Atlas-Based Topographical Scoring for Magnetic Resonance Imaging of Acute Stroke
Background and Purpose— The Alberta Stroke Program Early CT Score (ASPECTS), a 10-point scale, is a clinical tool for assessment of early ischemic changes after stroke based on the location and extent of a visible stroke lesion. It has been extended for use with MR diffusion-weighted imaging. The purpose of this work was to automate a MR topographical score (MR-TS) using a digital atlas to develop an objective tool for large-scale analyses and possibly reduce interrater variability and slice orientation differences.
Methods— We assessed 30 patients with acute ischemic stroke with a diffusion lesion who provided informed consent. Patients were imaged by CT and MRI within 24 hours of symptom onset. An MR-TS digital atlas was generated using the ASPECTS scoring sheet and anatomic MR data sets. Automated MR topographical scores (auto-MR-TS) were obtained based on the overlap of lesions on apparent diffusion coefficient maps with MR-TS atlas regions. Auto-MR-TS scores were then compared with scores derived manually (man-MR-TS) and with conventional CT ASPECTS scores.
Results— Of the 30 patients, 29 were assessed with auto-MR-TS. Auto-MR-TS was significantly lower than CT ASPECTS (P<0.001), but with a median difference of only 1 point. There was no significant difference between the auto-MR-TS and the man-MR-TS with a median difference of 0 points; 86% of patient scores differed by ≤1 point.
Conclusion— Auto-MR-TS provides a measure of stroke severity in an automated fashion and facilitates more objective, sensitive, and potentially more complex ASPECTS-based scoring.
Ischemic stroke results from reduced blood flow to brain tissue and can lead to infarction that is detectable by CT or by MR diffusion-weighted imaging (DWI).1 Neurological scoring methods such as the National Institutes of Health Stroke Scale score2 are used to quantify stroke severity to aid in the prediction of patient outcome and assessment of suitability for thrombolytic treatment with tissue plasminogen activator. Efforts have been made to improve prediction from the National Institutes of Health Stroke Scale score by combining it with imaging information such as with the Three-item Scale for the Prediction of Stroke Recovery, which is a composite of National Institutes of Health Stroke Scale, time from onset, and DWI infarct volume.3 The additional metric of DWI volume, however, is only a single feature from the breadth of information that could potentially be provided by MR stroke imaging.
The Alberta Stroke Program Early CT Score (ASPECTS) score is a topographical scoring system, that is, location-based.4 It is a 10-point scoring scale from which point deductions are made based on regional occupancy of an identifiable lesion on CT images.4 A small stroke lesion in a critical location can be far more serious than a large stroke encompassing a less critical region, which is what ASPECTS attempts to account for. The ASPECTS approach can be extended to DWI, because excellent intermodality agreement with CT has been demonstrated.5 Compared with CT, DWI has a higher contrast-to-noise ratio in acute infarction,6 which may facilitate computer automation and development of a topographical scoring method. Computer automation would provide an objective tool for large-scale analyses of data without pre-existing MR topographical score (MR-TS) data. Furthermore, computer automation may reduce interrater variability and slice orientation differences, which have been cited as sources of ASPECTS variability.7
The purpose of this work is to first extend the topographical ASPECTS method to MR DWI images, hereafter called MR-TS. Second, a computer automated (auto-MR-TS) implementation was developed and compared with manual scoring of the MR and CT images (man-MR-TS and CT ASPECTS). In a number of studies, ASPECTS has been validated and proven to show clear benefit7–9⇓⇓ and thus provides a strong clinical basis for our topographical scoring methods.
Magnetic Resonance Imaging
A retrospective study was performed using 30 patients with acute ischemic stroke with an acute DWI lesion. Informed consent was obtained. Patients were imaged by CT followed by MRI. All neuroimaging occurred within 24 hours of symptom onset. All MRI patient data were acquired on a 3-T scanner (Signa VH/i; General Electric Healthcare, Milwaukee, Wis) with a quadrature head coil. DWI images were acquired using a single-shot spin-echo echoplanar technique (b=1000 s/mm2, TR/TE/flip=7000 ms to 9000 ms/73.1 to 93 ms/90°, 192×115 or 144×144 acquisition matrix, 32 cm×19.2 cm or 24 cm×24 cm field of view, and 19 slices, 5-mm thick with a 2-mm gap or 27 contiguous 5-mm slices) to evaluate the infarct. All patients received ASPECTS scores based on their noncontrast-enhanced CT scans.
A 3-dimensional MR-TS digital atlas was generated by manually tracing regions on T1 anatomic data sets (MNI, www.bic.mni.mcgill.ca/brainweb) at a resolution of 1 mm×1 mm×2 mm using the ASPECTS scoring sheet shown in Figure 1.4 These data sets are averaged from 152 T1-weighted data sets acquired from healthy subjects. A published reference of anatomic regions and blood supply territories was also used to guide the region-drawing process.10
Like with ASPECTS, man-MR-TS scores are calculated based on the occupancy of the lesion within the scoring regions.4 There are 12 predefined regions per hemisphere for a total of 24 regions (Figure 1), but lesion occupancy in the anterior or posterior blood supply territories does not warrant deductions with ASPECTS, leaving 10 eligible regions. Points are deducted from 10 where visible occupancy of a scored region within the lesion warrants a point deduction. A stroke fellow (N.S.) with ASPECTS training was permitted to use both the apparent diffusion coefficient (ADC) maps and the raw DWI data for man-MR-TS, because both sets of images would be used in clinical practice.
Auto-MR-TS was performed regionally using ADC maps and the MR-TS atlas based on lesion-region overlay in registered space (Figure 2). This methodology is similar to processing methods applied by other groups for CT11 and MR12 data. The infarct region on the ADC map was defined using a computer-assisted volumetric methodology using region-growing methods that has been previously validated.13 This segmentation method allows for tortuous, or geometrically irregular, regions to be selected. ADC lesions were segmented by (1) selecting a lesion with the mouse cursor; (2) adjusting an intensity threshold; and (3) manually adding or removing voxels to the computer-selected region if necessary (in cases where leakage occurred). As a stopping criterion for (2), an upper threshold was set to 80% of the mean of a region of interest placed in contralateral normal tissue on a slice with clear ADC deficit. Nonlinear registration was performed with SPM2 (Wellcome Department of Imaging Neuroscience, www.fil.ion.ucl.ac.uk/spm, 2004) to register the ADC map and lesion into the MNI brain atlas space. The T2-weighted baseline image with no diffusion weighting served as the source image to obtain the registration parameters that were applied to the ADC map and lesion.
An MR-TS score was calculated from the overlap of the lesion within region territories by computing the intersection between the various regions and lesion data. Sufficient occupancy was required to warrant a deduction based on user-defined occupancy thresholds. These thresholds were intended to control for cases where occupancy was negligible by visual inspection or subject to small preprocessing errors. The region threshold was defined as a percentage of occupancy of the lesion volume that must be attained by a region for consideration, that is, a threshold of 5% dictates that at least 5% of the lesion volume must intersect with a given region to be considered for deduction. The lesion threshold, on the other hand, is defined as a percentage of occupancy of the region volume that must be attained by a lesion for consideration. An absolute threshold was also set based on the total number of voxels in the lesion and region intersection. The region threshold was set to 5%, the lesion threshold was set to 10%, and the absolute threshold was set to 60 voxels (equivalent to 0.12 mL). Only one of the thresholds had to be satisfied to warrant a point deduction. Nonparametric Friedman and Wilcoxon signed rank tests were used to compare auto-MR-TS, man-MR-TS, and ASPECTS (with α=0.05 chosen as the significance level).
Visualization of the (1) ASPECTS regions; (2) overlaid segmented stroke lesion; and (3) 3-dimensional rendering of the lesion were performed using MatLab (188.8.131.527; MathWorks, 2007). For qualitative validation of the MR-TS regions and the segmented stroke, a visualization scheme was developed with overlay of the stroke lesion and ASPECTS regions onto a roadmap image such as the registered ADC map.
One patient was excluded from stroke scoring due to the presence of MR image artifact. Specifically, ghosting was observed in the phase-encoded direction of the diffusion images, which effectively obscured the extent of the lesion on the ADC map so that proper lesion segmentation for auto-MR-TS could not be performed. Demographics and topographical scores are shown in Table 1 for the remaining 29 patients. These patients exhibited either no ghosting or only minor ghosting where the ghosting artifact was deemed to not interfere with the lesion segmentation. All scans were performed within 24 hours as per protocol, except for one patient (Patient 16) whose MR scan was performed 27 hours after stroke onset. The median delay time between CT and MR scans was 4.4 hours.
The median ASPECTS score for the 29 patients was 10 compared with a median of 8 from both auto-MR-TS and man-MR-TS (Table 1). These results indicate that the studied group represents patients with relatively mild strokes, because a threshold of ≤7 has been previously suggested to indicate high stroke severity and associated with poor patient outcome.4
There was a significant difference among the 3 scoring methods as summarized in Table 2 (Friedman test, P=0.0016, by scoring technique: auto-MR-TS, man-MR-TS, and ASPECTS). Auto-MR-TS and man-MR-TS were both significantly lower than ASPECTS (P<0.001), although the median difference was only 1 point between both MR-TS-derived scores and ASPECTS (Table 2). For the 4 patients with large discrepancies (difference ≥3 points), there were clear ADC lesions with changes not seen on CT (ie, ASPECTS=10).
For auto-MR-TS versus ASPECTS, 17 (59%) scores differed by ≤1 point, whereas 6 (21%) scores differed by ≥3 points. Five of the 6 patients (83%) with large discrepancies exhibited clear ADC lesions when no changes had been seen on CT (ie, ASPECTS=10). Thus, the score obtained from auto-MR-TS was consistently more severe than that obtained from ASPECTS. In only one patient was the score from auto-MR-TS less severe; in this case, it was higher by 1 point and the same as that determined by man-MR-TS (Patient 3).
There was no significant difference between auto-MR-TS and man-MR-TS (P=0.12) with a median difference of 0 points, and 25 (86%) scores differing by ≤1 point. Overall, consistent with the man-MR-TS versus ASPECTS comparison, auto-MR-TS yielded a score that was 1 point lower than from ASPECTS (median of 7 points versus 8 points). One score differed by 3 (Patient 13; Figure 3). In this patient, auto-MR-TS determined occupancy in the same regions as man-MR-TS as well as 3 additional adjacent regions (Figure 3). Figure 4 illustrates a case (Patient 28) in which man-MR-TS, auto-MR-TS, and ASPECTS yielded different scores (6, 7, and 8, respectively). Different deductions were made between the MR-TS methods because the lesion fell near the boundaries of a number of adjacent ASPECTS regions. The final scores, however, were in general agreement across methods (maximum difference of 2 points). The high contrast-to-noise ratio in the ADC maps (Figures 3 and 4⇓) facilitated lesion segmentation.
The overlay of auto-MR-TS scoring regions and stroke lesion onto a roadmap provided for qualitative validation of the regions and the lesion as depicted in Figures 3 and 4⇑. The overlay of the lesion allowed for qualitative verification that the scores were being computed correctly and that registration to atlas space was performed satisfactorily. This is seen as a useful adjunct step in verifying the automation process that can be applied to man-MR-TS as well.
This study demonstrated an automated topographical scoring system for MR DWI. The differences between both MR-TS methods and ASPECTS are most likely attributable to the inherent differences in pathophysiologically derived image contrast between CT and MR in infarct assessment. DWI6 and ADC14,15⇓ are more sensitive to the detection of early ischemic lesions suggesting that auto-MR-TS could simply be quantifying the lesion occupancy more sensitively. The high ADC lesion contrast allows for easier lesion segmentation that makes automation of MR-TS easier than for conventional ASPECTS. Furthermore, man-MR-TS and especially auto-MR-TS require negligible training compared with manual CT ASPECTS.16 Acquisition delay between CT and MR presents a confounding factor (median delay of 4.4 hours), although our results are in concordance with a previous comparison of manual MR and CT ASPECTS in which a median difference of 1 point was also observed despite there being a mean delay of only 1.7 hours for that study.5
The median ASPECTS scores were relatively high (8 for both MR-TS methods and 10 for CT ASPECTS), indicating that the studied group represents patients with relatively mild strokes, in which a score of ≤7 has been previously suggested to indicate high stroke severity associated with poor patient outcome independent of thrombolytic therapy (Table 2).4 It is important to note that MR stroke imaging and any associated topographical scoring will find greatest clinical importance if they add value to patients whose strokes are mild or moderate because most severe strokes can be more easily identified directly through clinical evaluation, CT scanning, and/or x-ray angiography.
The National Institutes of Health Stroke Scale score has seen widespread use and has been suggested to provide better prediction of patient outcome compared with lesion volume information acquired from MR stroke imaging.17 Topographical scoring may improve prediction of patient outcome, because it has been found, for instance, that functional outcome (dependence and morbidity) using ASPECTS scoring has proven to have high sensitivity and very high specificity (0.78 and 0.96, respectively). Furthermore, it has been shown that the tandem use of National Institutes of Health Stroke Scale and ASPECTS yielded improved prediction of outcome, better than either scoring technique alone, for patients treated with intravenous tissue plasminogen activator.18 The ASPECTS method has been extensively validated, has been used to manage acute therapy, and has also been used to compare different stroke imaging techniques.5,7–9,19,20⇓⇓⇓⇓⇓ For these reasons, we used the CT-based ASPECTS to develop our atlas-based MR-TS approach.
One limitation of our automated approach is that segmentation of regions demands some operator input. However, this input is minimized by the use of predefined ADC thresholds. Another limitation is that artifact can impinge on proper lesion segmentation. We excluded one patient due to aliasing artifact, although auto-MR-TS may be robust in the vast majority of cases (93% technical success rate in our study). Manual scoring, however, may still be necessary in some cases to distinguish between features arising from artifact or degraded signal information and features arising from stroke. In our judgment, these cases represent only a small portion of examinations and do not mitigate the main impetus for auto-MR-TS. It is important to consider that even if the auto-MR-TS method is not suitable, perhaps due to image artifact, our auto-MR-TS visualization scheme depicted in Figures 3 and 4⇑ can still be useful for guiding a final subjective score. The overlays can assist in arriving at a MR-TS score by indicating the general boundaries of the scoring regions, even when the automatic score is compromised by artifact.
We used region and lesion thresholds for occupancy calculations to control for false-positive deductions. We kept these thresholds as low as possible, at the same time still controlling for cases in which some overcalls seemed present, that is, cases in which occupancy was negligible. In some data sets for example, a large lesion encroached slightly into regions that subjectively appeared to not warrant a deduction. The thresholds also help control for false deductions based on registration and segmentations errors. Although the auto-MR-TS approach is inherently more objective than ASPECTS, it still allows for subjective tuning of threshold parameters both for the lesion segmentation and the occupancy calculations.
Auto-MR-TS may be extended to include augmented versions of ASPECTS such as a version that includes posterior circulation flow called posterior circulation ASPECTS (pc-ASPECTS).21 Note that with basic ASPECTS scoring, only the subcortical and middle cerebral artery blood supply territories are considered for scoring, whereas lesion occupancy in the anterior or posterior blood supply territory regions are disregarded in the score. Another important application would incorporate perfusion imaging to estimate a penumbral volume and an auto-MR-TS mismatch grade, extending on a study that showed how perfusion-diffusion MR ASPECTS mismatch scoring is effective for stroke assessment.9 Further augmentations may now be considered based on the advantages afforded by auto-MR-TS automation steps. The basic ASPECTS methodology had to balance reproducibility (related to ease of use by a manual scorer) and sensitivity (related to how finely the regions were allocated). Auto-MR-TS overcomes problems associated with subjectivity and ASPECTS training, which obviated a more complex scoring scheme.
In conclusion, auto-MR-TS using digital brain atlas mapping facilitates larger-scale studies and provides for an objective, reproducible, whole brain approach to quantitative stroke severity scoring.
Sources of Funding
This work was supported by the Alberta Heritage Foundation for Medical Research (AHFMR), the Canada Foundation for Innovation (CFI), and the Canadian Institutes of Health Research (CIHR). R.K.K. and J.C.K. are supported by AHFMR, Natural Sciences and Engineering Research Council of Canada, and Informatics Cir of Research Excellence Graduate Scholarship awards. A.M.D. is an AHFMR Scholar. R.F. is a Canada Research Chair and an AHFMR Senior Scholar. A.D. has received a research grant from Novo Nordisk Canada.
A.D. has received financial support from Sanofi-Aventis and Boehringer Ingelheim through the Speakers’ Bureau and from Boehringer Ingelheim as a consultant. There are no other conflicts to report.
- Received September 7, 2009.
- Revision received November 30, 2009.
- Accepted December 3, 2009.
- ↵National Institute of Neurological Diseases and Stroke. Stroke: Hope Through Research. 2005. Available at: http://www.ninds.nih.gov/disorders/stroke/detail_stroke.htm. Accessed on August 17, 2009.
- ↵Barber PA, Hill MD, Eliasziw M, Demchuk AM, Pexman JH, Hudon ME, Tomanek A, Frayne R, Buchan AM. Imaging of the brain in acute ischaemic stroke: comparison of computed tomography and magnetic resonance diffusion-weighted imaging. J Neurol Neurosurg Psychiatry. 2005; 76: 1528–1533.
- ↵Pexman JH, Barber PA, Hill MD, Sevick RJ, Demchuk AM, Hudon ME, Hu WY, Buchan AM. Use of the Alberta Stroke Program Early CT Score (ASPECTS) for assessing CT scans in patients with acute stroke. AJNR Am J Neuroradiol. 2001; 22: 1534–1542.
- ↵Demchuk AM, Coutts SB. Alberta Stroke Program Early CT Score in acute stroke triage. Neuroimaging Clin N Am. 2005; 15: 409–419, xii.
- ↵Butcher K, Parsons M, Allport L, Lee SB, Barber PA, Tress B, Donnan GA, Davis SM. Rapid assessment of perfusion-diffusion mismatch. Stroke. 2008; 39: 75–81.
- ↵Kretschmann HJ, Weinrich W. Cranial Neuroimaging and Clinical Neuroanatomy. Stuttgart, Germany: Thieme; 2004.
- ↵Maldjian JA, Chalela J, Kasner SE, Liebeskind D, Detre JA. Automated CT segmentation and analysis for acute middle cerebral artery stroke. AJNR Am J Neuroradiol. 2001; 22: 1050–1055.
- ↵Kosior JC, Dowlatshahi D, Idris S, Alzawahmah M, Tymchuk S, Hill MD, Frayne R, Demchuk AM. Quantomo: validation of a computer-assisted method used in the PREDICT Trial for volumetric analysis of hematoma in intracerebral hemorrhage. Proc 47th Annual Meeting of the American Society of Neuroradiology, Vancouver, BC, Canada, May 16–21, 2009.
- ↵Desmond PM, Lovell AC, Rawlinson AA, Parsons MW, Barber PA, Yang Q, Li T, Darby DG, Gerraty RP, Davis SM, Tress BM. The value of apparent diffusion coefficient maps in early cerebral ischemia. AJNR Am J Neuroradiol. 2001; 22: 1260–1267.
- ↵Pexman JH, Hill MD, Buchan AM, Demchuk AM, Barber PA, Simon JE, Coutts SB. Hyperacute stroke: experience essential when reading unenhanced CT scans. AJNR Am J Neuroradiol. 2004; 25: 516; author reply 516–518.
- ↵Weir NU, Pexman JH, Hill MD, Buchan AM. How well does ASPECTS predict the outcome of acute stroke treated with IV tPA? Neurology. 2006; 67: 516–518.
- ↵Demchuk AM, Hill MD, Barber PA, Silver B, Patel SC, Levine SR. Importance of early ischemic computed tomography changes using ASPECTS in NINDS rtPA Stroke Study. Stroke. 2005; 36: 2110–2115.
- ↵Puetz V, Sylaja PN, Coutts SB, Hill MD, Dzialowski I, Mueller P, Becker U, Urban G, O'Reilly C, Barber PA, Sharma P, Goyal M, Gahn G, von Kummer R, Demchuk AM. Extent of hypoattenuation on CT angiography source images predicts functional outcome in patients with basilar artery occlusion. Stroke. 2008; 39: 2485–2490.