Assessing the Relative Value of CT Perfusion Compared to Non-contrast CT and CT Angiography in Prognosticating Reperfusion-Eligible Acute Ischemic Stroke Patients

In the present study we sought to measure the relative statistical value of various multimodal CT protocols at identifying treatment responsiveness in patients being considered for thrombolysis. We used a prospectively collected cohort of acute ischemic stroke patients being assessed for IV-alteplase, who had CT-perfusion (CTP) and CT-angiography (CTA) before a treatment decision. Linear regression and receiver operator characteristic curve analysis were performed to measure the prognostic value of models incorporating each imaging modality. One thousand five hundred and sixty-two sub-4.5 h ischemic stroke patients were included in this study. A model including clinical variables, alteplase treatment, and NCCT ASPECTS was weak (R2 0.067, P < 0.001, AUC 0.605) at predicting 90 day mRS. A second model, including dynamic CTA variables (collateral grade, occlusion severity) showed better predictive accuracy for patient outcome (R2 0.381, P < 0.001, AUC 0.781). A third model incorporating CTP variables showed very high predictive accuracy (R2 0.488, P < 0.001, AUC 0.899). Combining all three imaging modalities variables also showed good predictive accuracy for outcome but did not improve on the CTP model (R2 0.439, P < 0.001, AUC 0.825). CT perfusion predicts patient outcomes from alteplase therapy more accurately than models incorporating NCCT and/or CT angiography. This data has implications for artificial intelligence or machine learning models.


INTRODUCTION
The pivotal intravenous thrombolysis trials, predominantly conducted more than a decade ago, did not use advanced imaging techniques such as perfusion imaging to identify a treatment target. Consequently, these trials provide limited guidance to the treating clinician when faced with the individual decision making (1). However, acute stroke treatment guided by more sophisticated imaging is being increasingly adopted in routine clinical practice following the positive endovascular therapy trials where the treatment targets were identified with CT angiography ± CT perfusion (2)(3)(4). Modern imaging not only identifies treatment targets such as a vessel occlusion and penumbral tissue (5,6), but can also allow clinicians to develop more sophisticated risk-benefit judgements at the individual patient level. Given recent evidence, the debate regarding imaging of vessels and tissue perfusion has evolved; initially, from whether it should be performed at all in routine practice, and now, to which modern imaging protocols provide the best balance between the decision support information provided and the time taken to acquire and interpret the data. CT angiography was a minimum requirement of the recent endovascular trials in order to identify an occluded vessel and to assess vascular access in patients being considered for endovascular therapy. This technique is ideal for identifying large vessel occlusions, although more advanced (multiple time point) CTA may be necessary to provide additional accuracy in grading occlusion severity (partial or complete) and for identifying collateral blood flow beyond an occlusion. However, CTA does not accurately identify patients with penumbral tissue, and is poor at identifying the infarct core (7). Therefore, "limited" multimodal CT protocols only using NCCT and CTA may lead to futile reperfusion treatment where there is no penumbra to salvage and/or when there is an established infarct core with minimal change on NCCT, which is common (8). Perfusion imaging can positively identify the ischemic core and penumbra, and is now well-validated at identifying patients most likely to benefit from reperfusion. However, perfusion imaging does not identify a vessel occlusion (unless simultaneous dynamic CTA is acquired). However, CTP can quantitatively estimate collateral blood flow by measuring the severity of the perfusion deficit (9). Currently there is a lack of data comparing the usefulness of the various imaging approaches in use to predict outcome of patients treated with intravenous thrombolysis.
The individual patient response to reperfusion treatment is well-documented in ischemic stroke, and modern imaging can identify subgroups of patients with varying treatment responses. As such, there is an enormous potential for an AI based decision support system to be accurate, fast and widely available, if it is shown to be consistently reliable or related to individual patient outcomes. Importantly, the data which is used to generate a prediction for decision support needs to be relevant to the underlying disease process. This highlights that if a more reasonable conceptualization of the underlying clinical problem is used to represent the features for Machine Learning (ML) (i.e., use perfusion imaging to measure ischemia in ischemic stroke patients), the accuracy of models generated by ML will be improved. This is important since the assessments of stroke patients have progressed from using a simple non-contrast CT as part of a diagnosis of exclusion to the routine use of CTA and CTP to positively identify ischemic stroke features such as a vessel occlusion or ischemic brain tissue. These advances have also paralleled advances in patient treatment such as the introduction of thrombectomy. The accuracy of outcome prediction models will also improve, resulting in more accurate outcome predictions that can then inform clinical practice for an individual patient in front of a clinician.
In the present study we sought to assess the ability of standard CT (NCCT) vs. "limited" multimodal CT (NCCT and CTA) vs. multimodal CT (NCCT, CTA and CTP) to identify treatment responsiveness in acute ischemic stroke patients being considered for intravenous thrombolysis in a statistical manner to potentially inform trials around which imaging modality provides the most power. We hypothesized that limited and multimodal CT would identify similar patients as being ideal for reperfusion therapy, and these would be superior to standard clinical and NCCT criteria in this regard.

METHODS
We retrospectivity analyzed consecutive acute ischemic stroke patients presenting to hospital within 4.5 h of symptom onset at 6 centers [(1) John Hunter Hospital, (2) Gosford Hospital, NSW, Australia, (3) Huashan Hospital, Shanghai, (4) the Second Affiliated Hospital of Zhejiang University, Hangzhou, China, (5) Sunnybrook Health Science Center, Toronto, Canada, and, (6) Roya Adelaide Hospital, Australia] between 2011 and 2014 were prospectively recruited for the International Stroke Perfusion Imaging Registry (INSPIRE). As part of this study, patients all underwent baseline multimodal CT imaging with non-contrast CT, CTA and CTP. Clinical stroke severity was assessed at baseline using the National Institutes of Health Stroke Scale (NIHSS). Eligible patients were treated with intravenous thrombolysis according to local guidelines and the clinical judgement of the treating physician. The modified Rankin scale (mRS) was assessed 90 days after stroke. Written informed consent was obtained from all participants, and the INSPIRE study was approved by the various sites' ethics committees.
In addition to the standard clinical and non-contrast CT criteria, CTA and CTP was routinely used as a decision-assistance tool in the process of determining thrombolysis suitability at the INSPIRE sites (10). The CTP criteria for treatment were based on qualitative assessment of the vendor software perfusion maps. If patients demonstrated any of the imaging characteristics listed below, they were considered less favorable candidates for thrombolysis. These were not absolute CTP contraindications to thrombolysis, but, in individual cases, the treating clinician may have chosen to withhold alteplase treatment: 1. Absent or very small perfusion lesion (on transit time maps) 2. An infarct core on CTP (determined by qualitatively low CBV and CBF) larger than 1/3 middle cerebral artery (MCA) territory (or >1/2 anterior or posterior cerebral artery territory), even if NCCT did not show the same extent of early ischemic change.

Imaging Acquisition Protocol
The baseline imaging protocols are shown in Table 1.

Imaging Analysis and Classification of Patients
All imaging was assessed by the INSPIRE core lab. Non-contrast CT scans were granted using the Alberta stroke programme early CT score (ASPECTS) system by two raters, and any disagreement resolved by a third. All perfusion imaging was post processed on commercial software MIStar (Apollo Medical Imaging Technology, Melbourne, Australia). Acute perfusion imaging was processed using single value deconvolution with delay and dispersion correction (11). Previously validated thresholds were applied in order to measure the volume of the acute perfusion lesion (relative delay time, DT >3 s) and acute infarct core (relative CBF <30%) (5). Penumbral volume was calculated from the volume of the perfusion lesion (DT threshold >3 s) minus the volume of the infarct core (relative CBF threshold <30% within the DT >3 s lesion), the volume of severely hypoperfused tissue (DT >6 s) was also recorded. The target mismatch criteria (based on the original DEFUSE2 criteria) (12) was also determined for each patient dataset (core <70 mL, penumbra >15 mL and Mismatch ratio >1.8, severely hypoperfused tissue <100 mL).
All acute CTA scans were analyzed in the core laboratory for occlusion site and severity using thrombolysis in cerebral infarction (TICI) grading system. Occlusion and collateral grade were also assessed by dynamic CTA using middle phase time points and displayed as maximum intensity projection images. This allowed accurate determination of antegrade or retrograde flow beyond an occlusion. We classified baseline vessel occlusion status as either (i) normal = TICI 3, (ii) partial = TICI 2a or 2b, or (iii) complete = TICI 1 or 0. Collateral vessel flow was assed using the 3 point Miteff scale defined as good (3), moderate (2), or poor (1) depending on the extent of contrast visualized distal to a vessel occlusion on computed tomography angiography (13). Collateral ratings and vessel occlusion status on baseline CTA were performed by two stroke neurologists, with any disagreement resolved by consensus with a third stroke neurologist.

Statistical Analysis
Statistical analyses were programmed using Stata v13.0 (StataCorp Ltd, College Station, TX). For this study we excluded patients from the database who were not eligible for alteplase on standard clinical and NCCT grounds. For the remaining patients, who were all clinically eligible for alteplase, we compared clinical and imaging variables between the treated and untreated patients using Kruskal-Wallis and Fisher exact tests where appropriate to identify potential confounding variables and compared the clinical outcomes between groups.
Next, the data was analyzed to determine the accuracy of a predefined imaging modality output at predicting clinical outcome for all patients, regardless of treatment, but with treatment group as a variable. We generated the following overlapping patient models for the purposes of identifying the accuracy of each imaging modality to compare the models using multiple logistic regression: Model AI, the standard CT model, including the variables: NCCT ASPECTS score, patient age, baseline NIHSS, treatment group and time to thrombolysis; and model AII: where NCCT ASPECTS was dichotomized to ≤7 and >7.
Model BI, the limited multimodal CT group, including: collateral grade, occlusion grade, NCCT ASPECTS, patient age, baseline NIHSS, treatment group, time to thrombolysis; and model BII: where occlusion grade was replaced by a dichotomized variable, presence/absence of a complete vessel occlusion and good collateral flow.
Model CI, the CTP group including: CTP ischemic core volume, penumbral volume, patient age, baseline NIHSS, treatment group NCCT ASPECTS, and time to thrombolysis and; model CII: with an additional predefined dichotomized variable of patients fulfilling target mismatch criteria.
Model DI, the multimodal CT group with the variables from models AI, BI and CI. Model DII contained the variables from models AII, BII and CII.
Lastly, linear regression was again performed to determine the treatment effect of alteplase therapy determined as odds ratios for each of the models above (AI, AII, BI, BII, CI, CII, DI, and DII). The assessed outcomes were (1) continuous mRS using ordinal regression; and binary regression for stratified mRS outcomes groups of (2) excellent outcome defined as mRS 0-1, (3) good outcome as mRS 0-2 and (4) poor outcome defined as mRS 5-6. A receiver operator characteristic curve analysis was also performed to determining the area under the curve for each model at predicting the outcomes (1-4).

RESULTS
The INSPIRE database consisted of 2,251 patients, of whom 1,562 were eligible for alteplase based on standard clinical criteria, with the majority of the 689 excluded due to patients being outside the 4.5 h treatment time window or with an uncertain time of stroke onset, established hypodensity on baseline NCCT, or other clinical exclusions (e.g., poor premorbid function, rapidly resolving symptoms; Figure 1). Of the 1,562 patients who were eligible for treatment based on standard clinical/NCCT criteria, 827 (53%) were treated with thrombolysis, and 735 (47%) did not receive thrombolysis after perfusion CT indicated a "less favorable" pattern, and individual patient characteristics were taken into consideration. Baseline clinical and imaging characteristics of the treated and untreated groups are shown in Table 2.

Model (A) Standard Clinical and NCCT Criteria
The AI model had modest accuracy in predicting continuous mRS outcome, and the overall predictive effect of the variables in the model was weak (R 2 0.067, P < 0.001, AUC 0.605). Prediction of the dichotomous mRS endpoints was also similar (mRS 0-1:  Table 3). The coefficients and significance of the individual variables in each model are available in Table 4.

DISCUSSION
In the present study we compared different CT modalities in acute ischemic stroke patients and identified that CTP has the best ability to predict patient outcome, particularly in relation to alteplase treatment. Dynamic CT angiography also had reasonable accuracy in predicting good outcome (but less so in predicting poor outcome) in patients with a complete occlusion and good collaterals. The lower accuracy of the CTA models may reflect the insensitivity of the CTA models at identifying infarction or the absence of penumbra tissue, and that the technique relies on reader interpretation in comparison to the automated, quantified output available from CTP. The range of data available from CTP (rather than simplified grading scores on CTA) enables better prediction of outcomes, in particular, those with larger infarct cores will likely have a poor outcome, and patients without penumbra are unlikely to benefit from reperfusion therapy. Surrogates for infarct core (NCCT ASPECTS) and presence of penumbra (good collaterals) used in the clinical and NCCT or CTA models do not provide the same prognostic information as the models incorporating CTP. The standard clinical criteria for thrombolysis therapy were poor at identifying patients who may benefit from therapy, even when including ASPECTS cut points. This may be due to the poor sensitivity and specificity in identifying patients with a proximal vessel occlusion, established infarct core or penumbral tissue. Importantly, the low predictive accuracy of these standard assessments does not imply that thrombolysis in these cases would be futile. However, it is well-recognized that at the individual patient level, the delivery of intravenous thrombolysis results in a meaningful clinical improvement in less than half the patients treated. Therefore, while these results were not unexpected, they do reflect why previous unselected trials of acute reperfusion therapies have relatively small treatment effects (1), as there was not a positive treatment target identification in the patient assessments.
The scores derived from CTA of course suffer from being qualitative and being prone to interrater variability, compared to standardized and automated output from CTP. These strengths of CTP are evident from this study, which demonstrated consistently higher accuracy at predicting patient outcome with reperfusion therapy, and avoiding potentially harmful or futile treatment (e.g., in patients with lack of target mismatch). While each imaging technique has its strengths and weaknesses, e.g., CTP maps cannot identify and grade severity of a vessel occlusion, many current generation CT scanners are capable of acquiring perfusion and dynamic angiography simultaneously, and so it is unlikely that clinicians will have to choose just one imaging modality for patient assessment in the future. In this study we have demonstrated that the combination of perfusion and angiography leads to a very high prognostic power for patient management and outcome prediction.
One important limitation of our study is the observational design as a source of data collection. A randomized trial of limited vs. full multimodal CT would be ideal to validate our results, where centers are randomized to the type of imaging that they performed. However, the viability of such a trial would be challenging. Variation in the post processing of perfusion imaging may also have considerable impact on study results such as ours. We performed this study using a single, well-validated imaging post processing platform in an attempt to control for this variation. Another issue to consider is the possibility of selection bias introduced prospectively with the use of qualitative interpretation of CTP by clinicians at our INSPIRE centers during the complex clinical decision-making involving patient selection for alteplase therapy. The effect of this bias is likely to result in a larger number of untreated patients who have significant pathology such as a large ischemic core, which would bias the results toward CTP. Next, this study may have been underpowered to show significant relationships with smaller effect sizes in secondary analysis (e.g., time to treatment). Lastly, during this study endovascular therapy was not readily available at our centers, a similar study with such patients may yield different results (14).
In conclusion, we have demonstrated that CTP is the imaging modality best able to identify patients who will benefit from alteplase therapy and those who will not. Surrogates for CTP infarct core or penumbra used in the other models are inferior to CTP, and hence have lower accuracy in predicting patient outcome.

DATA AVAILABILITY STATEMENT
Anonymized data that support the findings of this study are available from the corresponding author upon reasonable request.

ETHICS STATEMENT
The studies involving human participants were reviewed and approved by Hunter New England Human Research Ethics Committee. The patients/participants provided their written informed consent to participate in this study.

AUTHOR CONTRIBUTIONS
AB conceived and design of the study, contributed to the acquisition and analysis of data, interpreted the results, and prepared the original manuscript. CL, TK, KB, and MP participated the design of the study, data collection, and interpreted the results. LL and CC participated in data collection and analysis. XC, RV, NS, and QD participated in data collection and interpreted the results. All authors reviewed, edited the manuscript, and approved the final version.