Article Text

Download PDFPDF

Low signal, high noise and large uncertainty make CT perfusion unsuitable for acute ischemic stroke patient selection for endovascular therapy
  1. R Gilberto González1,2
  1. 1Department of Neuroradiology, Massachusetts General Hospital, Boston, Massachusetts, USA
  2. 2Harvard Medical School, Boston, Massachusetts, USA
  1. Correspondence to Dr R G González, Department of Neuroradiology, Massachusetts General Hospital, Boston, Massachusetts, 02114 USA; rggonzalez{at}

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

Introduction: I was wrong

Neurointerventionists are on the cusp of revolutionizing the care of the acute ischemic stroke patient. The stage is set for a major advance after 2 decades of experience and the development of extraordinarily effective endovascular devices. This advance depends on a clear demonstration of improved patient outcomes, but the evidence for improved outcomes after endovascular therapy is weak. This was most dramatically demonstrated by the halting of the IMS III trial ( Identifier: NCT00359424). There is an emerging consensus that the proper selection of patients for endovascular treatment has been lacking. Patient selection using neuroimaging has not been successful, most likely because ineffective neuroimaging methods have been employed.

I have been a strong advocate of using neuroimaging, especially CT perfusion (CTP), for patient selection. CTP is particularly attractive because it can be obtained immediately after non-contrast CT and C without moving the patient. But I was wrong. There is now strong evidence that some neuroimaging methods are effective but others are not. I have learned that the evidence is especially strong for the reliability of diffusion MRI and the unreliability of CTP for measuring the size of the infarct core in the individual patient. In this essay, I review the evidence that led to my current view on the best practices for imaging the stroke patient. My hope is that my education may help neurointerventionists maximally benefit their patients.

Massachusetts General Hospital neuroradiology and interventional neuroradiology consensus conference on advanced neuroimaging of stroke physiology for neurointervention

The greatest opportunity to improve overall outcomes in ischemic stroke is the successful treatment of patients with occlusion of major cerebral arteries. The vast majority of these patients involve occlusion of the proximal middle cerebral artery (MCA), and this condition will be the focus of this essay. Figure 1 illustrates the altered physiology produced by a right MCA occlusion. With occlusion, there is an immediate alteration in cerebral hemodynamics that is simultaneously accompanied by a neurological deficit, the stroke syndrome. Within the involved brain territory two regions arise that are commonly referred to as the infarct core (irreversibly injured brain) and the penumbra (hypoperfused, possibly symptomatic but viable brain). With the passage of time the core will grow into the penumbra. The average rate of core growth is 5.4 ml/h,1 but it can be highly variable, ranging from less than 1 ml/h up to much higher rates.2–4 The rate of core growth depends on the quality of the collateral circulation that is highly variable among individuals.

Figure 1

Middle cerebral artery (MCA) stroke physiology and methods to monitor it. The ischemic core and penumbra created by occlusion of the proximal right MCA changes with time. Shortly after the MCA occlusion, two regions of brain arise. The smaller region is designated as the core that represents irreversibly injured tissue. Surrounding this core is a region of reduced perfusion, the penumbra, which may give rise to significant neurological symptoms. The relative sizes of the core and the penumbra are determined by the quality of the collateral flow. With the passage of time there is shrinkage of the ischemic penumbra and a corresponding growth of the core. With good collaterals, the shrinkage of the penumbra and growth of the core is slow, while with poor collaterals very rapid growth of the core and shrinkage of the penumbra is observed. The physiology of the MCA depicted here may be evaluated using several neuroimaging methods including CT, CT angiography (CTA), diffusion weighted imaging (DWI) and CT perfusion (CTP). The rating of the effectiveness of each method for reliably identifying a part of the stroke physiology is shown on the figure. Class I indicates that there is strong evidence that the method is reliable and beneficial to the patient. Class IIb indicates that there is conflicting evidence on the reliability of the method with a trend suggesting that it is not reliable.

The principle imaging methods available to probe MCA stroke physiology are listed in figure 1. The members of the Neuroradiology and Interventional Neuroradiology Divisions at the Massachusetts General Hospital conducted a consensus conference that critically reviewed the evidence for the effectiveness of each method in the specific situation of MCA occlusion in a patient who is potentially eligible for intervention. Evidence was presented by faculty members that included the published literature as well as a review of 15 years of our emergency department experience in imaging acute stroke. The members of the divisions that did not present evidence formed a review committee that sat in judgment. The review committee subsequently deliberated and produced conclusions and recommendations. The committee rated the efficacy of each method, which are also listed in figure 1. The committee concluded that CTA is highly reliable for the detection of proximal MCA occlusions and was designated class I—that is, the method is reliable and useful for patient care. With respect to the core, three methods were evaluated. Diffusion weighted imaging (DWI) was considered the most reliable method for measuring the size of the infarct core and was designated as class I. Non-contrast CT was also rated as class I, but with the caveat that it is insensitive early in the course of the ischemic process. CTP was rated class IIb—that is, the evidence in support of its reliability for the measurement of the infarct core was equivocal with the weight of the evidence trending against it.

In the specific case of an MCA occlusion, it was judged that penumbral imaging was not necessary for the following reasons. First, the neurological examination provides reliable information on the approximate size of the penumbra. Second, this estimate can be corroborated with the imaging data. In the presence of a proximal MCA occlusion, the core and penumbra are dependent variables inversely related by the collateral circulation. If the size of the core is small, the penumbra will be large, and vice versa. Thus if the occlusion is established by CTA and the size of the core by DWI, then the size of the penumbra may be deduced with a precision that is sufficient for making clinical decisions.

The conclusions reached by the review committee were surprising to me and other members of Massachusetts General Hospital Neuro Divisions. Nonetheless, the evidence compelled us to change our emergency stroke imaging algorithm to reflect the best available evidence. The new algorithm is shown in figure 2. This algorithm has been used for nearly 2 years. It has been effective and it has developed strong support by Massachusetts General Hospital neurointerventionists and Stroke Service neurologists.

Figure 2

Massachusetts General Hospital (MGH) acute ischemic stroke imaging algorithm. All patients who present with a new onset of a significant neurological deficit undergo non-contrast CT (NCCT) scan that is followed immediately by CT angiography (CTA). If hemorrhage is excluded, then head and neck CTA is performed. While the patient is undergoing CT examination, it is determined whether the patient is able to undergo an MRI including that there are no contraindications. If the patient is able and the scanner is available, a diffusion MRI scan (DWI) is acquired. The next step is dependent on the findings on vessel imaging and DWI. If there is an occlusion of a major artery (internal carotid artery, proximal middle cerebral artery or basilar artery), and there is a small diffusion abnormality (defined as <70 ml in an anterior circulation stroke), then the patient proceeds to the interventional radiology suite for endovascular therapy if the patient meets all other criteria for such treatment. If endovascular therapy is not indicated, if the DWI abnormality is large or there is no large artery occlusion, then the patient will proceed to MRI perfusion. CT perfusion is provided to patients who are not able to undergo MRI. Perfusion imaging information obtained by CT or MRI helps to fully delineate the patient's physiology for consideration of other therapies. Patients that are eligible and are within the 4.5 h time limit for intravenous tissue plasminogen activator will receive the treatment in the CT scanner suite before proceeding to MRI, if that is the next step. IA, intra-arterial.

Evidence of poor reliability of infarct core measurement by CTP

The evidence is strong that the size of the infarct core is a major factor in the clinical outcomes of patients with MCA occlusions.5 6 Because of this, reliable determination of the size of the infarct core prior to the intervention is a critical factor in determining the probability of a good outcome if recanalization is successful. The Therapeutics and Technology Assessment Subcommittee of the American Academy of Neurology evaluated diffusion and perfusion imaging in acute ischemic stroke, and published evidence based guidelines.7 The subcommittee rated DWI as class I with level A evidence for the early identification of ischemic stroke. The subcommittee found insufficient evidence supporting the use of perfusion imaging for the same purpose.

While few claim that CTP matches DWI in identifying early ischemic lesions, many hope (as I once did) that it is ‘good enough’ to select patients for stroke treatment. This is understandable because virtually all patients will receive CT and many CTA. It is very convenient to follow this with CTP and avoid an MRI scan. But, this is justifiable only if CTP is reliable; the evidence indicates that it is not. The problem of CTP reliability is illustrated in figure 3A, which displays data recently presented at a national meeting.8 Figure 3 displays data from patients with acute ischemic stroke that were evaluated in our emergency department, and it compares infarct core volumes measured by DWI and CTP. All of the patients had proximal MCA and/or distal internal carotid artery occlusions documented on CTA. They all underwent CTP and diffusion MRI examinations. The figure demonstrates an excellent and statistically significant correlation (r2=0.87) between DWI and CTP measures of the core, but the scatter in individual values is very wide.

Figure 3

Statistical correlation but high variability between different estimates of core infarct size. (A) Ischemic infarct cores measured by diffusion weighted imaging (DWI) and CT perfusion (CTP). Infarct volumes were measured on DWI and CTP derived cerebral blood flow (CBF) images taken from the same patients. All patients had acute ischemic stroke due to occlusion of major anterior circulation arteries demonstrated by CT angiography (CTA). All had CTP studies performed immediately after the CTA, and an MRI afterward. A statistically significant correlation (p<0.01, r2=0.82) was found. However, a large variability between the DWI and the CTP measurements was commonly observed. (B) Interobserver variability of core infarct volumes measured on CTP derived cerebral blood flow images. Two different trained observers measured infarct volumes on CTP derived CBF images taken from the same patients. The patients were from the study depicted in figure 3. All patients had acute ischemic stroke due to occlusion of major anterior circulation arteries demonstrated by CTA. All had CTP studies performed immediately after CTA. While there is a positive correlation between the two measurements, a large variability between the two measurements was common in individual patients.

Figure 3B demonstrates interobserver variability in defining a volume based on CTP cerebral blood flow (CBF) images. The same post processed CBF images were presented to two trained individuals and each was asked to identify the margins of the presumed infarct core on the CBF images. This was then used to calculate volumes. Once again, while a statistically significant correlation is observed, there is wide variability between the two evaluators in their estimates of core infarct volumes in individual patients.

Basic imaging physics explains the unreliability of CTP for infarct core measurement

Imaging is clinically useful when it is capable of differentiating a pathological lesion from normal tissue in individual patients. Imaging physics informs us that the most important factor for distinguishing pathology is the relative imaging signal given by the pathology compared with normal in the presence of background noise. This is the signal-to-noise ratio (SNR) or, to be more precise, the contrast-to-noise ratio (CNR), which is the difference between the SNR of the pathology and the SNR of normal tissue. Understanding the relationship of SNR to the coefficient of variation provides an appreciation of how SNR is linked to measurement error. It turns out that mathematically, the SNR is the inverse of the coefficient of variation. Thus an imaging feature with an SNR of 10 has a coefficient of variation of 10% and is identifiable with ease, while one with an SNR of 1 has a coefficient of variation of 100%, and it may be difficult to identify or precisely demarcate its borders. Uncertainty looms large in images with poor SNR.

The problems associated with the determination of the early core infarct volume by CTP appear to be fundamentally related to the underlying physics. Images derived from CTP data, including CBF and CBV, are ‘noisy’. How noise affects the appearance of features in an image (such as an infarct) is well understood. The processing of CTP source images into derivative images such as CBF requires certain assumptions and multiple steps. While it is beyond the scope of this essay to describe the process, it is important to understand that that each assumption and each processing step adds uncertainty, reflected as noise in the final image. Added noise will reduce the CNR. It is well established that a CNR of 4 or greater allows a feature to be identified 100% of the time. A feature with any CNR of <4 may still be identified, but with less confidence.

The CNR of infarct cores on DWI images are far superior to CTP images. This is illustrated in figure 4. The figure shows the DWI and CTP derived CBF images from a patient with a documented left MCA stem occlusion. Regions of interest of the same size were placed over the center of the infarct core shown on DWI and in a corresponding location on the CBF image. The CNR of the core on DWI was more than 8, while the CNR on the CBF images was <1. Figure 4 displays the same CBF image data using different windows and levels. The top image on the right is displayed with a wide window. The lower two images are displayed using very narrow windows with the window center corresponding to 15% or 45% of the value measured in the normal part of the image. While the outline of the core is easily identified on the DWI, the boundaries are unclear on the CBF images.

Figure 4

Contrast-to-noise ratios (CNR) and distinctness of infarct volume borders on diffusion weighted imaging (DWI) and CT perfusion (CTP) derived cerebral blood flow (CBF) images. Images are from a patient with an acute stroke syndrome and documented left middle cerebral artery stem occlusion. Diffusion weighted image is at top left, while the others are the same CTP derived CBF image with different window settings. The top right image is the CBF image at a wide window. The bottom left CBF image is displayed with a very narrow window with the center set at a level of 15% of the mean signal within a normal appearing region. The bottom right CBF image is displayed with a very narrow window with the center set at a level of 45% of the mean signal within a normal appearing region. A region of interest (roi) was drawn within the DWI hyperintense area and the mean signal intensity and signal SD from within that roi was obtained. The mean signal intensity and signal SD from a roi in the contralateral hemisphere were also obtained. These values were used to calculate the signal-to-noise ratio and CNR. A similar procedure was done on the CBF image. CNR as above 8 for the DWI identified core, while it was <1 in the same region on the CBF image. The much higher clarity of the infarct border on the DWI compared with the CBF images is easily appreciated.

Many practitioners of CTP are well aware that there are problems with the technique. Meetings have been held to consider the issues, and recommendations have been published suggesting ways to improve it by standardizing parameters.9 One trend is to abandon absolute thresholds in favor of relative thresholds of CBV or CBF to distinguish the core from the penumbra. This has failed, as has been well documented in an important meta-analysis by Dani et al.10

The inherently poor SNR of CTP derived images is the fundamental flaw in the technique. Scientists may derive meaningful information from low SNR measurements by repeating a measurement multiple times and calculating the mean. This cannot be done with CTP scans in patients. This averaging effect explains the commonly reported high correlation between CTP derived images and a gold standard such as DWI. Examples of this effect are shown in figure 3. Some investigators extrapolate a high correlation in a population of measurements to high accuracy of the measurement in an individual. This is not valid. We must always be aware that we are treating individual patients and our measurements in that patient must have a sufficiently robust SNR and CNR to justify its use.


The use of CTP in patients with MCA occlusions must justify the added risks and costs of additional exposure to radiation and contrast material. I was wrong in my prior assessment of the value of CTP. Accumulated evidence strongly indicates that the error in the measurement of infarct core volume using CTP images is too high in an individual patient to be clinically reliable. Its use may be lead to erroneous decisions in which a patient with a low probability of good outcomes is treated while another with a high probability of good outcomes is not. The use of CTP for patient selection in clinical trials of endovascular therapy are not supported by the evidence; its use could very likely lead to trial failures unrelated to the true efficacy of the treatment being evaluated.



  • Linked articles 010415, 010416.

  • Competing interests None.

  • Provenance and peer review Commissioned; not externally peer reviewed.

Linked Articles