.07)

Relative change of DWI lesion sizec 1.08 (0.33) 0.01Goodness-of-fit (R2) � 0.83

aA positive coefficient indicates an increase in the final infarct sizewhen the value of the MRI parameter is increased. In contrast, anegative coefficient indicates a reduction in the final infarct size whenthe value of the MRI parameter is increased.bAnalysis was based on transformed data using GEE adjusting forthe acute DWI lesion and correlated MRI slices.cRelative change of lesion size was defined as: (subacute lesionsize-acute lesion size)/acute lesion size; NA � not applicable.

500 Lu et al.

served that a lower acute-phase rT2 value leads to alarger infarct at three months in patients who had notreceived tPA prior to acute MRI study.

In contrast to the acute rADC and rT2 values or theT2W image lesion size, the predictive abilities of acuteDWI lesion and the DWI-ISODATA mismatch are robustand independent of tPA treatment. The predictive abil-ity is 68% considering both the acute DWI lesion sizeand the acute DWI-ISODATA lesion mismatch. Only 2%predictive ability was added by including rADC, rT2,T2W image lesion, and the MRI parameters-by-tPAtreatment interactions that complicated the model (byadding seven additional variables to the model), withlittle improvement on predictive ability. The values ofADC or T2 have the same effects, compared to rADC orrT2, but had less predictive ability (67%) when rADCand rT2 were replaced with ADC and T2 values in thefinal acute multivariable model (Table 2). Adding therelative DWI lesion size change from acute to subacutephase increased the predictive ability to 83% (Table 3).Without the acute T2W image lesion size by tPA treat-ment interaction, the predictive ability still remained80%. Note that there is no direct comparison betweenthe acute and subacute multivariable models becauseof differences in numbers of MRI slices and in the pa-tient population. However, considering the subacutepatient population (88 slices from 32 stroke patients),the predictive abilities are only 60% without the inclu-sion of the acute-to-subacute change in DWI. There-fore, the change of DWI is an important predictor for theeventual infarct size.

The acute DWI lesion size was a strong predictor offinal infarct size, a finding that is in agreement withthose of other studies (28,29), which demonstrated thatthe DWI lesion volume in the acute phase of strokepredicts the infarct volume as shown by the T2W imageinfarct volume at outcome. In addition, the relativechange in the DWI lesion size from acute to subacutephase was a strong independent predictor of final in-farct size. This likely reflects that when there is DWIlesion expansion, the second measurement will matchthe final infarct size more closely.

The acute DWI-ISODATA mismatch lesion size wasalso a strong predictor of the final infarct size. Severalauthors (15,30) have used a different concept, PWI-DWImismatch, to indicate that the mismatch volume likelyrepresents viable tissue which, if left untreated, mayevolve to infarction. In untreated patients who demon-strate PWI-DWI mismatch in the acute phase, the DWIlesion will very likely expand over a time period of sev-eral hours to days. One weakness of the PWI in acuteischemia is that the PWI lesion volume probably in-cludes both critically hypoperfused tissue (ischemia)and hypoperfused (but not at a critical level—benignoligemia) tissue, which will never evolve into infarction.The ISODATA-DWI, -T2W, and -T1W composite imagescan identify subtle abnormalities in ischemic tissue notyet seen by any one of the individual MRI techniques(DWI, T2W images, or T1W images). We believe thishappens because there are contributions to image con-trast from various factors, such as perfusion, proteoly-sis, and changes in deoxyhemoglobin concentration,which are not visible by any one modality, but have a

cumulative effect on tissue signature when analyzedwith ISODATA. In this sense, an acute-phase mismatchbetween ISODATA and DWI may be an indication thatthe ISODATA lesion is identifying the portion of theperfusion lesion that undergoes some degree of isch-emic tissue injury or the MRI signs of a perfusion defi-cit, and, therefore, the DWI lesion has a potential togrow. The DWI-ISODATA mismatch may thus serve as asurrogate of the DWI-PWI mismatch with a better fea-sibility, because ISODATA only requires T1W images,T2W images, and DWI.

This study has not made a direct correlation betweenischemic lesion size and clinical neurological or func-tional scales, since its primary focus was the predictionof ischemic lesion size and not clinical correlates. Sev-eral previous studies (28,29,31) have demonstratedgood correlations of acute DWI lesion volumes withacute and chronic NIHSS scores and the correlation ofchronic T2W image infarct size and NIHSS scores. Ourprevious work has also demonstrated strong correla-tions between ISODATA lesion size and NIHSS scores inboth the acute and outcome phases, as well as betweenISODATA and DWI lesion size in the acute phase andISODATA and T2W image lesion size in the outcomephase (20). This suggests, albeit indirectly, that thefindings of the present study may be translated intomeaningful clinical correlations and outcomes.

ISODATA analysis can be used for any type and loca-tion of ischemic stroke. Previous work has indicatedthat it can be applied for brainstem stroke with gooddefinition of the ischemic lesion in that location (28).ISODATA is reproducible because the algorithm is un-supervised, as are all other processing methods used inthis study. The main limitation for the application of theISODATA in the emergency management of stroke maylie in the time needed to complete the analysis; it takesabout two hours to process ISODATA on a single slice,with the majority of this time (90%) spent for accuratecoregistration of the multiple images. With advances inthe accuracy of image acquisition, computer technol-ogy, and optimization of the software, it may be possibleto perform the analysis in a few minutes and thusaccommodate the requirements of the emergency clin-ical practice.

No effect of tPA on the final infarct size was found,possibly for the following reasons: 1) The acute phasemultiparametric imaging was obtained after treatmentwith tPA was already given. It is possible that differentfindings would appear if tPA was given before the acutephase study. 2) The study was designed to explore therelation of MRI parameters measured at acute or sub-acute times to the final infarct size. 3) The study wasnot designed to have sufficient power to detect the tPAeffect. It should be noted that only a marginal tPA effectwas observed on three-month infarct size on CT scansin the NINDS rtPA stroke study based on a total of 624patients (12). The medians and interquartiles (25% and75%) of the transformed three-month infarction were9.42 (5.48, 18.33) in the tPA-treated group and 10.14(5.74, 11.90) in the non-tPA-treated group, withP-value � 0.71 (24). In addition, three patients who hadhemorrhagic transformation after the initial strokes

Early MRI Predicts Final Infarct Size 501

were included in our study and only one had tPA treat-ment prior to the hemorrhage.

Of additional interest are the interactions betweenpreceding treatment with tPA acute T2W image lesionsize, rADC, and rT2 values. Patients who received tPAtreatment and had a lower rADC value developed alarger infarct size at three months. This probably indi-cates that the ischemic tissue injury was more severeand accompanied by inadequate reperfusion, as thelatter would be expected to produce higher ADC values.The findings from T2W imaging, i.e., that a larger acuteT2W image lesion would predict a smaller final infarctand that a lower acute rT2 value leads to a larger finalinfarct in untreated patients, are likely to have a similarinterpretation, indicating more complete and inade-quate reperfusion, respectively. Reperfusion, accompa-nied by edema, could be expected to produce a higherrT2 value and a larger T2W image lesion post-tPA treat-ment.

At present, response of the acute stroke to treatmentwith tPA can be judged at the three-month time pointbased on the clinical recovery. There have been no goodmarkers developed to distinguish responders from non-responders in the early phase following treatment. Thisissue is expected to become increasingly important assalvage treatments are developed for patients deemedto be nonresponders. The model presented here has thepotential to solve this problem, and allow early defini-tion of the potential nonresponders, thus allowing sal-vage treatments to be offered.

In conclusion, acute DWI and the acute DWI-ISO-DATA mismatch are the predictors of the final infarctsize at three months after stroke onset and are inde-pendent of preceding treatment with tPA. The acute–subacute DWI lesion size change increases the predic-tive ability of the model with regard to the final infarctsize. These findings could be used for the rational de-sign of image-based acute treatment trials.

ACKNOWLEDGMENT

We thank Lula Adams for editing.

REFERENCES1. Saur D, Kucinski T, Grzyska U, et al. Sensitivity and interrater

agreement of CT and diffusion-weighted MR imaging in hyperacutestroke. AJNR Am J Neuroradiol 2003;24:878–885.

2. Fiebach JB, Schellinger PD, Jansen O, et al. CT and diffusion-weighted MR imaging in randomized order: diffusion-weighted im-aging results in higher accuracy and lower interrater variability inthe diagnosis of hyperacute ischemic stroke. Stroke 2002;33:2206–2210.

3. Barber PA, Darby DG, Desmond PM, et al. Identification of majorischemic change. Diffusion-weighted imaging versus computed to-mography. Stroke 1999;30:2059–2065.

4. Lansberg MG, Albers GW, Beaulieu C, et al. Comparison of diffu-sion-weighted MRI and CT in acute stroke. Neurology 2000;54:1557–1561.

5. Watanabe Y, Mitomo M, Tokuda Y, et al. Eclamptic encephalopa-thy: MRI, including diffusion-weighted images. Neuroradiology2002;44:981–985.

6. Oppenheim C, Samson Y, Manai R, et al. Prediction of malignantmiddle cerebral artery infarction by diffusion-weighted imaging.Stroke 2000;31:2175–2181.

7. Parsons MW, Li T, Barber PA, et al. Combined (1)H MR spectroscopyand diffusion-weighted MRI improves the prediction of stroke out-come. Neurology 2000;55:498–505.

8. Arenillas JF, Rovira A, Molina CA, et al. Prediction of early neuro-logical deterioration using diffusion- and perfusion-weighted imag-ing in hyperacute middle cerebral artery ischemic stroke. Stroke2002;33:2197–2203.

9. Hermier M, Nighoghossian N, Adeleine P, et al. Early magneticresonance imaging prediction of arterial recanalization and lateinfarct volume in acute carotid artery stroke. J Cereb Blood FlowMetab 2003;23:240–248.

10. Ahlhelm F, Schneider G, Backens M, et al. Time course of theapparent diffusion coefficient after cerebral infarction. Eur Radiol2002;12:2322–2329.

11. DeWitt LD, Buonanno FS, Kistler JP, et al. Nuclear magnetic res-onance imaging in evaluation of clinical stroke syndromes AnnNeurol 1984;16:535–545.

12. NINDS. Effect of intravenous recombinant tissue plasminogen activa-tor on ischemic stroke lesion size measured by computed tomogra-phy. The National Institute of Neurological Disorders and Stroke(NINDS) rt-PA Stroke Study Group. Stroke 2000;31:2912–2919.

13. National Institute of Neurological Disorders and Stroke rt-PAStroke Study Group. Tissue plasminogen activator for acute isch-emic stroke. N Engl J Med 1995;333:1581–1587.

14. Marks MP, Tong DC, Beaulieu C, et al. Evaluation of early reper-fusion and i.v. tPA therapy using diffusion- and perfusion-weightedMRI. Neurology 1999;52:1792–1798.

15. Parsons MW, Barber PA, Chalk J, et al. Diffusion- and perfusion-weighted MRI response to thrombolysis in stroke. Ann Neurol2002;51:28–37.

16. Wu O, Koroshetz WJ, Ostergaard L, et al. Predicting tissue outcomein acute human cerebral ischemia using combined diffusion- andperfusion-weighted MR imaging. Stroke 2001;32:933–942.

17. Soltanian-Zadeh H, Pasnoor M, Hammoud R, et al. MRI tissuecharacterization of experimental cerebral ischemia in rat. J MagnReson Imaging 2003;17:398–409.

18. Jacobs MA, Knight RA, Soltanian-Zadeh H, et al. Unsupervised seg-mentation of multiparameter MRI in experimental cerebral ischemiawith comparison to T2, diffusion, and ADC MRI parameters and his-topathological validation. J Magn Reson Imaging 2000;11:425–437.

19. Jacobs MA, Mitsias P, Soltanian-Zadeh H, et al. MultiparametricMRI tissue characterization in clinical stroke with correlation toclinical outcome: part 2. Stroke 2001;32:950–957.

20. Mitsias PD, Jacobs MA, Hammoud R, et al. Multiparametric MRIISODATA ischemic lesion analysis: correlation with the clinicalneurological deficit and single-parameter MRI techniques. Stroke2002;33:2839–2844.

21. Mitsias PD, Ewing JR, Lu M, et al. Multiparametric iterative self-organizing MR imaging data analysis technique for assessment oftissue viability in acute cerebral ischemia. AJNR AJNR Am J Neu-roradiol 2004;25:1499–1508.

22. Adams HP, Jr., Bendixen BH, Kappelle LJ, et al. Classification ofsubtype of acute ischemic stroke. Definitions for use in a multi-center clinical trial. TOAST. Trial of Org 10172 in Acute StrokeTreatment. Stroke 1993;24:35–41.

23. Ghanei A, Soltanian-Zadeh H, Jacobs MA, et al. Boundary-basedwarping of brain MR images. J Magn Reson Imaging 2000;12:417–429.

24. Jacobs MA, Zhang ZG, Knight RA, et al. A model for multiparamet-ric MRI tissue characterization in experimental cerebral ischemiawith histological validation in rat: part 1. Stroke 2001;32:943–949.

25. McNeil D. Epidemiological research methods, 1st edition. NewYork: John Wiley & Sons; 1996. p 125–140.

26. Liang K, Zeger SL. Longitudinal data analysis using generalizedlinear models. Biometrika 1986;72:13–22.

27. Zeger SL, Liang K. Longitudinal data analysis for discrete andcontinuous outcomes. Biometrics 1986;42:121–130.

28. Barber PA, Darby DG, Desmond PM, et al. Prediction of strokeoutcome with echoplanar perfusion- and diffusion-weighted MRI.Neurology 1998;51:418–426.

29. Warach S, Dashe JF, Edelman RR. Clinical outcome in ischemicstroke predicted by early diffusion-weighted and perfusion mag-netic resonance imaging: a preliminary analysis. J Cereb BloodFlow Metab 1996;16:53–59.

30. Baird AE, Benfield A, Schlaug G, et al. Enlargement of humancerebral ischemic lesion volumes measured by diffusion-weightedmagnetic resonance imaging. Ann Neurol 1997;41:581–589.

31. Lovblad KO, Baird AE, Schlaug G, et al.. Ischemic lesion volumes inacute stroke by diffusion-weighted magnetic resonance imagingcorrelate with clinical outcome. Ann Neurol 1997;42:164–170.

502 Lu et al.