Predicting Final Infarct Size Using Acute and Subacute Multiparametric MRI Measurements in Patients With Ischemic Stroke

Transcription

1 JOURNAL OF MAGNETIC RESONANCE IMAGING 21: (2005) Original Research Predicting Final Infarct Size Using Acute and Subacute Multiparametric MRI Measurements in Patients 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, PhD 2 Purpose: To identify early MRI characteristics of ischemic stroke that predict final infarct size three months poststroke. Materials and Methods: Multiparametric MRI (multispin echo T2-weighted [T2W] imaging, T1-weighted [T1W] imaging, and diffusion-weighted imaging [DWI]) was performed acutely ( 24 hours), subacutely (three to five days), and at three months. MRI was processed using maps of apparent diffusion coefficient (ADC), T2, and a self-organizing data analysis (ISODATA) technique. Analyses began with testing for individual MRI parameter effects, followed by multivariable modeling with assessment of predictive ability (R 2 )on final infarct size. Results: A total of 45 patients were studied, 15 of whom were treated with tissue plasminogen activator (tpa) before acute MRI. The acute DWI and DWI-ISODATA mismatch lesion size, and the interactions of ADC, T2, and T2W imaging lesion with tpa remained in the final multivariable model (R 2 70%). A large acute DWI lesion or DWI ISODATA lesion independently predicted increase in the final infract size, with predictive ability 68%. Predictive ability increased (R 2 83%) when subacute MRI parameters were included along with acute DWI, DWI-ISODATA mismatch, and acute T2W image lesion size by tpa treatment interaction. Subacute DWI acute DWI lesion size predicted an increased final infarct size (P 0.01). Conclusion: Acute-phase DWI and DWI-ISODATA mismatch strongly predict the final infarct size. An acute-tosubacute DWI lesion size change further increases the predictive ability of the model. 1 Department of Biostatistics and Research Epidemiology, Henry Ford Health System, Detroit, Michigan, USA. 2 Department of Neurology, Henry Ford Health System, Detroit, Michigan, USA. 3 Department of Radiology, Henry Ford Health System, Detroit, Michigan, USA. Contract grant sponsor: National Institute of Neurological Disorders and Stroke (NINDS); Contract grant number: PO1 NS *Address reprint requests to: M.L., Department of Biostatistics and Research Epidemiology, Henry Ford Health Sciences Center, One Ford Place, 3E, Detroit MI mlu1@hfhs.org Received June 25, 2004; Accepted January 31, DOI /jmri Published online in Wiley InterScience ( Key Words: ischemic stroke; prediction; magnetic resonance imaging; generalized estimating equations (GEE); thrombolysis J. Magn. Reson. Imaging 2005;21: Wiley-Liss, Inc. TREATMENT OF acute ischemic stroke can be effective as long as the ischemic tissue is still viable. If, early after stroke onset, the ischemic tissue destined for infarction could be distinguished from tissue capable of recovery, therapeutic decision-making in acute ischemic stroke might be improved. MRI has higher sensitivity than computed tomography (CT) for identifying ischemic damage in cerebral tissue (1 3). In stroke, diffusion-weighted imaging (DWI) obtained within the first several hours after ictus is capable of detecting ischemic injury of brain tissue with a sensitivity over 90%, compared to less than 70% for CT (4). The acute ischemic lesion size determined by DWI is not only correlated with acute clinical neurological scores, but is also strongly correlated with the final infarct size measured by T2-weighted (T2W) imaging (4). The final infarct size and depth of apparent diffusion coefficient (ADC) reduction are also highly associated when combined with the early DWI lesion size (5); the depth of ADC decline improves the ability of MRI to predict the infarct size (6 9). Unfortunately, ADC strongly depends on the time elapsed since the onset of symptoms, thus complicating its predictive potential (10). For its part, the T2W image has long been recognized as a measurement of increase in total tissue water content and edema (11). In the acute and subacute period after the onset of a stroke, the T2W image may not provide a sensitive measurement of lesion size. However, an early T2W image lesion presence may be an important indicator of edema, which is strongly associated with a larger infarct size at three months after stroke (12). As for treatment, thrombolysis using tissue plasminogen activator (tpa) within three hours of symptom onset is an effective treatment for acute stroke (13). Recent studies (14,15) have demonstrated that candidates 2005 Wiley-Liss, Inc. 495

2 496 Lu et al. for intravenous thrombolysis in the three- to six-hour time frame can be identified based on the presence of perfusion-weighted imaging (PWI)-DWI mismatch. However, DWI and PWI studies are often limited by the feasibility of obtaining both DWI and PWI at the acute time of stroke, and by the technical complications of image coregistration. In our experience, only 30% of stroke patients have high quality quantitative perfusion images at the acute stroke time ( 12 hours of symptom onset); in the other study (16) only 14/94 (15%) of patients had a good quality PWI. Nevertheless, it would be significant progress in the treatment of stroke if a method could be found to identify a subgroup of stroke patients more than three hours postictus who would benefit from thrombolysis. The self-organizing data analysis (ISODATA) technique is a method of processing and analysis of an image basis set utilizing multiple imaging parameters. This algorithm, developed and implemented at our laboratory in an experimental model of ischemic stroke, is capable of characterizing the status of the ischemic tissue vis-à-vis tissue viability (17,18), and has been validated in human stroke (19). We have demonstrated that the ISODATA lesion is highly correlated with the PWI lesion at the acute time of stroke and that the three-month ISODATA infarct size is highly correlated with the T2W image infarct and the severity of the neurological deficit (20). Moreover, ISODATA can further grade the acute ischemic tissue status for ischemic damage, which enhances the ability to predict the ultimate ischemic tissue recovery (21). In this work, we sought to identify a set of MRI parameters that would predict the final infarct size three months after stroke. Data were collected from ischemic stroke patients in the acute and subacute phases of ischemic stroke; the patients include both those treated with tpa and not treated with tpa, within three hours of ictus. MATERIALS AND METHODS Subjects We recruited patients with acute neurological deficit compatible with ischemic stroke, whether treated or not with intravenous tpa, who could be studied with multiparametric MRI within 24 hours of symptom onset. These patients were enrolled in the MRI Stroke Registry. We excluded patients with cerebral hemorrhage and/or history of prior significant stroke or other neurological deficit that would hamper accurate follow-up assessments. Informed consent was obtained before enrollment. The study was approved by the Human Rights Committee of the Henry Ford Health System. Stroke onset was defined as the last time the patient was known to be without neurological deficit. Patients were evaluated by the study neurologist. Clinical neurological deficit was graded using the National Institutes of Health Stroke Scale (NIHSS) at the time each MRI study was obtained. MRI studies were performed at acute ( 24 hours after stroke onset), subacute (three to five days), and at outcome (three months) phases of stroke. The ischemic stroke subtype was defined according to the trial of Org (danaparoid) in acute stroke treatment (TOAST) classification (22) based on the entire diagnostic evaluation the patient underwent during hospitalization and follow-up. For the present analysis, we included all patients who met the following criteria: 1) acute-phase MRI completed within 24 hours of stroke onset, 2) outcomephase MRI completed at the three-month follow-up assessment, 3) diagnosis of supratentorial ischemic stroke, and 4) MRI data of sufficient quality to allow computer processing and analysis. MRI Methodology MRI studies were performed on a 1.5-T GE Signa MR scanner with echo-planar (EPI) capability (GE, Milwaukee, WI, USA). Each MRI study consisted of axial multispin echo T2W imaging, T1-weighted (T1W) imaging pre- and postcontrast gadolinium-diethylenetriamine pentaacetic acid (Gd-DTPA), and DWI with slice thickness of 6 mm. For T1W images and T2W images, the field of view (FOV) was mm, matrix size ; for DWI, FOV was mm, matrix size Additional parameters for each study were: 1) T1W images: TR/TE 600/14 msec; 2) T2W images: TR/TE 2800/30, 60, 90, 120 msec; 3) axial DWI was performed using an echo-planar sequence: TR/TE 10,000/101 msec, b-values 1000, 600, 300, 0 seconds/mm 2, number of excitations (NEX) 1. Throughout the acquisition of the images, the patient was monitored. The MRI protocol contains a broader range of images for exploratory purposes, with total acquisition time of about 45 minutes. MRI Selection and ISODATA Analysis The study neuroradiologist and the neurologist selected at most three slices (images) that would cover most, or all, of the ischemic lesion evident on the acute DWI; usually these were adjacent slices. Given the slice thickness (6 mm), one or two slices might suffice. For multiple lesions at different locations, slices were selected to cover most of, or all of, the largest infarction. Figures 1, 2 and 3 illustrate multiparametric imaging of three patients with ischemic stroke of various sizes, one treated with tpa, and with the measurements of lesion on acute phase DWI (left), acute phase ISODATA composite (middle), and three-month T2W image (right). MR image processing included image coregistration and warping (23), and ISODATA multiparametric segmentation (17 20,24). ISODATA included one protondensity-weighted image (TE: 30 msec), one T2W image (TE: 90 msec), one T1W image, and two DWI images (b-values: 600, 1000 seconds/mm 2 ). ISODATA is an unsupervised segmentation method related to the K- means algorithm, with additional splitting and merging steps that allow for the adjustment of cluster centers. The ISODATA process has been described in a previous publication (21). After convergence, ISODATA clusters were standardized to the characteristics of two known clusters: normal white matter and cerebrospinal fluid (CSF). Signa-

3 Early MRI Predicts Final Infarct Size 497 Figure 1. Acute-phase DWI (Left), ISODATA (Middle) and outcome-phase T2W (Right) images, demonstrating a large right hemispheric infarct and an old left hemispheric infarct. Please note the acute phase ISODATA-DWI mismatch and the size of the outcome-phase T2W ischemic lesion as well as the inhomogeneity of the acute-phase ISODATA lesion. The bar indicates the signature numbers and color-coding of the ISODATA-defined ischemic lesion signatures. ture 1 (represented as blue color on the color coded bar; Fig. 1) was assigned to normal white matter and signature 12 (represented as white color; Fig. 1) to CSF. Thus, ischemic tissue clusters were identified with a signature number, in a range of 2 11 (see the color coded bar, between green and red codes; Fig. 1). The acute ISODATA signature value per slice was calculated adjusting for the lesion size among the regions based on the following formula: 1 Total ISODATA area M m 1 ISODATA Signature value m ISODATA area m, (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 lesion area, the ISODATA region corresponding to the lesion was overlaid on the ADC and T2 map, and the mean value 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 mean ratios (with contralateral values) of ADC and T2 (radc and rt2, respectively). For each of the different MRI weightings, we calculated the lesion area per MRI slice, as the sum of each lesion pixel area, multiplied by the number of lesion pixels. We report lesion size as mean ( SD) of the lesion area. Data Collection MRI data were collected from stroke patients in the acute, subacute (when possible), and final (three months) phases after stroke. Our focus was to use the acute or the subacute MRI to predict the final infarct size. The final infarct size was measured per MRI slice using the three-month T2W image. We included MRI measurements of ISODATA signature value, ADC/rADC, T2 value/rt2, and the ISO- DATA, DWI, and T2W image lesion sizes at the acute time. Given the impact of tpa on MRI parameters (21) obtained posttreatment, and its benefit on functional and lesion recovery (12,13), we included the variable of tpa 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 left mesial temporal lobe infarct. The bar indicates the signature numbers and color-coding of the ISODATA-defined ischemic lesion signatures.

4 498 Lu et al. Figure 3. Acute-phase DWI (left), ISODATA (middle), and outcome-phase T2W (right) images, demonstrating a small right internal capsule lacunar infarct in a patient who received treatment with tpa prior to MRI. The bar indicates the signature numbers and color-coding of the ISODATA-defined ischemic lesion signatures. of stroke onset to time of acute MRI as a potential predictor for the final infarct, given that the MRI parameters may be dependent on time. The ISODATA-DWI mismatch was calculated at the acute phase, defined as the lesion size difference between ISODATA lesion and DWI lesion. In addition, we calculated the changes and relative changes in MRI parameters at subacute from the acute (e.g., subacute MRI acute MRI) on a subgroup of patients with MRI obtained at both time points, yielding a total of 23 possible MRI parameters of interest. Of the 23 MRI parameters, 10, including the time of stroke to MRI, were collected at the acute phase; 13 MRI changes were collected at the subacute phase. Statistical Analysis Approach Since infarct size is not normally distributed, a similar data transformation, proposed in the National Institute of Neurological Disorders and Stroke (NINDS) rt-pa Stroke Study Group for CT infarct (12) was performed, thus producing a distribution more approximately normal. To avoid a possible variable confounding (25), or the treatment effect on acute MRI parameters, we tested the MRI parameter difference between the tpa-treated group and non-tpa-treated groups at the 0.05 level. We first tested the individual MRI parameter effect (e.g., ADC variable alone) or the effect of the MRI parameter by tpa treatment interaction (e.g., ADC with the treatment of tpa interaction) on the final infarct size, followed by multivariable modeling for independent predictive abilities. An MRI parameter with the tpa treatment interaction indicates that the predictive ability of the MRI parameter on the infarct size is dependent on whether or not patients received tpa treatment. We used the generalized estimation equation (GEE) (26,27), assuming independent subject (patients) and common correlation among slices per subject. All analyses were adjusted for the acute DWI lesion size by including the acute DWI lesion size in the regression model as a covariate. Variables with an individual effect, or an interaction between the variable and the treatment of tpa at the 0.15 level, were included in the first multivariable model. To avoid redundancy, the highly correlated variables (r 0.70), such as ADC and radc, had to be evaluated separately. Two-way or three-way effects (MRI parameter interactions or MRI parameters and tpa treatment interactions) were considered only if there was a one-way effect at the 0.05 level or a two-way MRI variable interaction at the 0.10 level, respectively. A final multivariable model included individual MRI parameters with P-value 0.05 or MRI parameter interactions with P-value A positive coefficient of the MRI parameter indicated an increase in final infarct size when the value of the MRI parameter is increased. A negative coefficient, in contrast, indicates a reduction in the final infarct size when the MRI parameter value is increased. We assessed the predictive ability by measuring the squared multiple correlation coefficient (R 2 ), which reflects the degree of variation in the outcome explained by the predictor variables, and is related to the correlation coefficient r R 2. R 2 ranges from 0 to 1.0, with a value close to 1 indicating that the model explains most of the variation in the outcome. Due to the paucity of MRI data collection at the subacute phase, we conducted two different multivariable models. The first model considered acute MRI parameters only (the acute multivariable model), and the second model considered both MRI at the acute and subacute time. Note that acute or subacute MRI parameters may predict a large reduction in infarct size in a patient with a single or multiple MRI slices, but it may not reflect the whole picture of this patient s recovery, because of the MRI 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. The median acute NIHSS was 7 (range: 2 26). A total of 15 patients (31%) had received treatment with intravenous tpa within three hours of stroke onset, and prior to the acute-phase MRI. The mean time interval from stroke onset to MRI study was 10.7 ( 5.6) hours. Populations

5 Early MRI Predicts Final Infarct Size 499 of the ischemic stroke subtypes were: cardiogenic embolism (N 17; 38%), large vessel atherosclerosis (N 3; 7%), lacunar stroke (N 12; 27%), undetermined etiology (N 10; 22%), other determined etiology (N 3; 7%). A total of 123 acute MRI slices were included from these 45 patients. A total of 32 of the 45 patients, 11 treated with tpa, had both acute- and subacute-phase MRI data on 88 slices. MRI parameters were first compared between the tpa treated group and the untreated group. The only MRI measures which differed between tpa treated and untreated groups were both the change and relative change in T2 value from the acute to the subacute with P-values of 0.02 and 0.04, respectively. Effect of Individual MRI Parameters on Final Infarct Size As noted, MRI parameters were first tested for their individual predictive effect with regard to the final infarct size, followed by multivariable modeling for independent predictive abilities. At the acute phase, the MRI variables, DWI and ISODATA lesion size, and the stroke onset MRI time interval had individual effects for prediction of the final infarct size, with P-values In addition, ADC and radc values, T2 and rt2 values, T2W image lesion size and ISODATA-DWI lesion mismatch also had predictive effects for the final infarct size (P 0.15); however, these effects were dependent on the treatment of tpa presence and absence (an MRI by tpa interaction). There was no ISODATA weighted signature effect predictive of the final infarct size (Table 1a). At the subacute phase, only two MRI-related parameters, the relative change of DWI lesion size (lesion size Table 1b Test of Individual Association Between Subacute-phase MRI Parameters and Final Infarct Size* Variables of interest P-value a for individual MRI parameter effect P-value a for MRI parameter by tpa Interaction Change of ISODATA lesion size b Relative change of ISODATA lesion size b Change of DWI lesion size b Relative change of DWI 0.01 c 0.89 lesion size b Change of T2W image lesion size b Relative change of T2W image lesion size b Change of ADC value Relative change of ADC value Change of radc value Relative change of radc value Change of T2 value 0.12 c 0.85 Relative change of T2 value Change of rt2 value Relative change of rt2 value *Change means the change in lesion size from acute to subacute phase. Relative change means the relative change in lesion size from acute to subacute phase which is defined as: (acute-subacute) phase lesion size/acute phase lesion size. a Analysis was based on transformed data using GEE adjusting for the acute DWI lesion and correlated MRI slices. b Measured by area per slice c Variables included in the first step multivariable model. Table 1a Test of Individual Association Between Acute-phase MRI Parameters and Final Infarct Size Variables of interest P-value a for individual MRI parameter effect P-value a for MRI parameter by tpa Interaction Tissue signature ADC value b radc b T2 value b rt b ISODATA lesion 0.01 b 0.37 size c DWI lesion size e 0.01 b 0.99 T2W image lesion b size c DWI-ISODATA b mismatch size c Stroke onset-mri 0.12 b 0.44 time interval Treatment with tpa 0.71 NA d a Analysis was based on transformed data using GEE adjusting for the acute DWI lesion and correlated MRI slices. b variables, or variable by tpa interaction, included in the first step multivariable model. c measured by area per slice. d NA not applicable. difference at subacute from the acute divided by acute lesion size), and the change in T2 value (the difference at subacute from the acute phase) had effects in prediction of final infarct size with P-values 0.15 (Table 1b). No MRI by treatment tpa interaction was observed at the P 0.15 level. Multivariable Modeling Results Using the backward model selection, and utilizing only the acute-phase MRI data (123 slices from 45 patients), the DWI lesion size (P 0.01), the DWI-ISODATA mismatch (P 0.01), as well as the interactions of radc by tpa 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 multivariable model. These findings mean that a larger lesion size on acute-phase DWI predicted a large final infarct [with a coefficient and (SD) of 0.90 (0.10)]. A DWI lesion size larger than the ISODATA lesion size predicted a decrease in the size of the final infarct, while a DWI lesion size smaller than the ISODATA lesion size predicted an increase in the size of the final infarct [coefficient 0.08 (0.03)]. The predictive ability of the acute-phase DWI and ISODATA-DWI lesion size mismatch on the final infarct size were independent of tpa treatment. However, the predictive abilities of the acute-phase radc,

6 500 Lu et al. Table 2 Acute Multivariable Model for Prediction of Final Infarct Size Variables in the final model Coefficients a (SD) P value b DWI lesion size 0.90 (0.10) 0.01 tpa treatment NA NA DWI-ISODATA mismatch 0.08 (0.03) 0.01 size Interaction of T2w image 0.01 lesion size by tpa treatment With tpa presence 0.21 (0.07) With tpa absence 0.07 (0.07) Interaction of radc by tpa 0.04 treatment With t-pa presence (7.71) With t-pa absence 2.14 (1.51) Interaction of rt2 with tpa 0.03 treatment With t-pa presence (6.73) With t-pa absence 4.40 (1.99) Goodness-of-fit (R 2 ) 0.70 a A positive coefficient indicates an increase in the final infarct size when the value of the MRI parameter is increased. In contrast, a negative coefficient indicates a reduction in the final infarct size when the value of the MRI parameter is increased. b Analysis was based on transformed data using GEE adjusting for the acute DWI lesion and correlated MRI slices. NA not applicable. rt2, or the T2W image lesion size on the final infarct size were dependent on whether treatment with tpa was given to the patient prior to the acute-phase MRI study. In patients who received tpa prior to the acute-phase MRI studies, a higher rt2 value [coefficient 10.43(6.73)], or a smaller T2W Image lesion size [coefficient 0.21 (0.07)] predicted a larger final infarct size (P 0.02, based on subgroup analysis). On the other hand, for patients who had not received tpa, a lower rt2 value [coefficient 4.40 (1.99)] predicted a larger final infarct size (P 0.02), while no effect of the acute-phase T2W image lesion size was detected (P 0.35). Similar effects were observed with the radc value, but the effects were marginal (P 0.11, based on subgroup analysis). No other two-way or three-way interactions were detected. The predictive ability (R 2 ) of the final model was 70% (Table 2), and 68% considering both the acute DWI and the acute DWI-ISODATA mismatch in a multivariable model. As a result, including radc, rt2, T2W image lesion and the MRI parameters-by-treatment interactions added only 2% more predictive ability as to the size of the final infarct. Using the same model analysis approach, we constructed the final multivariable model based on all acute and subacute phase data (88 slices from 32 patients with both MRI in the acute and subacute phases). The acute DWI lesion size (P 0.01), the DWI-ISODATA mismatch (P 0.01), the T2W image lesion size by tpa interaction (P 0.08), as well as the relative change in the DWI lesion size from the acute to subacute phase (P 0.01) remained in the final model (Table 3). The predictive ability of the final model (R 2 ) was 83%. The predictive abilities of the acute-phase parameters (DWI lesion size, DWI-ISODATA mismatch, and T2W image lesion size) on the final infarct size were similar compared to their predictive abilities in the acute-phase multivariable model. In addition to these acute-phase MRI predictors, the combined acute-phase and subacute-phase MRI parameters predicted the final infarct size with a higher predictive ability of 83% (increased from 70%). No other interactions were detected at the 0.10 level. DISCUSSION This study identifies those early MRI characteristics that can predict the final infarct size (e.g., at three months after stroke). A total of 23 parameters taken from MRI images obtained at the acute and subacute phases of stroke were assessed as possible predictors for the infarct size. With a rigorous statistical analysis approach, we developed two predictive models, one (the acute model) including only parameters from acute phase of stroke and one (the final model) incorporating parameters from both the acute and subacute phases of stroke. With these analyses we replicated previous studies findings that the final infarct size depends on the acute DWI lesion, and also demonstrated that it additionally depends on the acute DWI-ISODATA lesion mismatch and the change in the DWI lesion size from the acute to the subacute phase. We also demonstrated that the final infarct size also depends on the acute phase T2W image lesion size, as well as the acute phase radc and rt2 values. However, the predictive abilities of the latter parameters are influenced by whether treatment with tpa was applied prior to acute-phase image acquisition. Patients who received tpa treatment and had a lower radc value developed a larger infarct size at three months. Also, patients who received tpa treatment and had a larger T2W image lesion would have a smaller infarct, compared to the tpa-treated patients with a smaller T2W image lesion. We also ob- Table 3 Final Multivariable Model, Utilizing Acute and Subacute-phase Parameters, for the Prediction of Final Infarct Size Variables remaining in the final model Coefficients a (SD) P value b Acute-phase DWI lesion size 1.24 (0.08) 0.01 tpa treatment NA NA Acute-phase DWI-ISODATA.10 (0.03) 0.01 mismatch size Acute-phase T2w image lesion size 0.08 by tpa interaction With tpa presence 0.19 (0.07) With tpa absence 0.03 (0.07) Relative change of DWI lesion size c 1.08 (0.33) 0.01 Goodness-of-fit (R 2 ) 0.83 a A positive coefficient indicates an increase in the final infarct size when the value of the MRI parameter is increased. In contrast, a negative coefficient indicates a reduction in the final infarct size when the value of the MRI parameter is increased. b Analysis was based on transformed data using GEE adjusting for the acute DWI lesion and correlated MRI slices. c Relative change of lesion size was defined as: (subacute lesion size-acute lesion size)/acute lesion size; NA not applicable.

7 Early MRI Predicts Final Infarct Size 501 served that a lower acute-phase rt2 value leads to a larger infarct at three months in patients who had not received tpa prior to acute MRI study. In contrast to the acute radc and rt2 values or the T2W image lesion size, the predictive abilities of acute DWI lesion and the DWI-ISODATA mismatch are robust and independent of tpa treatment. The predictive ability is 68% considering both the acute DWI lesion size and the acute DWI-ISODATA lesion mismatch. Only 2% predictive ability was added by including radc, rt2, T2W image lesion, and the MRI parameters-by-tpa treatment interactions that complicated the model (by adding seven additional variables to the model), with little improvement on predictive ability. The values of ADC or T2 have the same effects, compared to radc or rt2, but had less predictive ability (67%) when radc and rt2 were replaced with ADC and T2 values in the final acute multivariable model (Table 2). Adding the relative DWI lesion size change from acute to subacute phase increased the predictive ability to 83% (Table 3). Without the acute T2W image lesion size by tpa treatment interaction, the predictive ability still remained 80%. Note that there is no direct comparison between the acute and subacute multivariable models because of differences in numbers of MRI slices and in the patient population. However, considering the subacute patient population (88 slices from 32 stroke patients), the predictive abilities are only 60% without the inclusion of the acute-to-subacute change in DWI. Therefore, the change of DWI is an important predictor for the eventual infarct size. The acute DWI lesion size was a strong predictor of final infarct size, a finding that is in agreement with those of other studies (28,29), which demonstrated that the DWI lesion volume in the acute phase of stroke predicts the infarct volume as shown by the T2W image infarct volume at outcome. In addition, the relative change in the DWI lesion size from acute to subacute phase was a strong independent predictor of final infarct size. This likely reflects that when there is DWI lesion expansion, the second measurement will match the final infarct size more closely. The acute DWI-ISODATA mismatch lesion size was also a strong predictor of the final infarct size. Several authors (15,30) have used a different concept, PWI-DWI mismatch, to indicate that the mismatch volume likely represents viable tissue which, if left untreated, may evolve to infarction. In untreated patients who demonstrate PWI-DWI mismatch in the acute phase, the DWI lesion will very likely expand over a time period of several hours to days. One weakness of the PWI in acute ischemia is that the PWI lesion volume probably includes both critically hypoperfused tissue (ischemia) and hypoperfused (but not at a critical level benign oligemia) tissue, which will never evolve into infarction. The ISODATA-DWI, -T2W, and -T1W composite images can identify subtle abnormalities in ischemic tissue not yet seen by any one of the individual MRI techniques (DWI, T2W images, or T1W images). We believe this happens because there are contributions to image contrast from various factors, such as perfusion, proteolysis, and changes in deoxyhemoglobin concentration, which are not visible by any one modality, but have a cumulative effect on tissue signature when analyzed with ISODATA. In this sense, an acute-phase mismatch between ISODATA and DWI may be an indication that the ISODATA lesion is identifying the portion of the perfusion lesion that undergoes some degree of ischemic tissue injury or the MRI signs of a perfusion deficit, and, therefore, the DWI lesion has a potential to grow. The DWI-ISODATA mismatch may thus serve as a surrogate of the DWI-PWI mismatch with a better feasibility, because ISODATA only requires T1W images, T2W images, and DWI. This study has not made a direct correlation between ischemic lesion size and clinical neurological or functional scales, since its primary focus was the prediction of ischemic lesion size and not clinical correlates. Several previous studies (28,29,31) have demonstrated good correlations of acute DWI lesion volumes with acute and chronic NIHSS scores and the correlation of chronic T2W image infarct size and NIHSS scores. Our previous work has also demonstrated strong correlations between ISODATA lesion size and NIHSS scores in both the acute and outcome phases, as well as between ISODATA and DWI lesion size in the acute phase and ISODATA and T2W image lesion size in the outcome phase (20). This suggests, albeit indirectly, that the findings of the present study may be translated into meaningful clinical correlations and outcomes. ISODATA analysis can be used for any type and location of ischemic stroke. Previous work has indicated that it can be applied for brainstem stroke with good definition of the ischemic lesion in that location (28). ISODATA is reproducible because the algorithm is unsupervised, as are all other processing methods used in this study. The main limitation for the application of the ISODATA in the emergency management of stroke may lie in the time needed to complete the analysis; it takes about two hours to process ISODATA on a single slice, with the majority of this time (90%) spent for accurate coregistration of the multiple images. With advances in the accuracy of image acquisition, computer technology, and optimization of the software, it may be possible to perform the analysis in a few minutes and thus accommodate the requirements of the emergency clinical practice. No effect of tpa on the final infarct size was found, possibly for the following reasons: 1) The acute phase multiparametric imaging was obtained after treatment with tpa was already given. It is possible that different findings would appear if tpa was given before the acute phase study. 2) The study was designed to explore the relation of MRI parameters measured at acute or subacute times to the final infarct size. 3) The study was not designed to have sufficient power to detect the tpa effect. It should be noted that only a marginal tpa effect was observed on three-month infarct size on CT scans in the NINDS rtpa stroke study based on a total of 624 patients (12). The medians and interquartiles (25% and 75%) of the transformed three-month infarction were 9.42 (5.48, 18.33) in the tpa-treated group and (5.74, 11.90) in the non-tpa-treated group, with P-value 0.71 (24). In addition, three patients who had hemorrhagic transformation after the initial strokes

8 502 Lu et al. were included in our study and only one had tpa treatment prior to the hemorrhage. Of additional interest are the interactions between preceding treatment with tpa acute T2W image lesion size, radc, and rt2 values. Patients who received tpa treatment and had a lower radc value developed a larger infarct size at three months. This probably indicates that the ischemic tissue injury was more severe and accompanied by inadequate reperfusion, as the latter would be expected to produce higher ADC values. The findings from T2W imaging, i.e., that a larger acute T2W image lesion would predict a smaller final infarct and that a lower acute rt2 value leads to a larger final infarct in untreated patients, are likely to have a similar interpretation, indicating more complete and inadequate reperfusion, respectively. Reperfusion, accompanied by edema, could be expected to produce a higher rt2 value and a larger T2W image lesion post-tpa treatment. At present, response of the acute stroke to treatment with tpa can be judged at the three-month time point based on the clinical recovery. There have been no good markers developed to distinguish responders from nonresponders in the early phase following treatment. This issue is expected to become increasingly important as salvage treatments are developed for patients deemed to be nonresponders. The model presented here has the potential to solve this problem, and allow early definition of the potential nonresponders, thus allowing salvage treatments to be offered. In conclusion, acute DWI and the acute DWI-ISO- DATA mismatch are the predictors of the final infarct size at three months after stroke onset and are independent of preceding treatment with tpa. The acute subacute DWI lesion size change increases the predictive ability of the model with regard to the final infarct size. These findings could be used for the rational design of image-based acute treatment trials. ACKNOWLEDGMENT We thank Lula Adams for editing. REFERENCES 1. Saur D, Kucinski T, Grzyska U, et al. Sensitivity and interrater agreement of CT and diffusion-weighted MR imaging in hyperacute stroke. AJNR Am J Neuroradiol 2003;24: Fiebach JB, Schellinger PD, Jansen O, et al. CT and diffusionweighted MR imaging in randomized order: diffusion-weighted imaging results in higher accuracy and lower interrater variability in the diagnosis of hyperacute ischemic stroke. Stroke 2002;33: Barber PA, Darby DG, Desmond PM, et al. Identification of major ischemic change. Diffusion-weighted imaging versus computed tomography. Stroke 1999;30: Lansberg MG, Albers GW, Beaulieu C, et al. Comparison of diffusion-weighted MRI and CT in acute stroke. Neurology 2000;54: Watanabe Y, Mitomo M, Tokuda Y, et al. Eclamptic encephalopathy: MRI, including diffusion-weighted images. Neuroradiology 2002;44: Oppenheim C, Samson Y, Manai R, et al. Prediction of malignant middle cerebral artery infarction by diffusion-weighted imaging. Stroke 2000;31: Parsons MW, Li T, Barber PA, et al. Combined (1)H MR spectroscopy and diffusion-weighted MRI improves the prediction of stroke outcome. Neurology 2000;55: Arenillas JF, Rovira A, Molina CA, et al. Prediction of early neurological deterioration using diffusion- and perfusion-weighted imaging in hyperacute middle cerebral artery ischemic stroke. Stroke 2002;33: Hermier M, Nighoghossian N, Adeleine P, et al. Early magnetic resonance imaging prediction of arterial recanalization and late infarct volume in acute carotid artery stroke. J Cereb Blood Flow Metab 2003;23: Ahlhelm F, Schneider G, Backens M, et al. Time course of the apparent diffusion coefficient after cerebral infarction. Eur Radiol 2002;12: DeWitt LD, Buonanno FS, Kistler JP, et al. Nuclear magnetic resonance imaging in evaluation of clinical stroke syndromes Ann Neurol 1984;16: NINDS. Effect of intravenous recombinant tissue plasminogen activator on ischemic stroke lesion size measured by computed tomography. The National Institute of Neurological Disorders and Stroke (NINDS) rt-pa Stroke Study Group. Stroke 2000;31: National Institute of Neurological Disorders and Stroke rt-pa Stroke Study Group. Tissue plasminogen activator for acute ischemic stroke. N Engl J Med 1995;333: Marks MP, Tong DC, Beaulieu C, et al. Evaluation of early reperfusion and i.v. tpa therapy using diffusion- and perfusion-weighted MRI. Neurology 1999;52: Parsons MW, Barber PA, Chalk J, et al. Diffusion- and perfusionweighted MRI response to thrombolysis in stroke. Ann Neurol 2002;51: Wu O, Koroshetz WJ, Ostergaard L, et al. Predicting tissue outcome in acute human cerebral ischemia using combined diffusion- and perfusion-weighted MR imaging. Stroke 2001;32: Soltanian-Zadeh H, Pasnoor M, Hammoud R, et al. MRI tissue characterization of experimental cerebral ischemia in rat. J Magn Reson Imaging 2003;17: Jacobs MA, Knight RA, Soltanian-Zadeh H, et al. Unsupervised segmentation of multiparameter MRI in experimental cerebral ischemia with comparison to T2, diffusion, and ADC MRI parameters and histopathological validation. J Magn Reson Imaging 2000;11: Jacobs MA, Mitsias P, Soltanian-Zadeh H, et al. Multiparametric MRI tissue characterization in clinical stroke with correlation to clinical outcome: part 2. Stroke 2001;32: Mitsias PD, Jacobs MA, Hammoud R, et al. Multiparametric MRI ISODATA ischemic lesion analysis: correlation with the clinical neurological deficit and single-parameter MRI techniques. Stroke 2002;33: Mitsias PD, Ewing JR, Lu M, et al. Multiparametric iterative selforganizing MR imaging data analysis technique for assessment of tissue viability in acute cerebral ischemia. AJNR AJNR Am J Neuroradiol 2004;25: Adams HP, Jr., Bendixen BH, Kappelle LJ, et al. Classification of subtype of acute ischemic stroke. Definitions for use in a multicenter clinical trial. TOAST. Trial of Org in Acute Stroke Treatment. Stroke 1993;24: Ghanei A, Soltanian-Zadeh H, Jacobs MA, et al. Boundary-based warping of brain MR images. J Magn Reson Imaging 2000;12: Jacobs MA, Zhang ZG, Knight RA, et al. A model for multiparametric MRI tissue characterization in experimental cerebral ischemia with histological validation in rat: part 1. Stroke 2001;32: McNeil D. Epidemiological research methods, 1st edition. New York: John Wiley & Sons; p Liang K, Zeger SL. Longitudinal data analysis using generalized linear models. Biometrika 1986;72: Zeger SL, Liang K. Longitudinal data analysis for discrete and continuous outcomes. Biometrics 1986;42: Barber PA, Darby DG, Desmond PM, et al. Prediction of stroke outcome with echoplanar perfusion- and diffusion-weighted MRI. Neurology 1998;51: Warach S, Dashe JF, Edelman RR. Clinical outcome in ischemic stroke predicted by early diffusion-weighted and perfusion magnetic resonance imaging: a preliminary analysis. J Cereb Blood Flow Metab 1996;16: Baird AE, Benfield A, Schlaug G, et al. Enlargement of human cerebral ischemic lesion volumes measured by diffusion-weighted magnetic resonance imaging. Ann Neurol 1997;41: Lovblad KO, Baird AE, Schlaug G, et al.. Ischemic lesion volumes in acute stroke by diffusion-weighted magnetic resonance imaging correlate with clinical outcome. Ann Neurol 1997;42: