Infarct Volume Prediction by Early Magnetic Resonance Imaging in a Murine Stroke Model Depends on Ischemia Duration and Time of Imaging
Background and Purpose—Despite standardization of experimental stroke models, final infarct sizes after middle cerebral artery occlusion (MCAO) vary considerably. This introduces uncertainties in the evaluation of drug effects on stroke. Magnetic resonance imaging may detect variability of surgically induced ischemia before treatment and thus improve treatment effect evaluation.
Methods—MCAO of 45 and 90 minutes induced brain infarcts in 83 mice. During, and 3 and 6 hours after MCAO, we performed multiparametric magnetic resonance imaging. We evaluated time courses of cerebral blood flow, apparent diffusion coefficient (ADC), T1, T2, accuracy of infarct prediction strategies, and impact on statistical evaluation of experimental stroke studies.
Results—ADC decreased during MCAO but recovered completely on reperfusion after 45 and partially after 90-minute MCAO, followed by a secondary decline. ADC lesion volumes during MCAO or at 6 hours after MCAO largely determined final infarct volumes for 90 but not for 45 minutes MCAO. The majority of chance findings of final infarct volume differences in random group allocations of animals were associated with significant differences in early ADC lesion volumes for 90, but not for 45-minute MCAO.
Conclusions—The prediction accuracy of early magnetic resonance imaging for infarct volumes depends on timing of magnetic resonance imaging and MCAO duration. Variability of the posterior communicating artery in C57Bl6 mice contributes to differences in prediction accuracy between short and long MCAO. Early ADC imaging may be used to reduce errors in the interpretation of post MCAO treatment effects on stroke volumes.
Experimental studies in models of disease are the underpinning of modern biomedicine. Worryingly, robustness, reproducibility, and thus predictiveness of experimental medicine have recently been questioned,1–3 a discussion which lead to a widespread concern about rigor in science (eg, Economist October 19, 2013, "How science goes wrong"). Stroke research has been the epitome of these concerns because it experienced an almost complete failure to transfer successful treatments in experimental models of stroke to more effective treatment of patients at a grand scale.4 This apparent lack of predictiveness of biomedical research has been attributed to many factors, but poor experimental design, statistical quality, and bias feature top of the list.5–7
In most animal stroke studies, the treatment effect is evaluated by statistical comparison of infarct sizes of treatment versus placebo groups. Group means are compared, and treatment is declared effective depending on a predefined type I error level (α), most commonly P<0.05. Notably, experimental stroke studies, in contrast to clinical stroke trials, almost never predefine criteria for the type II error (β). Given α and β, the acceptance of a treatment effect (ie, the rejection of the null hypothesis) depends on the effect size, which relates to a property of the experimental drug, and the variance of the measured infarct sizes, which relates to biological factors as much as to the properties of the model and its execution. A distinct advantage of experimental stroke models is that parameters with high variability in the clinical setting, such as onset, location and duration of vessel occlusion as well as timing of lesion size determination (eg, by magnetic resonance imaging [MRI]) can be standardized, potentially reducing variability of results. Despite these advantages, variability of infarct sizes is substantial, owing to the variability of the surgical procedure8,9 and interanimal variability of brain blood vessel anatomy10–12 among other factors. Because of the substantial infarct size variability (SDs usually ≈30% of the mean) and small group sizes (between 4 and 10) in experimental stroke research, it is likely that false-positive results and an overestimation of effect sizes abound in publications.6 At the same time, treatment effects may frequently remain undetected.
What can we do to cope with the large variability of infarct sizes besides increasing group sizes and improving our experimental paradigm and skill of experimentalists? One approach would be to measure the early ischemic lesion volume before onset of treatment with noninvasive brain imaging and then compare early ischemic lesion volumes between treatment and placebo groups or infarct volumes predicted by early ischemic lesion volumes with those observed. Furthermore, pretreatment determination of early lesion volumes may allow unbiased exclusion of outliers, such as animals with failed middle cerebral artery occlusion (MCAO) or extreme variations of early ischemia volumes.
Here, we tested whether a pretreatment MRI approach may be used to gain information on pretreatment ischemic lesion variability in a common murine stroke model and to what extent early ischemic lesions predict final infarct sizes. We characterized the early time courses of commonly used, widely available MRI parameters for 2 different MCAO durations and 3 early measurement time points. We subsequently used these data to evaluate the potential of pretreatment MRI to improve the analysis of experimental stroke treatment studies.
Materials and Methods
We performed all animal experiments in accordance with international guidelines. The local official committee approved all animal experiments (Landesamt für Arbeitsschutz, Gesundheitsschutz und technische Sicherheit, Berlin, Germany). Eighty-three males, 8- to 10-week-old C57Bl/6 (Charles River, Sulzfeld, Germany) or 129S6/SvEv (BfR, Berlin, Germany) mice were used. All experiments were performed according to Standard Operating Procedures of the Department of Experimental Neurology.
Mice were placed on a heating pad and anesthetized with isoflurane in a mixture of oxygen and dinitrogen monoxide (N2O). A coated 5.0-nylon monofilament was introduced into the left external carotid artery (ECA) and carefully advanced to the origin of the left MCA via the internal carotid artery. The common carotid artery as well as the ECA were ligated and remained occluded during reperfusion. A single, experienced surgeon performed MCAO.
To evaluate the effect of MCAO duration and timing of MRI on infarct prediction accuracy, we performed experiments on 6 experimental groups (Figure 1). MCAO durations of 45 and 90 minutes were chosen to induce moderate and large infarct sizes, respectively. MRI was performed during MCAO or at 3 or 6 hours after MCAO onset. Group sizes were 10 to 20 animals (Table I in the online-only Data Supplement).
MRI and Image Processing
Details on acquisition and processing of MRI are provided as online-only Data Supplement. Cerebral blood flow (CBF; a flow sensitive alternating inversion recovery protocol),13,14 apparent diffusion coefficient (ADC), and T1 and T2 MRI were acquired. Early lesion volumes were delineated by 2 raters blinded to final infarcts. Rater independent evaluation of early lesion volumes was performed for validation of rater-dependent results. Final infarcts were delineated on T2-weighted (T2w) images obtained 48 hours after MCAO.
Infarct Probability Analysis and Contribution of Posterior Slices
From T2w images at 48 hours, coregistered to a standard brain, we calculated maps, which show the percentage of animals exhibiting infarction for each voxel (Figure 2). In the following, we refer to these as spatial infarct frequency maps.15 As mice of the 45-minute MCAO group with MR during MCAO were SV129 and all other C57Bl/6, we restricted this analysis to the 4 groups with MRI 3 or 6 hours after MCAO.
To evaluate the contribution of infarcts in posterior slices to variability of infarct sizes, we determined the contribution of infarct volumes of individual brain slices to the total infarct volume (infarct volume in single slices divided by total infarct volume, Figure 2, C1). The same analysis was performed for the contribution to the SD (Figure 2, C2). The number of animals needed to detect a difference in infarct volumes of 30% was calculated separately for all slices using mean and SD of infarct volumes of the respective slice (with α set to 0.05 and the power to 0.8, Figure 2, C3). To evaluate whether ADC measurement during MCAO could detect involvement of posterior slices, we analyzed the correlation of early ADC lesion volumes with final infarcts for slices ≈3 to 5 mm from the rhinal fissure (Figure 2, C5, core MCA territory), versus ≈6 to 8 mm from the rhinal fissure (MCA/posterior cerebral artery [PCA] watershed and PCA territory, Figure 2, C6).
Coefficient of Determination of Infarct Sizes by Early MRI Lesion Sizes
We calculated the coefficients of determination (r2) of visually delineated lesion sizes for early MRI with T2w imaging at 48 hours for all groups. To assess reliability of this procedure, 2 raters, blinded to the results of each other, evaluated lesion sizes and coefficients of determination were compared. Animals with no infarct on T2w imaging at 48 hours were excluded.
Infarct Volume Prediction
We determined receiver-operating characteristics (ROC) curves and calculated area under the curve (AUC) for the following prediction models: (1) early ADC lesion, (2) spatial infarct frequency, and (3) a combination of both. ROC curves were generated from infarct probability maps by calculating sensitivity and specificity of infarct prediction on a voxel-wise basis for each probability threshold. A detailed description of the calculation of infarct probability maps is provided in the online-only Data Supplement. We did not apply the mismatch concept (prediction of lesion growth by mismatch between early ADC and CBF lesions) because early ADC lesions did not represent infarct core, our CBF method could not distinguish between benign oligemia and critical hypoperfusion on an individual animal basis and 4 of our 6 groups had MRI after reperfusion.
Analysis of the Impact on Treatment Effect Evaluation
To illustrate possible effects of the variance in early ADC lesions on experimental stroke studies, we created an artificial 2 group design from our data by using permutations resulting in data sets with all possible assignments of animals of 1 of our experimental groups to 2 groups (eg, for division of 11 animals of experimental group 1 in 2 groups of 5 and 6 animals, respectively, there are 462 different possible allocations): The T2w lesions at 48 hours of these 2 groups were then compared using a t test. If a significant difference was detected (P<0.05), this was considered a type I error, as all animals had been treated equally and the detected difference was, therefore, a chance finding. In case of a significant difference, we subsequently compared early ADC lesion volumes using a t test. If P<0.05 in this comparison, we considered that early MRI had identified a difference in early ADC lesion volumes as a confounder leading to a type I error. We used this simple approach to illustrate more intuitively the effect of early ADC lesion differences on chance findings of infarct sizes differences for different MCAO durations and MR measurement time points. For comparison of 2 small samples, Janusonis et al16 have suggested ANCOVA to correct for pretreatment baseline variability. We therefore added ANCOVA as an additional test to evaluate the impact of early ADC lesion volume.
Lesion Sizes and Variability of Lesion Sizes
Of 83 animals, 3 (3.6%) died before imaging at 48 hours and 1 had artifacts on T2w imaging precluding reliable lesion volume determination at 48 hours. These 4 animals were excluded from further analysis. Of the remaining 79 animals, 8 (10%) had no T2w lesion at 48 hours, indicating failed MCAO. Excluding animals with failed MCAO reduced the coefficient of variation from 56% to 44% for 45-minute MCAO and from 51% to 36% for 90-minute MCAO. Final lesion sizes along with SD and coefficient of variation are shown in Table I in the online-only Data Supplement.
Of the 8 animals with no lesion, 4 had ADC imaging during MCAO and none of the 4 had an early ADC lesion. In contrast, all 24 animals with MRI during MCAO and infarcts at 48 hours exhibited an early ADC lesion, indicating that early ADC can be used to exclude animals with failed MCAO. As evidenced by the reduction of coefficients of variation, exclusion of animals with failed MCAO improves statistical assessment of treatment studies.
Figure 2 shows infarct probability maps for 45-minute MCAO, 90-minute MCAO, and the difference in infarct probability for these MCAO durations. The anterior slices covering the core MCA territory showed similar infarct areas with little variability. This is reflected in a high contribution of these slices to infarct volume and low contribution to SD (Figure 2, C1 and C2). In contrast, posterior slices contributed less to mean infarct volume, but substantially to SD of infarct volumes. Hence, analyzing single slices, the number needed to detect a significant difference in an experimental stroke study is low in anterior slices but much higher in posterior slices (Figure 2, C3). The difference in mean infarct volumes between 45- and 90-minute MCAO was low in anterior slices—7% ≈3 to 4 mm from the rhinal fissure—but increased to 81%, ≈9 mm from the rhinal fissure. The initial ADC lesion during MCAO correlated well with the ischemic T2w lesion at 48 hours in posterior slices for 90-minute MCAO, but not for 45-minute MCAO because most mice did not develop infarctions in the posterior slices despite substantial initial ADC lesions in the 45-minute MCAO group.
Early MRI Characteristics
In-group average MRI (coregistered to an average brain template) of early ADC, CBF, T1map, and T2w along with T2w images at 48 hours are shown in Figure I in the online-only Data Supplement. As visible on these images, ADC and CBF deliver discernable lesions in the early phase, whereas T2-weighted imaging shows pronounced hyperintense lesion areas at 48 hours after MCAO. The ADC lesion largely disappears at 3 hours after 45 minutes of MCAO. For 90-minute MCAO, a part of the ADC lesion recovers although ADC further decreases in the core MCA territory. The ADC lesion reappears at 6 hours after 45-minute MCAO.
Figure 3 illustrates changes in lesion volume and absolute values of ADC and CBF in the early phase during or shortly after MCAO. ADC dropped to values below those of the nonischemic hemisphere within minutes after MCAO. Again, ADC lesion shrinkage at 3 hours and regrowth at 6 hours is visible. During reperfusion, at 3 hours after onset of MCAO, ADC had recovered completely (45-minute MCAO) or partly (90-minute MCAO, Figure 3; Figure II in the online-only Data Supplement).
The complete recovery of ADC lesions at 3 hours after 45-minute MCAO is reflected not only in visually determined lesion volumes and absolute ADC values of the ischemic hemisphere but also in the absolute ADC values in the region that later developed a T2w lesion (Figure 3, middle row) and in ADC histograms (Figure II in the online-only Data Supplement). At 6 hours after MCAO, ADC declined again in both groups (Figure 3; Figure II in the online-only Data Supplement). CBF was largely reduced during MCAO in the ischemic hemisphere and recovered only partly on reperfusion at 3 and 6 hours after onset of MCAO. The perfusion deficit persisting after removal of the thread that occludes the MCA is explained by the permanent occlusion of the common and ECA. Reperfusion occurs via collaterals, for example, the posterior and anterior communicating arteries. At 6 hours after MCAO, average hemispheric CBF reached ≈60% of the nonischemic hemisphere regardless of occlusion time. CBF within the later ischemic T2w lesion did not differ relevantly from CBF of areas, which did not proceed to infarction (Figure 3). This provides evidence that reperfusion in the MCA territory was complete to the degree possible when ECA and common carotid artery remain occluded. Furthermore, we found no relevant correlation of average hemispheric CBF with ADC values at 3 or 6 hours after MCAO. We have no clear explanation for the overall reduced CBF in group 2 compared with the other groups, but repetitive anesthesia with short recovery period may have contributed to this finding.
Correlation of Early MRI With Final Infarcts
To evaluate which parameter could best be used in the early evaluation of ischemia, visually determined lesion volumes of early ADC, CBF, T1map, and T2w imaging were compared with T2w lesion volumes determined 48 hours after MCAO. Lesion delineation was performed blinded to final infarcts. Correlation of early lesion sizes with T2w lesion sizes at 48 hours was consistently high for MR during (r2=0.87 and 0.79 for 2 raters, respectively) or at 6 hours (r2=0.85 and 0.78) after 90-minute MCAO (Figure 4). There was a weak-to-moderate correlation of early ADC lesions with late T2w lesions for 45-minute MCAO and measurement at 6 hours after MCAO (r2=0.35 and 0.54). Of note, when excluding animals with no T2w lesion at 48 hours, there was no significant correlation of ADC lesions during MCAO with T2w lesions at 48 hours for 45-minute MCAO. The other parameters did not correlate well with late T2w lesions for all time points. For CBF, the reason is most likely the fact that persisting occlusion of common carotid artery and ECA caused a persisting CBF reduction covering the entire hemisphere. T1map and T2w imaging did not show robust lesions in the early phase in our study.
Infarct Prediction by Early MRI
We evaluated 3 different models predicting infarction (T2w lesion at 48 hours) 3-dimensionally (3D) on a voxel-wise basis. ROC curves were determined for prediction via early ADC, spatial infarct frequency, and a combination of early ADC and spatial infarct frequency. Figure 5A shows the AUCs for different occlusion times, MRI measurement time points, and prediction models. AUCs were high (consistently >0.90) for prediction by spatial infarct frequency or a combination of spatial infarct frequency and ADC. Prediction by early ADC alone was good for MRI during MCAO and at 6 hours after MCAO, with AUC values ≈0.8. In a further step, we calculated the root mean square deviation of predicted infarct volumes from measured infarct volumes for optimal infarction probability thresholds as determined from ROC curves. The results were compared with the root mean square deviation derived from linear regression analysis of visually obtained ADC lesion volumes (Figure 5B). The best prediction was obtained for visually determined ADC lesions for 90-minute MCAO when MRI was obtained during MCAO or at 6 hours after MCAO. At first glance, there is a discrepancy between highest AUC (Figure 5A) for prediction via spatial infarct frequency maps and the smallest root mean square deviation obtained from linear regression analysis of visually determined ADC lesions. However, this is explained by the fact that voxels with equal ADC may or may not proceed to infarction, which reduces AUC for a voxel-wise prediction model. However, if the number of voxels that proceed to infarction despite normal ADC (or number of voxels which recover despite reduced ADC) is proportional to the final infarct, prediction by linear regression can still perform well. Furthermore, it has to be kept in mind that AUC depends equally on sensitivity and specificity. In a model that produces small infarcts, ROC analysis may suggest a good prediction if infarcts sizes are generally overestimated, as sensitivity will be high, but specificity not much reduced because of a high number of voxels correctly identified not to proceed to infarction (true negatives). Then, the root mean square deviation of predicted infarcts may nevertheless be relatively high with respect to average infarct volume.
Impact on the Evaluation of Treatment Effects in Experimental Stroke Studies
By simulation we assessed, whether early MRI may identify differences in early ischemic lesion volumes as a confounder for differences in final infarct sizes. To this end, all theoretically possible ways of partitioning 1 experimental group (ie, from 1 population) into 2 groups (samples) were tested for a chance finding of a significant difference of T2w lesion sizes at 48 hours (ie, falsely concluding that the 2 samples come from different populations). The results are shown in Table. For example, group 1 (MCAO 90 minutes and MRI during MCAO) contained 11 animals and thus 462 different divisions in 2 groups of 5 and 6 animals, respectively, were possible. A t test revealed a significant difference (P<0.05) of T2w lesions sizes in 23 (5.0%) of these 462 possible divisions. As the animals came from the same population, these represent chance findings of a treatment effect.
In case of significant chance findings, we evaluated further the impact of early ADC to the finding by (1) a t test comparing early ADC lesion volumes and (2) ANCOVA with early ADC lesion volumes as covariate.16
In 15 of these 23 cases, early ADC lesions were significantly different, thus indicating that the detected difference at 48 hours was because of a difference of the early ischemic lesion. ANCOVA failed to confirm allocation treatment effect in the majority and indicated a significant effect of early ADC on final infarct volume in all 23 cases. Using these strategies, significant confounding by early ADC was present and after adjustment for confounding, significant group differences disappeared in the majority of cases for groups 1 and 3, but not for the other groups (Table).
The main findings of our study are: (1) ADC decreases within minutes during transient MCAO in mice, but shows a triphasic time course with full or partial recovery after reperfusion and secondary decline within a few hours, depending on duration of MCAO. The secondary decline in ADC is not paralleled by a secondary decline in CBF. (2) In C57/bl6 mice, the posterior slices covering MCA/PCA watershed and PCA territory have a major contribution to infarct size variability and infarct size differences between 45- and 90-minute MCAO. (3) Accuracy of infarct size prediction using ADC depends strongly on timing of MRI and duration of MCAO. Early ADC lesion volumes during MCAO and at 6 hours after MCAO accurately predict infarct volumes for 90 minutes MCAO. (4) Chance findings of infarct volume differences can be explained by differences in early ADC lesion volume for 90-minute MCAO, but not for 45-minute MCAO. Hence, depending on MCAO duration and timing of MRI, one potential confounder leading to type I errors in treatment effect evaluation may be identified using early ADC imaging.
Characteristics of Early MRI in Mouse Stroke
The observed triphasic ADC time course with an initial rapid decline, followed by recovery after reperfusion and secondary decline within hours confirms earlier reports in rats.17–21 Full recovery of ADC at 3 hours after MCAO onset has been reported for MCAO duration of 8 to 30 minutes in rats and mice.22–24 Of note, this does not reflect complete tissue recovery, as MR spectroscopy may still detect significant changes in neurochemical profiles.22,25 The initial as well as the secondary decline of ADC has been shown to be associated with ATP depletion but not secondary deterioration of perfusion in rats.26 Our CBF data at 6 hours after reperfusion (Figure 3; Figure II in the online-only Data Supplement) confirm this notion for mice. Reperfusion was not relevantly different in areas proceeding to infarction when compared with those that did not. This indicates complete reperfusion to the degree possible with persisting common and external carotid artery occlusion (with CBF at ≈60% of the contralateral hemisphere). We cannot rule out the possibility, however, that the persisting hypoperfusion, although safely above levels of immediate ischemic threat, may have contributed to secondary ischemic damage.
ADC recovery was largely complete 3 hours after MCAO onset for 45-minute MCAO (Figure 3; Figure II in the online-only Data Supplement). This may indicate almost exclusive secondary ischemic damage (with temporary restoration of energy metabolism and reversal of cytotoxic edema after reperfusion) for this occlusion duration. A similar ADC time course (with temporary ADC recovery) has been demonstrated in patients with stroke undergoing successful intra-arterial thrombolysis.27 Interestingly, ADC reversal28,29 as well as ADC lesion growth30 have both been well described in human stroke studies using MRI, but reports on secondary ischemic injury after early ADC reversal are rare. This may be because of imaging protocols of stroke studies, which do not include repetitive MRI in the first hours after reperfusion. Hence, temporary ADC reversal may simply occur undetected. However, it could also indicate that secondary ischemic damage after initial temporary restoration of energy metabolism is less frequently encountered as pathophysiological sequence in patients with stroke when compared with experimental stroke studies.
The reversal of ADC on reperfusion has implications for infarct prediction: although ADC lesion volumes determined during 90-minute MCAO accurately predicted T2w ischemic lesions at 48 hours, ADC reversal largely prohibited infarct prediction by ADC when lesions were determined at 3 hours after MCAO onset. Secondary ADC decline again permitted accurate prediction when ADC lesions were determined 6 hours after 90-minute MCAO (Figure 4). The time course of ADC after reperfusion has to be taken into account when planning an experimental stroke study that includes infarct prediction via early ADC measurement. Our data on 45-minute MCAO indicate that accurate prediction of infarct sizes may not be possible for short occlusion times. One tempting interpretation is that for short occlusion times, variability of pathophysiological events leading to secondary ischemic damage but not initial ischemic lesion volumes determine final infarct volume variability.
Involvement of Posterior Slices/PCA Territory
In contrast to most humans, the PCA is supplied via the internal carotid artery in mice. As ECA, common carotid artery, and MCA are occluded in the most commonly used MCAO models, tissue fate in the PCA territory depends on collateral supply from the posterior communicating artery (Figure 2) and details of the MCAO model used.31 The high variability of the posterior communicating artery calibre and patency in C57Bl/6 mice32 explains our observation of high variability of infarct volumes in posterior slices (Figure 2). This observation is also in line with studies demonstrating large SDs compared with mean infarct volumes in posterior slices in mice.11,33,34 Evaluating only anterior slices in experimental stroke studies may, therefore, improve treatment effect detection, if the evaluated treatment effect equally reduces infarct volumes in anterior when compared with posterior slices. However, in our study the difference in anterior slice infarct volumes between 45- and 90-minute MCAO was low compared with posterior slices. This may indicate that the time course of ischemic damage is different for the PCA/MCA watershed and PCA territory in our MCAO model. A potential explanation is that collateral flow via the posterior communicating artery may be sufficient to prevent permanent ischemic damage commonly in 45-minute MCAO but variably in 90-minute MCAO, in line with results reported by McColl et al.35 It could be that not only variability of infarct volumes in control groups but also neuroprotective treatment effects are larger in the posterior slices in the MCAO model. Our data therefore indicate that posterior slices should routinely be evaluated in experimental stroke treatment studies in mice.
Infarct Prediction Strategies
We compared 2 different infarct prediction strategies: voxel-wise prediction using ROC analysis and prediction by linear regression analysis of lesion volumes. AUC values from ROC analysis in our study were largely in line with previous reports for prediction by early ADC or a combination of early ADC with spatial infarct frequency data in rats.15,36–38 Interestingly, prediction by spatial infarct frequency alone yielded AUCs consistently higher than those determined from early ADC. One shortcoming of our study, which may contribute to this finding, is the fact that we measured T2w lesions 48 hours after MCAO, when significant edema was noted for large infarcts (42±12 μL, ≈20% of the average volume of a hemisphere, for 90-minute MCAO). Infarct prediction by spatial infarct frequency alone as judged by AUC was also superior for 45-minute ischemia, where the average volume of edema was lower (21±12 μL). A factor that might have decreased prediction accuracy by early MRI is inaccuracy of coregistration of early MRIs with T2w images at 48 hours. Of note, an almost perfect correlation has been demonstrated between infarcts determined by T2w MRI at 48 hours and triphenyltetrasodiumchloride histology (r=0.94)39 and T2w images between 48 and 168 hours after MCAO and hematoxylin/eosin histology (r=0.98).40 We found an almost perfect correlation of edema with infarct volume, indicating that edema did not relevantly reduce infarct size prediction by linear regression.
To our knowledge, previous reports have not published AUCs derived form spatial infarct frequency prediction alone (a strategy that does not use early MRI). Our data indicate that prediction via early MRI does not necessarily outperform prediction by spatial infarct frequency, which can be generated from late MR or histology without the need for early MRI. However, as equal infarct volumes are then predicted for all animals, there is no correlation of predicted with measured infarct volumes for this approach.
We found that linear regression of visually determined ADC lesions achieved the smallest root mean squared deviations of predicted from measured infarct volumes for 90-minute MCAO when MRI was performed during MCAO or at 6 hours after MCAO (Figure 5B). This finding was consistent between 2 independent raters blinded to final lesions and experimental group. This indicates that depending on MCAO duration, MRI time point and prediction model, regression analysis of early with late lesion sizes may be superior to voxel-wise prediction by ROC analysis. One possible explanation is that the volume with initially normal ADC that proceeds to infarction as well as the volume of initially decreased ADC that does not proceed to infarction is correlated to final lesion volume. Although ROC would classify these volumes as false-negative or false-positive, thereby decreasing sensitivity, specificity, and AUC, the prediction accuracy by regression analysis may not be relevantly reduced.
Type-I Errors and Early MRI
Our study suggests that depending on MCAO duration and timing of MRI, early ADC lesion volume determination could contribute to avoiding type I errors. We evaluated this by randomly allocating animals of 1 experimental group to 2 groups and testing for a chance finding of a significant infarct volume difference. We found such differences for 5% of possible allocations, as expected at a significance level of 0.05 (Table). For 90-minute MCAO and MRI during MCAO or at 6 hours after MCAO, in about two thirds of such chance findings of final infarct volume differences, early ADC volumes also exhibited significant differences. ANCOVA confirmed an effect of group in only a small minority of these cases and indicated an effect of early ADC lesion volumes in nearly all (Table). Thus, in a treatment study with no effect of the treatment, early MRI would have indicated that chance differences in early ischemic lesion volume had confounded final infarct volume differences in 90-minute MCAO groups with MRI performed during or at 6 hours after MCAO.
In contrast, for 45-minute MCAO, chance findings of final infarct volume differences could not be uncovered by early ADC imaging reliably (Table). One possible explanation may be derived from our 3D analysis of infarct volumes. ADC lesions variably detected in posterior slices during 45-minute MCAO frequently did not exhibit infarctions at 48 hours (Figure 2, C6), in contrast to animals with 90-minute MCAO. Correlation of infarct sizes with early ADC lesions in 45-minute MCAO was also weak in the anterior slices. Thus, variability of transformation of an early ADC lesion into final infarct beyond 6 hours after transient ischemia may play a larger role for 45-minute MCAO than for 90-minute MCAO and this cannot be detected by early ADC imaging.
Implications for Experimental Stroke Studies
Our results do not argue for routine early MRI in experimental stroke studies. However, our data indicate that early imbalances in ischemic lesion volumes are a major source of type I errors for long duration MCAO. Thus, detecting differences in early ischemic lesion volumes may help to reduce these errors and could be used as one tool among others to improve quality of experimental stroke studies. Although imbalances in early ischemic lesion volumes can also contribute to type I errors for short MCAO, our data indicate that other factors may play a larger role. Thus, early MRI is less promising for experimental stroke studies with short occlusion times. When MCAO duration is short enough to leave some animals with no infarction despite successful occlusion of the MCA, early MRI may still be used to confirm successful MCAO. However, confirming successful MCAO may also be done more easily using Laser-Doppler-Flowmetry. Our data show that reduction of variability by exclusion of animals with failed MCAO can be substantial.
In any case, if MRI is performed for other reasons before treatment, ADC imaging may be added and timing may be adapted to enable evaluation of early ADC lesion volumes. When performed during MCAO, the confounding factor of anesthesia necessary for MRI can be minimized by rapid transfer to the MR scanner after MCAO and a short imaging protocol. Finally, early MRI may deliver important information in experimental stroke studies with relevant mortality. In these studies, early ADC can clarify whether deaths were related to infarct size and thus help to improve interpretation of a potential treatment effect in surviving animals.
Weaknesses of Our Study
We started performing the experiments in SV129 mice (group 4), but had to switch to C57Bl/6 mice for logistical reasons. Hence, the results of this group have to be compared with the other groups with care. It has been shown that infarct volumes and cerebrovascular anatomy vary with genetic background11,12 and larger infarcts have been observed in the C57Bl/6 compared with the SV129 strain.41 However, a study by Pham et al found no differences of infarct sizes, ADC, or CBF at any time point between SV129 and C57Bl/6 mice after transient MCAO.42
MRI was performed using isoflurane anesthesia for noninvasiveness and easier handling of animals. Isoflurane can increase CBF13 and reduce infarct sizes. It cannot be excluded that a substantial CBF increase alters the ADC time course and that altered ADC time courses or infarct volumes influence the accuracy of infarct prediction using early ADC measurements. This may increase or decrease prediction accuracy. When applying our protocol to experimental studies, it should therefore be kept in mind that anesthesia may influence the correlation between early ADC lesions and final infarcts.
We measured T2w lesion sizes 48 hours after MCAO onset. Although determination of the infarct volume at this time point enabled low mortality (only 3 mice died before 48 hours after MCAO), significant edema was noted. This may have contributed to inaccuracy of voxel-by-voxel infarct prediction by early ADC. Infarct determination at a later time point, or a more sophisticated coregistration procedure could have reduced this error. Correction of lesion sizes for edema did not significantly alter infarct size prediction by linear regression analysis, as edema volume was highly correlated with infarct volume (r2=0.93).
We used simple prediction models that are easy to implement but may be less accurate when compared with more sophisticated models. Using complex mathematical analyses (artificial neural networks and support vector machines), Huang et al15 demonstrated increased prediction accuracy in rats when combining CBF, ADC, and spatial infarct frequency information. In line with our study, prediction was more accurate for ADC than for CBF and accuracy increased with MCAO duration.38 Desai et al43 showed that prediction could be improved by measuring ADC changes at multiple time points. Wu et al37 and Bouts et al44 found high prediction accuracy in rats using a general linear model combining perfusion and diffusion information. Christensen et al45 have demonstrated increased prediction accuracy of perfusion imaging by more sophisticated CBF parameter analysis.
Our goal was not to achieve the highest AUC for a voxel-by-voxel infarct prediction model. Rather, we aimed at most accurately predicting overall infarct sizes (with accuracy defined as lowest difference between predicted and measured infarct sizes). Our data show that highest AUC does not necessarily translate in minimal mean difference between predicted and measured infarct sizes. In contrast, although infarct prediction by spatial infarct frequency yielded high AUCs comparable with those of sophisticated prediction models for rats,15,37,38 the mean difference between predicted and measured infarct sizes was smaller when prediction was simply derived from linear regression for visually determined early ADC lesions. These arguments notwithstanding improvements in MRI measurement and prediction algorithms can further strengthen the potential for early MRI to improve the statistical evaluation of experimental stroke studies.
Despite standardized procedures in animal experiments, infarct sizes show considerable variability in the intraluminal thread model in mice. This variability is larger in the posterior parts of the brain in C57Bl/6 mice, most likely because of variability of the posterior communicating artery territories. Early MRI may be used to partly control this variability. ADC is the most promising parameter. The accuracy of early ADC lesions to predict final infarcts depends on the timing of MRI and duration of MCAO. Our data suggest that accuracy is highest for MRI during or at 6 hours after 90-minute MCAO. Transient ADC recovery precludes accurate infarct prediction by ADC for a few hours after reperfusion, and infarct prediction is less accurate for shorter occlusion times.
Prediction of final infarct volumes using early ADC measurement may be used to improve power of experimental designs, specifically by detecting imbalances in early ischemic lesion volumes, by excluding animals with failed MCAO from further analysis and by improving the evaluation of treatment studies with relevant mortality.
We thank Marco Foddis for excellent MCAO surgery.
Sources of Funding
This study was supported by the European Union’s Seventh Framework Programme (FP7/2008–2013) under grant agreements no. 201024 and no. 202213 (European Stroke Network), and the German Ministry for Health and Education (Bundesministerium für Bildung und Forschung [BMBF]).
Guest Editor for this article was Miguel A. Perez-Pinzon, PhD.
The online-only Data Supplement is available with this article at http://stroke.ahajournals.org/lookup/suppl/doi:10.1161/STROKEAHA.114.007832/-/DC1.
- Received October 29, 2014.
- Revision received August 25, 2015.
- Accepted September 2, 2015.
- © 2015 American Heart Association, Inc.
- Barone FC,
- Knudsen DJ,
- Nelson AH,
- Feuerstein GZ,
- Willette RN.
- Zhang H,
- Prabhakar P,
- Sealock R,
- Faber JE.
- Huang S,
- Shen Q,
- Duong TQ.
- Dijkhuizen RM,
- Knollema S,
- van der Worp HB,
- Ter Horst GJ,
- De Wildt DJ,
- Berkelbach van der Sprenkel JW,
- et al
- van Lookeren Campagne M,
- Thomas GR,
- Thibodeaux H,
- Palmer JT,
- Williams SP,
- Lowe DG,
- et al
- Bardutzky J,
- Shen Q,
- Henninger N,
- Schwab S,
- Duong TQ,
- Fisher M.
- Berthet C,
- Lei H,
- Gruetter R,
- Hirt L.
- Olivot JM,
- Mlynash M,
- Thijs VN,
- Purushotham A,
- Kemp S,
- Lansberg MG,
- et al
- An H,
- Ford AL,
- Vo K,
- Powers WJ,
- Lee JM,
- Lin W.
- Nagakane Y,
- Christensen S,
- Brekenfeld C,
- Ma H,
- Churilov L,
- Parsons MW,
- et al
- Kitagawa K,
- Matsumoto M,
- Yang G,
- Mabuchi T,
- Yagita Y,
- Hori M,
- et al
- Koumura A,
- Hamanaka J,
- Kawasaki K,
- Tsuruma K,
- Shimazawa M,
- Hozumi I,
- et al
- Jacobs MA,
- Knight RA,
- Soltanian-Zadeh H,
- Zheng ZG,
- Goussev AV,
- Peck DJ,
- et al
- Christensen S,
- Mouridsen K,
- Wu O,
- Hjort N,
- Karstoft H,
- Thomalla G,
- et al