|Home | About | Journals | Submit | Contact Us | Français|
Magnetic resonance imaging (MRI) has been shown to improve the diagnosis and management of patients with brain disorders. Multiparametric MRI offers the possibility of noninvasively assessing multiple facets of pathophysiological processes that exist simultaneously, thereby further assisting in patient treatment management. Voxel-based analysis approaches, such as tissue theme mapping, have the benefit over volumetric approaches in being able to identify spatially heterogeneous co-localized changes on multiple parametric MR images that are not readily discernible. Tissue theme maps appear to be a promising tool for integrating the plethora of novel imaging contrasts that are being developed for the non-invasive investigation of the different stages of disease progression into easily interpretable maps of brain injury. We describe here various implementations for combining multiparametric imaging and their merits in the evaluation of brain diseases.
Magnetic resonance imaging (MRI) plays important roles in the diagnosis and prognosis of patients with brain disorders ranging from acute conditions such as stroke and traumatic brain injury (TBI) to chronic illnesses such as multiple sclerosis and epilepsy. As a step towards personalized medicine, many have speculated that imaging can be used to develop treatment plans for these patients that take into account individual variation1–5. MRI also offers the potential of noninvasive longitudinal monitoring of disease progression, for example in the case of brain tumors when repeated biopsies can be burdensome. MRI has frequently been speculated to provide potential biomarkers for monitoring the effects of therapies at both pre-clinical and clinical stages, and for optimizing patient selection in clinical trials of novel drugs 6. Further, some investigators have advocated the use of imaging as a surrogate endpoint for long term outcome in clinical trials due to the ease of translation of results between pre-clinical and clinical settings7.
Combining multiple imaging contrasts that reflect different aspects of pathophysiological processes may provide additional insight over what can be gleaned from a single imaging parameter. In addition, individual MRI parametric maps typically present single snapshots in time of disease processes that are highly dynamic. For example, in acute stroke patients, to better assess tissue viability and potential recovery after therapeutic intervention in regions of reduced cerebral blood flow (CBF), extra data on changes in cerebral blood volume (CBV) (e.g. whether it is reduced8 or elevated9) may be critical.
There are several methods for multimodal imaging analysis. Some involve evaluating diverse imaging datasets that are typically acquired at the same imaging session, while other approaches may concentrate on the investigation of longitudinally acquired single parametric maps. One strategy for analyzing multiparametric data involves extracting relevant features, such as volume or signal intensity, and then performing statistical tests or modeling on those factors. Figure 1 demonstrates such an approach in which one compares lesion volumes between acute trace-weighted diffusion-weighted MRI (DWI) and peak time of the deconvolved residue function (Tmax) of perfusion-weighted MRI (PWI). This can be followed, according to one’s hypothesis, by statistical tests such as analysis of variance to determine if there exists a difference in volumes between groups, or multivariate regression to predict an outcome variable. Region of interest (ROI) analysis can also be performed in which one compares signal intensity in different patient cohorts, e.g. treated or non-treated. These volumetric regional approaches have been used to investigate several brain disorders ranging from acute stroke10–12, brain tumors13, migraine14, multiple sclerosis15, dementia16, CADASIL17, Alzheimer’s disease18 and epilepsy19, 20 to name a few. A benefit of a volumetric or ROI based methodology is that one does not need spatially aligned datasets and can concentrate on particular features of interest, excluding a priori noisy data. However, as the number of quantitative novel MRI contrasts continues to increase, thereby increasing the amount of images to interpret, readily identifying regions of abnormal tissue in multitude of images would be cumbersome and unfeasible in acute clinical settings.
Another strategy for performing multimodal analysis is a voxel-wise approach (see Figure 2). In this technique, one transforms quantitative imaging values for each voxel <i,j,k> from each parametric map, Xijk,1…N, to create a single new imaging map T for which the value at each voxel <i,j,k> is a function, , of N parameters. This function may either be a linear or nonlinear. Similar to the volumetric approach, feature selection must also be performed for the voxel-based approach, however, in this case it primarily involves choice of which imaging parameters to combine. The benefit of this method is that one can investigate voxel-wise changes that may otherwise be obscured when taking an aggregate value in an ROI (e.g. mean or median value). This heterogeneity has been shown to occur for water self-diffusivity or apparent diffusion coefficient (ADC) 10, 21–23, and other parameters such as CBF and CBV, which not only vary by tissue pathology24–28 but also by tissue type. Further, it may be the co-localized changes in these parameters that can provide the best insight into which tissue is most severely injured and which tissue is most likely to respond to therapy. Indeed, the heterogeneity in the pathophysiology of human stroke is thought by some to account for the limited success of therapeutic interventions29, 30. Furthermore, integration of lesion location with lesion size was found to create a better estimator of clinical outcome than volume alone31. This weighting of location and lesion load is exactly what talented clinicians do intuitively at the bedside for a variety of brain disorders as they look at the scans of a given patient. Therefore a single image that combines a myriad of potential markers of tissue viability into a simple imaging signature or theme map that can be easily understood and used to assess tissue status accurately and objectively would be an ideal tool for quantifying the impact of new therapeutic interventions, and possibly for assisting the clinical decision-making process. One disadvantage of a voxel-wise approach is that accurate co-registration of the data is necessary. Fortunately, several algorithms and software exist for performing co-registration (e.g. MNI Autoreg, McConnell Brain Imaging Center, Montreal Neurological Institute32; FLIRT, FMRIB Centre, Department of Clinical Neurology, University of Oxford33, 34; Automatic Image Registration, Laboratory of Neuro Imaging, UCLA35, 36), and MRI manufacturers have developed automatic MRI slice positioning protocols (e. g., AutoAlign) that enable the precise and consistent alignment of scans among different individuals and repeated imaging of the same individidual37, 38.
Early advances in terms of integration of multiple MRI datasets to create a tissue theme or signature map have been made in the field of cerebral ischemia. However, there has been increasing interest in the use of theme maps in other brain diseases. We will use primarily the framework of cerebral ischemia and brain tumors to discuss some of the approaches developed for theme map calculation. However, these approaches can readily be extended to other brain diseases.
The imaging data used to create tissue signature models depends on several factors such as the application, the disease to be studied and pertinent collected imaging modalities. Therefore detailed descriptions of particular MRI parametric maps that can be used, e.g. DWI and PWI, will not be discussed in this article. However, we do note that whichever maps are used, one important consideration for theme map generation is whether one should use a “weighted” image (i.e. raw acquired MRI signal intensity data) or a quantitative “parametric” map (i.e. derived maps calculated from raw acquired data). Both “weighted” imaging parameters and “parametric” maps have been found to perform with varying degrees of success, though the use of surface coils in data acquisition have often precluded the use of weighted images due to image inhomogeneities that can lead to incorrect classification. Most techniques normalize imaging parameters with respect to a reference tissue that is stable across scan sessions and subjects, e.g. normal contralesional white matter. This is useful particularly for cross-subject comparisons. Normalization is typically performed as a pre-processing step on the input data prior to training the model or on a post hoc basis to standardize the output results. The risk with this approach is that it assumes the differences between “weighted-images” are linear scaling effects uniformly spatially distributed, which are often not the case, especially when multi-channel array head coils are used. Non-uniformity correction prior to normalization can compensate for some of these artifacts, but not completely. Therefore, it would be preferable to generate models with parametric maps if the equivalent information was available, e.g. T2-maps instead of T2-weighted maps. One of the drawbacks of parametric maps is that longer acquisition times are often required to obtain the data necessary for calculating the quantitative maps.
Tissue theme map calculation can be broadly classified into two categories: supervised or unsupervised algorithms. Unsupervised algorithms use the inherent data heterogeneity within a given dataset to automatically determine distinct features between different tissue classes, e.g. infarcted versus non-infarcted tissue. Supervised algorithms, in contrast, “learn” to distinguish tissue classes using training sets that are representative of data acquired from patients with the same disorder in which the state of tissue injury is known, e.g. measured with histopathology or on follow-up imaging. In supervised learning algorithms, the outcome states are a priori determined and a training dataset is used to develop a model that minimizes misclassification rates in the training data. With supervised algorithms, one needs to be careful not to “overfit” on the training data in order for the model to be applicable to other datasets. “Overfitting” of an algorithm typically occurs when one uses too many variables, such that the variables fit the noise of the particular training data set rather than the physiological process one is seeking to model. The over-fitted algorithm will typically do very well on its training data set, but poorly on new data sets. Consequently, results of supervised learning algorithms need to be cross-validated on an independent dataset in order to evaluate the models’ generalizability. This is used to avoid bias that would otherwise occur if the model’s performance were evaluated on the same data that was used to train the model. Despite the extra overhead, a benefit of a supervised approach is that the output is typically well-defined. With unsupervised algorithms that automatically segment the tissue into different classes or clusters, as additional parameters are added, the model results may become more difficult to interpret 39, or generate clusters that might not have a readily discernible biological meaning. On the other hand, unsupervised algorithms do not need a training dataset and hence are not concerned with overfitting a model to the training data. We will broadly cover examples of both approaches. It should be noted that learning algorithms are an active area of investigation among the machine learning community. Given the breadth and variety of available algorithms, the reader is referred to Duda 39 or Bishop 40 for more details.
One of the earliest examples of theme map generation using imaging for improving our understanding of disease progression was developed by Welch et al 22. Based on data from experimental stroke literature 41, 42, Welch et al formulated an MR-based theme map characterizing the stages of lesion evolution in cerebral ischemia. Six signatures of tissue stages were identified that represented (A) normal tissue with normal ADC and normal T2-weighted MRI (T2WI), (B) tissue that may or may not be salvageable with reduced ADC and normal T2WI, (C) tissue at high risk of necrosis with reduced ADC and elevated T2WI, (D) tissue that is transitioning to necrosis with normal ADC and elevated T2WI, (E) tissue that is irreversibly necrotic with elevated ADC and elevated T2WI, and (F) tissue with established cell necrosis with elevated ADC and normal T2WI. This approach combined both a priori suppositions about biology and broad clustering of imaging parameters. They tested this tissue signature model, derived from the experimental stroke literature, on MRI data collected from eight patients (scanned between 24 to 116 h post-stroke onset). Upper and lower threshold values for ADC and T2WI were chosen based on normal gray and white matter tissue values in the contralateral hemisphere, and used to create thresholds for tissue classification. The thresholds were selected using an iterative supervised approach (Figure 3A). Voxels for which the coregistered values fell into a particular region were labeled accordingly (Figure 3B) and mapped to create a theme map displaying the spatial distribution of the tissue signatures (Figure 3D). Clearly evident is the heterogeneity of the distribution of the signatures in the stroke patient in Figure 3. Welch et al noted that the B signature, which often surrounded signatures C and D, represents tissue that may still be salvageable, while C and D show tissue that is at high risk or transitioning to necrosis. This work demonstrated that using multiparametric maps provide images that correlate with hypothesized stages of disease evolution that would not be detectable using standard ROI analysis approaches, since it was the interaction of the two parameters that were important for staging the tissue’s pathological state.
A similar multiparametric approach was used by Moffat et al to analyze a single imaging modality, ADC, across multiple time points for stratification of brain tumor response to therapy43. Moffat et al created functional diffusion maps (fDM) that mapped changes on co-registered ADC maps acquired in patients with malignant brain tumor, before and three weeks after initialization of therapy (radiation and/or chemotherapy). Patients were assigned a green signature if there was no change in ADC, a blue signature if ADC decreased (assumed to be an indication of increased cellularity or water redistribution without significant cell lysis) and a red signature if ADC increased (representing cells undergoing necrosis or apoptosis). Thresholds for whether a change in ADC was significant or not, were determined from a training set composed of data from five other patients. The green signature was thought to represent tissue resistant to therapy, while the other signatures were associated with partial response to therapy as verified on follow-up imaging four weeks after therapy. The percent of the total volume that was assigned red or blue signatures was found to be 100% sensitive and specific in identifying which patients would exhibit partial recovery from those with stable or progressive disease, in a group of 20 patients, at only 3 weeks into therapy. A later study in a cohort of 60 patients with high grade glioma receiving radiotherapy, demonstrated that fDM was also predictive of overall survival, with patients exhibiting increases in ADC to be most likely to be alive at 1 year44. Another study used the same approach to map changes in relative CBF and CBV values (normalized with respect to contralesional white matter) before and after therapy in 44 high grade glioma patients undergoing radiotherapy, generating parametric response maps (PRM), a generalized form of the fDM 45. The PRM was found to be more predictive at 1 and 3 weeks of survival at 1 year than examining percent change in CBF and CBV using ROI analysis in the tumor volume. This was attributed to the heterogeneity of blood flow and blood volume between normal gray and white matter and diseased tumor tissue which ROI analysis obscured. In addition to early prediction of lack of effectiveness of therapy, a multiparametric spatial mapping technique can potentially be used for guiding therapies that can be targeted at specific tissue 46.
Wu et al used a generalized linear model (GLM) to predict tissue infarction for stroke patients given standard of care (N=14) 47, that is non-thrombolytic or catheter-based therapy. Tissue outcome was modeled as a binary variable, for which 1 represented infarcted tissue and 0 represented non-infarcted tissue in the training data set. The output of this algorithm was the probability, P, of tissue belonging to the infarcted or non-infarcted class given acute imaging presentation. P was a combination of input maps, x, which consisted of acute DWI, ADC, T2WI, CBF, CBV and MTT normalized to contralateral tissue on a voxel-wise basis utilizing a logistic function:
The predictor, η(x), was a linear function of x:
for which β was a vector of calculated coefficients and α was the bias or intercept term. The coefficients α and β that characterized the GLM were calculated from the training data group using an iterative re-weighted least squares algorithm. The calculated coefficients were then used to make predictions for new datasets using Equation 1 and Equation 2. Since additional terms may produce a better mathematical fit that yields no additional biological insight by overfitting to the training data, the decision to include additional parameters or not was based on the Akaike Information Criterion (AIC), a measure that was a function of both training error and complexity 40. Due to the limited number of patients in the study, instead of cross-validation in which a second dataset is used to evaluate the generalizability of the model, the predictive performance was evaluated using jack-knifing or a leave-one-out approach in which the coefficients for each patient were calculated using data from the other patients in the group as training data 48. This approach was found to more accurately predict tissue infarction by follow-up than any single imaging modality when thresholding the probability map at its optimal operating point. An example risk map generated using this approach is shown in Figure 4 along with input imaging parameters. Rather than using a hard classification of tissue with a probability greater than a single threshold, e.g. 50%, as expected lesion (e.g. a binary mask consisting of values of 0 or 1), one can also interpret the maps as the likelihood of tissue dying without interventional therapy (e.g. a continuous variable from 0 to 100%). An alternate interpretation is that the output represents degree of existing tissue injury at the time of imaging. For example, for Figure 4 one sees that the infarction risk map shows a heterogeneous region of injured tissue, with regions of higher infarction risk (red) matching areas of low ADC and increased trace DWI, consistent with regions of the infarcted core. Notably the predicted lesion is smaller in size than the MTT abnormality, often assumed to represent potentially salvageable tissue, but corresponds well with tissue that was found to be infarcted on follow-up imaging. The probability values in these regions are notably lower than in the “core” region. Again, if one used just a simple volumetric mismatch approach between acute DWI and acute MTT, the heterogeneous distribution of tissue risk would have been obscured.
Another benefit of the GLM approach is that the results have a straightforward interpretation independent of the number of imaging input parameters, namely the probability that the tissue will be classified as abnormal. Figure 5 shows an example output including Tmax as an additional parameter. The output appears very similar to Figure 4 and can be interpreted in the same manner. Both overlays are displayed showing only tissue with P values > 50%. It appears that for the particular patient in Figure 4, the inclusion of Tmax did not significantly change lesion classification despite the volume with abnormal Tmax in this patient being much larger than the areas with abnormal MTT and predicted high infarction risk. However, in a study of 80 patients, the inclusion of Tmax, on average, significantly improved the predictive performance of the model over using the original six parameters49. The GLM approach was also used to evaluate differences between patients given standard of care (N=11) and those given thrombolytic therapy (N=27) 50. The authors showed that patients who were treated with thrombolytic therapy and experienced early reperfusion had smaller lesion volumes than were predicted by the GLM, resulting in lower accuracy than in those patients not given thrombolytics. This suggests that the GLM can be used to evaluate therapeutic efficacy of new drugs by comparing actual outcome with expected outcome. Deviations from predicted outcome would be an indication of the treatment affecting its biologic target, which for certain diseases may be regionally specific, e.g. neuroprotective drugs that selectively targets gray matter in potentially salvageable tissue around a core of infarcted tissue.
One limitation of the GLM is the assumption of linearity, which may not be correct for all input parameters. For example, both abnormally high ADC values, and abnormally low ADC values, are predictive of tissue infarction 10. Both low and high CBV have also been observed in the acute stage in areas that became infarcted on follow-up 51. Therefore, models such as generalized additive models (GAM) 52, the functional form of which is derived from their data, have been speculated to potentially provide a better estimate of infarction state over models that assume an a priori linear dependence, such as the GLM 53.
Rose et al developed a parametric normal classifier 54 to predict tissue outcome in which the output was represented as two classes: infarcted or non-infarcted55. The input consisted of an eight-parameter vector X in voxels from abnormal MTT tissue: DWI, relative DWI, CBF, relative CBF, CBV, relative CBV, MTT, and relative MTT. Relative values were calculated with respect to contralateral hemisphere. The model was trained on 10 datasets to produce probability distribution functions assuming normal distributions, and validated on 7 datasets for the purposes of cross-validation. For the 10 patients used to develop the training data, jack-knifing was used for validation. These models were found to be able to predict lesion growth. This approach was also used for tissue classification using ADC, CBF index, T2, and proton density imaging data from an experimental stroke model, however in this case there were 3 output classes consisting of abnormal and normal tissue and cerebral spinal fluid (CSF) 56.
There are also approaches that allow for improved prediction but at the cost of more complexity in the modeling of the output as a function of the input images. The hierarchical mixture of experts (HME) is a tree-structured architecture consisting of gating and expert networks 57, the gating networks of which select the appropriate expert based on the input data. The HME was shown to improve prediction of tissue outcome compared to the GLM 58. Non-parametric approaches such as k-nearest neighbors (kNN) algorithm have been proposed for combining multivariate data for both stroke56, 59 and for tumor patient evaluation 60. The kNN algorithm, which does not have a previous assumption regarding data distribution, classifies each new input vector according to the label of its k-nearest neighbors in the training data, in which distance is often measured with a Euclidean metric. For a human stroke study (N=14) 59, the label was “infarcted” or “non-infarcted” tissue on follow-up imaging (more than 5-days from onset), while for the tumor study in grade IV tumor patients (N=26) 60, the label was “enhanced” or “non-enhanced” tissue on post-radiotherapy imaging (within one month of conclusion of therapy) for tissue that was originally non-enhancing. The number of labels can be more than two, as was the case for the experimental stroke study for which the labels were normal, abnormal, or CSF 56. For predicting tissue infarction in stroke patients, kNN was shown to perform similarly to the GLM for input vectors consisting of normalized acute DWI, ADC, T2WI, CBF, CBV and MTT 59, using a jack-knifing for cross-validation. The stroke study 59 used features that were determined by previous studies to be useful 47. For predicting tissue enhancement in tumor patients, the input vector consisted of 38 features including fractional anisotropy, ADC, CBV, contrast agent bolus curve shape parameters, and proton spectroscopic data (spectral heights and width) acquired 2 weeks before initiation of therapy. Here, the authors weighted the features using genetic optimization due to the known sensitivity of the kNN to parameter scaling and noise from numerous features included in the model. The genetic algorithm demonstrated that using only 7 of the 38 features could produce an accurate prediction of post-treatment contrast-enhancement, using a leave-one-out approach for validation.
An example of a fully unsupervised classification approach is the K-Means (KM) algorithm. The basic KM algorithm clusters voxels by minimizing intracluster distance metrics, in which the number of clusters, K, is a priori specified. Distance metrics that have been typically used include Euclidean or Mahalanobis distance metric. KM has been applied to both lesion prediction in experimental stroke models 61 and differentiating benign and malignant lesions in human patients62. Carano et al showed in an experimental stroke model that using KM to cluster acute quantitative T2 maps, ADC, relative CBF and proton density signal into three classes (CSF, normal and abnormal tissue), produced results that better correlated with post-mortem histological findings than results based on a simple ADC threshold. Srinivasan et al used KM to segment ADC into two (high, low) or three (high, low, intermediate) clusters in patients with head and neck lesions and found a statistically significant difference in the low clusters between malignant and benign lesions 62.
A limitation of KM clustering is that it requires that the number of classes (K) is established prior to processing. The Iterative Self Organizing Data Analysis (ISODATA) technique extends the KM algorithm by dynamically adjusting the number of classes based on distance metrics between centers of different clusters (inter-cluster) and between members of the same cluster (intra-cluster) 63, 64. Clusters with centers that are relatively close will be combined, while a single cluster with members that are relatively far apart will be split. This is repeated until convergence measured by cluster variances within a threshold is achieved or an iteration limit is reached. Because of its lack of reliance of a priori determined number of clusters, ISODATA can be used to show heterogeneity of the values of a parametric map at a single time point. Figure 6 illustrates ISODATA classification procedures. Nagesh et al used an ISODATA approach to combine four trace DWI data sets acquired at b-values of 0, 300, 600 and 900 s/mm2 to examine heterogeneity of ADC in nine stroke patients imaged within 10h of stroke onset 21. They found multiple signatures within ischemic tissue that corresponded to reduced, pseudo-normal and elevated ADC. ISODATA was also used to combine data from multiple parametric maps. Jacobs et al65 used this approach to characterize tissue over time in ten stroke patients scanned at three time points: acute (<= 12 hours), subacute (3 to 5 days), and chronic (3 months). Imaging parameters to the ISODATA segmentation algorithm consisted of two T2WI acquired at echo times of 30 and 90 ms, 1 T1 weighted image (T1WI), and two DWI acquired at b-values of 600 and 1000 s/mm2. The volume of ischemic tissue damage, as determined by acute ISODATA, was highly correlated with the follow-up clinical outcome of the patient. Moreover, features that were detectable on the combined ISODATA map were not evident on the individual DWI, ADC or T2 maps. This suggests that ISODATA can be used to accurately and non-invasively characterize the evolution of ischemic tissue damage. Another study found that the acute ISODATA determined lesion correlated with the acute PWI abnormality in stroke patients seen within 36 h from stroke onset (N=11) 66.
One limitation of ISODATA is the need for post-hoc interpretation of the clusters. Soltanian-Zadeh et al have proposed an extension to ISODATA that orders the clusters fo standardization of output results 67. The approach consists of creating an MRI score that reflects the Euclidean distance between each ISODATA cluster, i, and normal reference tissue signature, EDi,Normal, and CSF signature, EDi,CS, in which the MRI score is the integer part of
This standardization allows one to quantitatively compare ISODATA images and evaluate degree of ischemic injury across studies (see Figure 6D). Mitsias et al showed that patients with lower MRI scores were likely to demonstrate at least 20% smaller lesions on follow-up imaging than originally identified on acute ISODATA maps in a study of acute stroke patients seen within 24 h of onset (N=48), with a subset receiving thrombolytic therapy (N=15) 68. Another study used DWI and normalized ISODATA mismatches in stroke patients (N=45) to predict tissue expansion at 3 months 69. Other improvements to the ISODATA algorithm include spatial contiguity constraints and use of the Mahalanobis function in place of the Euclidean function for distance metric 70.
Tissue theme maps have much promise in increasing our understanding of the mechanisms of brain tissue’s response to injury and to therapeutic intervention. These maps have the benefit of providing a metric that can be readily translated from pre-clinical lab bench settings to clinical bedside settings. The inherent spatial resolution in tissue theme maps can be used to track the unique evolution of different tissue regions in response to therapies. By being able to identify early on whether a therapy is effective or not, e.g. affecting its expected biologic target, a prospectively determined ineffective treatment can be stopped and replaced with one that may show a response, for example, in the case of brain tumors. Alternatively, theme maps have the potential to noninvasively identify which patients are most likely to respond to therapeutic intervention, e.g. by demonstrating presence of still salvageable tissue after stroke, thereby improving triage of patients and patient management. Although examples presented here are primarily from the stroke and neuro-oncology field, there is little doubt that similar tissue theme map approaches can be extended to study other brain disorders. Further, continuing advances in both development of new MRI contrasts and algorithms for combining these parameters suggest that the potential of theme maps to assist in the diagnosis and prognosis of brain disorders have only just been broached.
Sources of Support:
OW: The author is supported in part by the National Institutes of Health (NIH), grants R01NS59775, R01NS063925, P50NS051343, R01NS051412, R21HL089065 and NCRR [P41RR14075]. The contents of this manuscript are solely the responsibility of the authors and do not necessarily represent the official views of the NIH.
RMD: The author is supported in part through funding from the Royal Netherlands Academy of Arts and Sciences, and the European Union’s Seventh Framework Programme (FP7/2007-2013) under grant agreements n° 201024 and n° 202213 (European Stroke Network).
The study is supported in part by the National Institutes of Health (NIH), grants R01NS59775, R01NS063925, P50NS051343, R01NS051412, R21HL089065 and NCRR [P41RR14075] and funding from the Royal Netherlands Academy of Arts and Sciences, and the European Union’s Seventh Framework Programme (FP7/2007-2013) under grant agreements n° 201024 and n° 202213 (European Stroke Network). The contents of this manuscript are solely the responsibility of the authors and do not necessarily represent the official views of the NIH. The authors would like to thank Robert E Irie for manuscript review.