Objectives Automated CT perfusion mismatch assessment is an established treatment decision tool in acute ischemic stroke. However, the reliability of this method in patients with head motion is unclear. We therefore sought to evaluate the influence of head movement on automated CT perfusion mismatch evaluation.
Methods Using a realistic CT brain-perfusion-phantom, 7 perfusion mismatch scenarios were simulated within the left middle cerebral artery territory. Real CT noise and artificial head movement were added. Thereafter, ischemic core, penumbra volumes and mismatch ratios were evaluated using an automated mismatch analysis software (RAPID, iSchemaView) and compared with ground truth simulated values.
Results While CT scanner noise alone had only a minor impact on mismatch evaluation, a tendency towards smaller infarct core estimates (mean difference of −5.3 (−14 to 3.5) mL for subtle head movement and −7.0 (−14.7 to 0.7) mL for strong head movement), larger penumbral estimates (+9.9 (−25 to 44) mL and +35 (−14 to 85) mL, respectively) and consequently larger mismatch ratios (+0.8 (−1.5 to 3.0) for subtle head movement and +1.9 (−1.3 to 5.1) for strong head movement) were noted in dependence of patient head movement.
Conclusions Motion during CT perfusion acquisition influences automated mismatch evaluation. Potentially treatment-relevant changes in mismatch classifications in dependence of head movement were observed and occurred in favor of mechanical thrombectomy.
- CT perfusion
Data availability statement
The datasets generated and/or analyzed during this research are available from the corresponding author upon reasonable request.
Statistics from Altmetric.com
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.
CT and MRI perfusion imaging is a well-established treatment decision tool in acute ischemic stroke.1 For instance, a perfusion analysis with a predefined target mismatch profile (eg, ischemic core <70 mL; mismatch ratio ≥1.8; tissue at risk ≥15 mL) was found to identify patients benefiting from mechanical thrombectomy (MT) even in extended time windows.2 Using a similar perfusion imaging based mismatch criterion, the WAKE-UP investigators identified patients who still benefit from intravenous thrombolysis in extended time windows.3
However, possible errors in automated mismatch calculation could lead to imprecise treatment decisions at the individual patient level outside of such randomized controlled trials. Automated mismatch estimation remains a surrogate for the underlying ground truth perfusion with multiple model assumptions and sources of error. They include the CT acquisition protocol, scanner noise or respectively denoising, the influence of head movement and the choice of deconvolution versus non-deconvolution approaches.4 Even the optimal parameters that represent ischemic core and penumbra and the influence of the post-processing software are still under debate.5–7 In clinical decision-making, the physician should be aware of these problems. Error bars or confidence intervals for the mismatch ratio, which are so far unknown, should be considered.
While it may be possible to analyze the influence of the post-processing software within the clinical setting,8 neither the underlying ground truth perfusion can be deducted in this setting, nor can noise-free images be acquired. Also, absolute control of the patient’s head movement is not achievable. One possibility to address these shortcomings is the analysis of simulated perfusion data.9
In the present study, we therefore sought to assess the reliability of the mismatch calculation depending on the mismatch ratio using a realistic CT brain-perfusion-phantom. By adding real measured CT noise and artificial head movement to the data, we examined the impact of two potential sources of error, which are hardly known or controlled for in clinical practice. Primary outcome measures were differences between evaluated and ground truth simulated infarct core volumes, penumbral volumes and mismatch ratios. The secondary outcome measure was a change in mismatch classification according to the DEFUSE-3 target mismatch criteria.
Material and methods
Part 1: phantom data
All simulations were carried out using Matlab (Matlab R2018b, Mathworks, Natick, MA). Realistic, noise-free, CT perfusion data were simulated by adopting an established hybrid digital brain perfusion phantom.10 11 It combines anatomical details as obtained from real MRI with dynamic tissue-attenuation curves (TAC) as simulated by convolution following the indicator-dilution theory.12 The phantom’s applicability to simulate realistic CT perfusion data was shown by Riordan et al 10
A detailed insight into the indicator-dilution theory and deconvolution-based perfusion analysis is given by Fieselman et al.13 Briefly, we consider the passage of a bolus of non-diffusible contrast agent given at a time t=0 to the feeding arteries of a volume-of-interest. The contrast agent will follow different paths through the vascular bed with a characteristic distribution of transit times (described by the probability density function ) and a resulting specific mean transit time (MTT). The tissue residue function now describes the fraction of contrast agent present within the volume-of-interest at a given timepoint t≥0:
Within this study, the residue function was simulated to be for t<MTT followed by a mono-exponential decay. Such a model assumes that the contrast agent remains strictly intravasally, that is, no blood–brain barrier disruptions are evident.14 Following the indicator-dilution theory, the contrast agent concentration within the volume-of-interest is then given by:
where is the concentration of contrast agent in the feeding arterial vessel, CBF is the cerebral blood flow, and ρ is the tissue density (in our study approximated by 100 g/100 mL brain parenchyma). The correction factor accounts for differences in hematocrit between large vessels and capillaries (in this study set to the quotient of and ) and the fact that the contrast agent dilutes only within the plasma.10 When introducing the flow scaled residue function , equation (2) can be written as:
Thereby, describes the main parameters used in CT perfusion analysis (scaled by the above named correction factors): CBF is the maximum of the flow scaled residue function, CBV (cerebral blood volume) is the area under the flow scaled residue function, MTT is the first moment of the function, and Tmax is the time to the maximum of the function. To simulate the data, the TAC of the arterial input function has been obtained from real CT perfusion data as described before.10 The parameters for CBF and CBV are obtained from the literature (see online supplemental table 1). MTT followed according to the indicator-dilution theory by .
In our study, four different tissue compartments were considered (healthy tissue, oligemia, penumbral tissue, and infarct core) to simulate seven different realistic stroke scenarios within the left middle cerebral artery (MCA) territory. The scenarios ranged from a small infarct core within the basal ganglia (9 mL of infarct core, mismatch ratio of 11.6) up to a large infarct core (73 mL of infarct core, mismatch ratio of 1.4). Figure 1 gives an overview of the simulated scenarios. The parameters for CBV and CBF are given in the online supplemental. To account for Tmax being a parameter of mainly bolus delay,1 the TACs were shifted additionally by 2, 5 or 7 s for oligemia, penumbra and infarct core, respectively. Macrovascular collateral supply was simulated by delayed (4 s) and reduced contrast filling (25% compared with normal arterial vessels) of the MCA branches distal to the MCA M1 segment occlusion. In a first step, perfusion data were simulated as isotropic with 1 mm3 voxel size at a repetition rate of 0.1 s.
Next, artificial motion was applied by rotating these simulated volumes around all three orthogonal axes. Thereby two scenarios were considered. For subtle head movement, rotational angles were chosen randomly in between −0.1°≤α≤0.1°, whereas for strong head movement, random angles lay in between −0.25°≤α≤0.25°. Hence, for a given slice acquisition number n the resulting rotation is given by with . To respect fixation of the patient’s head within the head holder, this sum was restricted to a range of maximal ±2° for subtle and ±4° for strong head movement. Afterwards, data were reconstructed at 5 mm slice thickness (ST) and a resolution of 512×512 voxels in x-/y-plane.
Finally, realistic noise was acquired with a CT perfusion scan of a cylindrical, water filled phantom. As in the clinical setting, the scan was obtained using a 64-slice CT scanner (Siemens Somatom, Siemens Healthineers, Erlangen, Germany) at 180 mAs, 80 kV with a z-coverage of 8 cm (resulting in 16 slices with ST of 5 mm). In total, 30 full acquisitions were acquired at a repetition rate of 1.6 s, leading to a total of 480 noise samples. With a mean of 0 HU (Hounsfield units) (water) and an SD of ±10 HU, these noise samples could then be combined with the simulated perfusion data (each noise sample with the simulated data for the respective slice acquisition time point). Figure 2 gives an overview of the simulation workflow.
Perfusion analysis was carried out using RAPID (RAPID, iSchemaView, Golden, Colorado, USA), a widely established, US Food and Drug Administration (FDA) approved software for the processing of acute stroke perfusion imaging.2 It offers a fully automated CT stroke workflow including denoising of the CT data and application of automated movement correction and segmentation. The arterial input function is detected automatically and, according to equation 4, deconvolution is applied to obtain parameter maps of Tmax, CBV and CBF. The mismatch ratio is calculated by volumetric analysis of the volume with Tmax >6 s and the infarct core with relative CBF <30%.
Part 2: real CT perfusion data
To test the transferability of the simulation results to the clinical setting, we analyzed how often such head movement occurs in real, acute stroke CT perfusion imaging. Second, the impact of artificial in-plane head movement on the mismatch ratio was also investigated in real, clinically obtained CT perfusion data. Therefore, imaging data of all patients who underwent CT perfusion imaging before mechanical thrombectomy for acute occlusion of the MCA M1 segment between January 2017 and June 2020 at our institute were processed with RAPID. The occurrence of patient head movement was then categorized from the RAPID analysis report. In concordance with the above described simulations, head movement was classified as not relevant (detected head rotation <1°), subtle head movement (−2° <detected maximal head rotation <2°, rotation per full temporal acquisition <1°) and strong (rotation <−2° or >2°).
Next, artificial head movement was added to the raw CT perfusion data for patients with prior minimal, not relevant head movement. Due to the slice thickness of 5 mm (see above), only in-plane movement could be added to the clinical data. Artificial in-plane translation was simulated by a random x-/y-translation of the patient’s head in between −1 mm ≤translation ≤1 mm every new slice acquisition (restricted to a maximum translation of ±2 mm for subtle movement and ±5 mm for strong movement). Further, in-plane rotation along the z-axis was added as described above. Afterwards, data were reprocessed with RAPID.
All statistical analysis was carried out using R* (The R Project for Statistical Computing, V3.1.2). Differences between simulated (or original in the case of real CT perfusion data) and evaluated infarct core volume, Tmax >6 s volume and mismatch ratio were assessed by Bland-Altman statistics. Linear regressions were carried out to identify whether evaluated volumes or respectively mismatch ratios correlate linearly with ground truth. Statistical significance was reached at p<0.01. Confidence intervals are given as 95% CI.
Phantom data: impact of CT scanner noise on mismatch evaluation
When comparing ground truth simulated volumes with evaluated volumes, we found that CT scanner noise alone had only minor effects on the mismatch evaluation. In detail, linear correlation between ground truth and evaluated infarct core size was given (r2=0.99, p<0.001), as well as between simulated and evaluated mismatch ratio (r2=0.99, p<0.001 (cases with undefined mismatch ratios due to no detected infarct cores excluded)). Bland-Altman statistics revealed a mean difference of −4.1 (−11.2 to 2.9) mL between simulated and estimated infarct core, −1.4 (−12 to 8.8) mL between simulated and evaluated Tmax-lesion size, and +0.4 (−0.9 to 1.6) between simulated and evaluated mismatch ratio. This minor tendency to overestimate mismatch ratios was evident, especially for small infarct cores and respectively large mismatch ratios. One critical classification was observed for scenario E (table 1). While ground truth simulated mismatch ratio barely met the DEFUSE-3 mismatch criterion with a simulated mismatch ratio of 1.8, the evaluated mismatch ratio was found to be 1.7.
Phantom data: impact of head movement on mismatch evaluation
The addition of subtle movement to the perfusion data led to a mean difference between simulated and estimated infarct core volume of −5.3 (−14 to 3.5) mL, to a mean difference between simulated and estimated Tmax-lesion size of +9.9 (−25 to 44) mL, and consequently to a mean difference between simulated and evaluated mismatch ratio of +0.8 (−1.5 to 3.0). While linear correlation between simulated and evaluated infarct cores (p<0.001, r2=0.99) and mismatch ratios (p=0.001, r2=0.94) remained present, one clinically critical misclassification was observed, where a patient would have been classified as benefitting from MT while he was not eligible to benefit from this therapy according to the ground truth mismatch profile (table 1, scenario F). This effect became even more evident when strong movement was added. In this case, the mean difference was −7.0 (−14.7 to 0.7) mL for the infarct core and +35 (−14 to 85) mL for the Tmax-lesion. Consequently, the mean difference between simulated and evaluated mismatch ratio was +1.9 (−1.3 to 5.1). While a strong linear correlation between simulated and evaluated infarct cores was still present (p<0.001, r2=0.97), linear correlation between simulated and evaluated mismatch ratios was weaker (p=0.01, r2=0.82) (see figure 3). Concerning the DEFUSE3-mismatch criteria, we observed clinically relevant misclassifications in two of seven scenarios (scenario F and G, both in favor of MT). Importantly, there was no clinically relevant underestimation of mismatch ratios.
Real CT perfusion data
Of 57 patients who met the inclusion criteria, CT perfusion data were evaluable in 53 patients (92%, two scans were excluded due to severe bolus delay of >25 s, in one patient CT perfusion acquisition was incomplete, and in one patient a segmentation error occurred due to acute subdural hematoma, which was wrongly classified as infarct core). In 23 of the remaining 53 patients, no relevant head movement was detected (43%), whereas subtle head movement was found in 16 patients (30%). Strong head movement was detected in 14 of 53 patients (26%).
After the introduction of artificial in-plane movement, we found that also in real CT perfusion, infarct core estimates and penumbral volumes depend on the occurrence of motion. Mean difference between Tmax-lesion size was +4.3 (−12 to 21) mL for subtle and +6.6 (−32 to 45) for strong movement. While no difference between mean infarct core size was found after introduction of subtle in-plane motion (−0.3 (−4.1 to 3.5) mL), a tendency towards smaller infarct core estimates is noted after the introduction of strong movement (−3.2 (−11 to 4.4) mL). Consequently, we found that the introduction of head movement led to an overestimation of mismatch ratios (mean difference +0.9 (−2.7 to 4.6) for subtle head movement and +1.6 (−2.0 to 5.2) for strong head movement (patients with no detected infarct cores (n=11 for subtle head movement and n=14 for strong head movement) were not included due to undefined mismatch ratios)). Figure 4 displays an exemplary patient where the introduction of artificial in-plane head movement led to an overestimation of mismatch ratio. Nevertheless, in this study population, no potentially treatment relevant misclassification was observed.
Automated CT perfusion mismatch assessment is an established treatment decision tool and has proven to identify patients potentially benefitting from endovascular stroke therapy and intravenous thrombolysis in extended time windows.2 3 15 While perfusion-based treatment decisions are hence valid at a group level, there may still be misclassifications at the individual patient level.
Already it has been noted that automated mismatch evaluation depends on the software16 and a general tendency to underestimate infarct cores has been noted before.17 So, Hoving et al compared acute CT perfusion infarct core volumes with follow-up diffusion weighted imaging (DWI) infarct volumes after MT and found that infarct core volumes tended to be underestimated by 4.4 mL in median.18 This finding agrees perfectly with our results on simulated data. Still, one potential drawback of the above-named studies is that the infarct core, as predicted by CT perfusion analysis, is compared with the final infarct core detected later on follow-up imaging. While positron emission tomography-CT (PET-CT) makes it possible to approximate physiological brain perfusion,19 this technique is not feasible in the acute setting due to the considerable time delay and logistic effort.
Hence, in the clinical setting, the ground truth of brain perfusion on an individual basis is unknown. In this context, a major strength of our study is that the ground truth is used as simulation input. We point out that such simulations are an established approach to validate perfusion analysis software.9 10 Manniesing et al used a similar digital brain perfusion phantom and found a dependency of contrast-to-noise ratio from the lesion size of penumbral tissue and the affected tissue class.20 Kudo et al found in simulated data that estimated CBV and CBF values differ from simulated values and depend on the analysis software.9
To our knowledge, our study is the first to investigate the influence of noise and motion on the clinically relevant mismatch ratio. While, overall, differences between simulation input and evaluated volumes were small, we still found that already minor head movement influences mismatch evaluation. This finding is of special importance since head motion frequently occurs in patients with acute ischemic stroke and possible agitation.21 We observed such head movement in 57% of CT perfusions. Overall, a tendency to underestimate infarct cores and to overestimate penumbral volumes in dependency of head movement was noted. Most importantly, misclassifications according to the DEFUSE-3 target mismatch profile were observed always in favor of MT. Hence, patients are unlikely to be withheld from MT due to the occurrence of head movement. We point out that in clinical decision-making, such a selection bias towards MT is preferable compared with false refusal from MT. For instance, Sarraj et al found that MT is associated with higher rates of functional independence and lower mortality, even in patients with large core sizes across various thresholds and infarct core definitions.22 Our findings even suggest that patients who do not exhibit any head movement during CT perfusion imaging and miss the mismatch criteria (which were established on stroke cohorts where motion occurs) might have been deemed eligible to benefit from MT, if they would have been examined with head motions.
Our study has several limitations. First, we simulated the tissue attenuation curves using a convolution model and (after the introduction of noise and motion) analyzed these data using a deconvolution model. Potentially, this leads to trivial inversions and the results can be unrealistically optimistic.20 In our analysis, this problem may, however, be of minor importance, since we did not compare absolute CBF/CBV values but we compared lesion sizes. Moreover, our results on simulated data are corroborated by studying the effect also in real CT perfusion. We point out that the simulated motion is of an artificial nature. No streak artifacts, image blurring or additional partial volume effects are introduced. Moreover, beam hardening from the skull as well as beam hardening from the introduced iodine contrast agent may influence CT perfusion evaluation.10 While these artifacts are hard to simulate, they may lead to further reductions in image quality. As such, beam hardening and the occurrence of streak artifacts can either lead to complete exclusion of voxels from mismatch analysis or to reduced reliability of perfusion values and may even mimic ischemic lesions.23 The influence of contrast agent bolus delay and bolus shape on perfusion analysis was minimized by using an arrival-time-insensitive perfusion algorithm. Thereby, Tmax should represent a bolus-shape-independent parameter of perfusion delay.24 Finally, the quality of motion correction will depend on image quality (ie, contrast and signal to noise ratio), resolution (with better results for isotropic, high resolution data)25 and the motion correction algorithm itself. Hence, technical advances in all these areas could help to reduce the impact of movement on CT perfusion mismatch analysis. While the major program for mismatch analysis in the DEFUSE-3 study was RAPID,2 we point out that our results may not directly apply to other software. For instance, Koopman et al found that ischemic core volumes and the frequency of infarct overestimation depend on the utilized software packages and the applied smoothing filters.26 Lastly, we point out that our study focused on the technical limitations of automated CT perfusion mismatch evaluation. Other effects, such as the impact of time from symptom onset to imaging or the extent of collaterals on optimum perfusion thresholds, cannot be addressed in our simulation study.
CT perfusion mismatch analysis has been shown to be influenced by subtle to strong patient head movements. Thereby, a tendency to overestimate mismatch ratios in dependence of head movement was noted. Potentially treatment relevant changes in mismatch classifications were observed. Noticeably, these occurred in favor of MT, removing safety concerns. Patients were unlikely to be withheld from MT due to the occurrence of head movement during CT perfusion acquisition.
Data availability statement
The datasets generated and/or analyzed during this research are available from the corresponding author upon reasonable request.
Patient consent for publication
Contributors AP, MB, SH and MMö initiated the project. AP led the research, conducted the simulations, data acquisition, statistical analysis and wrote the manuscript. MB, SH, MMö were involved in the study design and concept. FS, CW, MMu, CH assisted data acquisition. All authors discussed the results, commented on the paper, and approved the final version of the manuscript.
Funding The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-for-profit sectors.
Competing interests None declared.
Provenance and peer review Not commissioned; internally peer reviewed.
Supplemental material This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.