Validation of a Novel Multiphase CTA Perfusion Tool Compared to CTP in Patients With Suspected Acute Ischemic Stroke

Abstract

Background

We recently developed a novel machine learning‐based algorithm using multiphase computed tomography angiography (mCTA) to generate perfusion maps of the brain, similar to computed tomography perfusion (CTP) (ie, multiphase CTA perfusion [mCTAp]). Here, we aim to validate the clinical utility of mCTAp in detection of brain ischemia and its side, extent, and location.

Methods

In this prospective multi‐reader‐multi‐case analysis, we included baseline images: mCTAp (StrokeSENS‐algorithm) and CTP (4D; GE Healthcare) from 121 randomly selected patients whose scans were not part of algorithm‐development. After excluding 2/121 scans because of poor image‐quality, 3 experienced radiologists read time to maximum, and relative cerebral blood flow‐maps generated by the test (mCTAp) and reference (CTP) modality. The 2 reading sessions were separated by 5 days although the reading order was randomized. Core laboratory imaging assessments – that used non contrast computed tomography, mCTA, and CTP – were considered as ground‐truth. A mixed‐effects statistical model with “reader” as random effects variable was used to calculate the area under the curve (with 95% CI), sensitivity, and specificity for both modalities (mCTAp/CTP) for ischemia detection, affected side, and occlusion location. The time required for interpretation and inter‐rater variability in assessments were compared across the 2 modalities.

Results

Area under the curves (95% CI) for detecting ischemia using mCTAp and CTP were 0.85 (95% CI, 0.8–0.9) and 0.84 (0.8–0.9) respectively (P=0.43). Area under the curves for the affected side were 0.94 (0.92–0.97) and 0.96 (0.94–0.98) (P=0.69), respectively; for detecting large vessel occlusion were 0.84 (0.8–0.9) and 0.86 (0.8–0.9), (P=0.31), respectively; M2‐or‐distal occlusion were 0.79 (0.73–0.84) and 0.88 (0.83–0.92) (P=0.22), respectively, for anterior cerebral artery‐occlusion 0.82 (0.66–0.98) and 0.93 (0.82–1.00) (P=0.15), respectively, and for posterior cerebral artery‐occlusions 0.9 (0.8–1) and 0.99 (0.98–0.99) (P=0.01), respectively. The median (interquartile range [IQR]) time for image interpretation was 62 seconds (IQR, 46–78) and 59 seconds (IQR, 42–69) for mCTAp and CTP, respectively, (P=0.15). Fleiss` Kappa‐values for inter‐rater reliability in detecting ischemia were 0.5 and 0.8 for mCTAp and CTP, respectively.

Conclusion

mCTAp shows similar performance and interpretation times compared to CTP in assisting readers to detect brain ischemia, affected side, and occlusion location, but mainly as it relates to proximal vessel occlusions. The proposed tool still needs further refinement for distal vessel occlusions. Nonetheless, mCTAp is a promising tool as it allows for acquisition of brain perfusion maps with lower radiation exposure, acquisition time, and contrast dose compared with additional CTP.

Nonstandard Abbreviations and Acronyms

CTP

computed tomography perfusion

mCTA

multiphase computed tomography angiography

NCCT

non contrast computed tomography

Tmax

time to maximum

Clinical Perspective

• What Is New?

  A perfusion tool that provides brain perfusion maps by using baseline non contrast computed tomography and multi‐phase computed tomography aniography (ie, multiphase computed tomography angiography perfusion) as input, shows similar performances as computed tomography perfusion in aiding the reader to find the occlusion and its anatomic location. This tool mainly relates to large vessel occlusions and needs further refinement for distal vessel occlusions.

• What Are the Clinical Implications?

  By leaving out an additional scan at baseline (computed tomography perfusion), we achieve similar performances, hence reducing radiation exposure, iodine contrast usage, and acquisition time.

After the publications of the DEFUSE 3 ¹ (Endovascular Therapy Following Imaging Evaluation for Ischemic Stroke) and DAWN ² (DWI or CTP Assessment With Clinical Mismatch in the Triage of Wake‐Up and Late Presenting Strokes Undergoing Neurointervention With Trevo) trials, many centers worldwide have implemented computed tomography perfusion (CTP) volumetrics (penumbra, core) in treatment decision protocols for late window acute ischemic stroke (AIS) patients. However, such advanced imaging has several limitations (eg, limited brain coverage, complex post‐processing software). Recent literature, including the recent MR CLEAN‐LATE (Multicenter Randomized Clinical Trial of Endovascular Treatment for Acute Ischemic Stroke in the Netherlands) randomized controlled trial, shows that CTP use may not have additional value when selecting patients for randomized controlled trial (EVT), when compared with the use of non contrast CT (NCCT), magnetic resonance imaging or CTA based imaging paradigm. ³ , ⁴ , ⁵ Nonetheless, CTP provides an easily interpretable assessment of brain perfusion, especially when readers lack the expertise to interpret NCCT or CTA based imaging in a timely manner.

Simultaneously, machine learning‐based tools have been used to enhance the ability of simpler imaging modalities such as NCCT and CTA in predicting the fate of ischemic brain more accurately. Multi‐phase CTA is a time‐resolved imaging modality with faster imaging acquisition, lower radiation exposure, less need for intravenous contrast, and higher spatial resolution than CTP. Interpreting collaterals on all 3 phases of a mCTA does however require some expertise. We recently used ML‐based techniques on multiphase computed tomography angiography (mCTA) to develop perfusion maps that resemble similar maps available from CTP. The current study is designed to compare the ability of radiologists to assess perfusion maps generated from mCTAp with maps provided from CTP in patients with suspected AIS. The aim of mCTAp is not to outperform CTP, but rather to investigate the similarity of the produced perfusion maps with CTP.

Methods

The data that support the findings of this study are available from the corresponding author upon reasonable request. Multiphase CTA perfusion (mCTAp) is a novel technique that leverages the temporal ability of multiphase CTA to provide time to maximum (Tmax), cerebral blood flow (CBF), and relative cerebral blood volume‐perfusion maps. Technical aspects of mCTAp are provided further down in this paper.

In the current study, we compare mCTAp with CTP in its clinical applicability. For this study, we included a total of 121 out of 606 participants derived from the Prove‐IT (Precise and Rapid Assessment of Collaterals Using Multi‐Phase CTA in the Triage of Patients with Acute Ischemic Stroke for IA Therapy) study. Scans from these patients were not used for the development of the mCTAp tool. The Prove‐IT study was a prospective multi‐center, hospital‐based cohort study of 606 consecutive patients with AIS, presenting within 12 hours of stroke symptom onset with evidence of intracranial occlusion on routine CTA, over 3 years (ClinicalTrials.gov, Identifier: NCT02184936). Results of the Prove‐IT study has been published elsewhere, ⁶ where further details about enrollment period and inclusion/exclusion criteria can be found.

The data selection flowchart of the current study is shown in Figure 1. All included participants had CTP (CT Perfusion 4D; General Electric Healthcare, Waukesha, WI, provided by GE CTP 4D, version 11.3.81) and mCTA with mCTAp maps generated by the StrokeSENS algorithm, version 1.4. Both imaging studies were performed at the time of stroke imaging work‐up, before EVT treatment, and within 10 minutes of each other. All readers assessed all cases by using the index test (mCTAp, reading arm 1) and subsequently using the reference test (4D CTP, reading arm 2) with a 5‐day interval in between (Figure 2 shows the study design flow chart).

Figure 1.

Data selection flowchart (N=119).

Figure 2.

Study design flowchart (N=119).

Three experienced readers visually analyzed Tmax and relative cerebral blood flow perfusion maps of both mCTAp and CTP modalities subsequently, of all patients. Three experienced radiologists (F.Be. with 4 years of experience, F.Ba. with 12 years of experience, and I.A. with 6.5 years of experience in reading stroke imaging) were blinded to all clinical as well as any other imaging information, receiving only the mCTAp or CTP post‐processed images at the subject‐level in each reading session. Both reading sessions were set to have minimal differences in reading conditions. This meant that both sessions are done on the same desktop computer (using Osirix‐viewer MD, version 12.5) and under the same ambient lighting conditions that would normally be applied when reading acute stroke scans in the emergency setting. For each subject‐level read, the readers were asked to record (i) if there is any observable ischemia (ie, hypoperfusion deficit), (ii) the side of ischemia (if present), as well as (iii) the anatomic location of the occlusion (if present). Before the reading sessions started, the readers reached consensus on how to define the anatomic occlusion locations with patterns of hypoperfusion only. This was based on anatomic knowledge as well as on previous published literature. ⁷ A hemispheric deficit (larger than 50% of the hemisphere is affected) was considered as a large vessel occlusion (LVO; internal carotid artery [internal carotid artery (ICA), internal carotid artery ‐ T (thrombus extending in both A1 and M1); ICA‐T, or internal carotid artery ‐ L (thrombus extending in either A1 or M1) ICA‐L] and/or M1 segment of the medial cerebral artery [MCA]). A non‐hemispheric deficit located in the M1 region (eg, lentiform hypoperfusion region) was considered as MCA territory and hence included in the LVO‐pool, since strictly speaking this could be an LVO with good collaterals. A non‐hemispheric (<50% of the hemisphere is affected) hypoperfusion region, located anterolateral/frontobasal or more posteriorly (temporal/parietal) was considered as an anterior or posterior M2 occlusion, respectively, and a small peripheral wedge‐shaped deficit was considered as an M3 occlusion or even more distally. Posterior cerebral artery (PCA) and anterior cerebral artery (ACA) occlusions were considered as any occlusion in the PCA and ACA territory, respectively. In case of any indeterminate results, a fourth, independent, expert (B.M. with >15 years of experience in stroke imaging) was consulted for consensus understanding. The reading sessions were time recorded, where the reader set the timing of the start of the read (ie, opening the image) and of the end of the read (ie, having all questions answered). Automated calculations provided the total duration of the read and the duration per reader, for all readers.

To generate the reference standard for this study, the original core laboratory readings collected as part of the Prove‐IT‐study were considered. Ground truth readings were conducted by 2 experienced neuro‐radiologists, who read the images by consensus. ⁶ Unclear findings were further reviewed and interpreted by a senior neuroradiologist. ⁶ In all truthing scenarios, the truthers had access to baseline clinical information (clinically affected hemisphere, National Institutes of Health Stroke Scale [NIHSS], time since symptom onset) as well as baseline imaging, including NCCT, mCTA, and CTP. Although truthing the data, they were blinded to any clinical outcomes. See Figure 3, for example of a left distal M1 occlusion (A) and a left M3 occlusion (B) showing the mCTAp left – CTP middle – ground truth [mCTA] right.

Figure 3.

Tmax provided by mCTAp and by CTP.

Examples are provided of a left distal M1 occlusion (A) and a left M3 occlusion (B) showing the mCTAp left – CTP middle – ground truth [mCTA] right. CTP indicates computed tomography perfusion; mCTAp, multiphase CTA perfusion; and Tmax, time to maximum.

NCCT and mCTA

NCCT with 5 mm slice thickness was obtained, followed by mCTA. The first phase of mCTA was acquired from arch to vertex (CTA head and neck; conventional single phase CTA). This first phase was followed by skull base to vertex coverage for the second (peak venous) and third (late venous) phase of mCTA acquisition. Axial images with 1 mm overlap and multiplanar axial, coronal, and sagittal reconstructions with 3 mm thickness, 1 mm intervals, and 1 mm overlap for the first phase were obtained, along with axial minimum intensity projections for all 3 phases with 24 mm thickness and 4 mm intervals. The first phase is scanned during the peak arterial phase and is mostly acquired in less than 7 seconds. The second phase was acquired after a delay of 4 seconds, allowing for table repositioning to the skull base, and the scanning duration for each additional phase was 3.4 seconds. Detailed mCTA acquisition parameters have been published previously. ⁸

Index Test: StrokeSENS mCTA Perfusion

The mCTAp algorithm used in this study (Figure 3A, B – left), namely StrokeSENS mCTA Perfusion (StrokeSENS mCTAp; Circle Cardiovascular Imaging, Calgary, Canada), utilizes both NCCT and mCTA volumes as input. First, preprocessing operations are performed to ensure all volumes were read properly and aligned with each other. These included gantry tilt correction, inter‐slice distance variation correction, co‐registration between the mCTA phases, and co‐registration between the mCTA and NCCT volumes. Second, StrokeSENS mCTA perfusion calculated the voxel‐wise contrast time curve for all tissue voxels and used them to measure the CBF, cerebral blood volume, and time to peak maps by using the direct method (ie, a non‐deconvolution‐based formulation of perfusion parameters), without the need for arterial input factor information. Finally, these maps were used as input to a machine learning‐approach (ie, a neural network) to generate the final Tmax and CBF perfusion maps. The CBF was then normalized to create the relative cerebral blood flow maps. The AI model adds on to the non‐deconvolved output maps (time to peak/cerebral blood volume /CBF) by filtering out noise and generating estimations of Tmax and relative cerebral blood flow that are accurate relative to what would be produced by CTP. More details about the neural network and the StrokeSENS mCTA Perfusion algorithm can be found in prior published literature. ⁹

Reference Test: 4D GE CTP

The CTP source data were centrally processed in a non‐acute setting for each study, using a delay‐insensitive deconvolution software (CTP 4D, GE Healthcare, Waukesha, WI. Version 11.3.81). The protocol consisted of 45 mL of iodinated contrast agent injection at a rate of 4.5 mL/s, followed by a 40 mL saline bolus injected at a rate of 6 mL/s. Image acquisition started 5 seconds after contrast injection and 24 passes over 66 seconds were performed with 5 mm section thickness and a cranio‐caudal coverage of 8 cm. For further details, please refer to previous published literature. ⁶

Statistical Analysis

Baseline characteristics of the included patients are provided in Table 1 with descriptive statistics, conducted according to the nature of the data. Reader‐averaged area under the curve (AUC), ¹⁰ sensitivities and specificities were provided by using mixed effect regression models for each variable (ie, ischemia detection, left side, LVO, M2 or distal, ACA, and PCA). The reader is included as a random‐effects variable in these analyses. Each variable of interest (ie, ischemia detection, side, and occlusion location) was analyzed as a separate model. Delong tests, after bootstrapping, were conducted for comparisons of the reader averaged AUCs for both modalities, per model. Agreements between the readers among the subjects in both arms were analyzed using Fleiss` Kappa statistics. The time of image interpretations for each reader are presented as medians (interquartile range [IQR]) and were compared between the 2 modalities using Mann–Whitney U test. Median reading (IQR) times of all readers are compared between the 2 imaging arms. A P value of <0.05 was considered statistically significant. All analyses were performed using Stata/MP 16.1 (StataCorp, College Station, TX).

Table 1.

Baseline Characteristics of Total Cohort of Included Patients, N=119

Characteristics	n (%)
Female, n (%)	62 (53)
Age, median (IQR)	71 (61–79)
Baseline NIHSS, median (IQR)	13 (8–23)
Time from onset to CT, median (IQR)	134 (81–266)
Ischemia present, n (%)	96 (81)
Side, n (%)
• Left	53 (45)
• Right	43 (36)
Anatomic location of the occlusion, n (%)
• LVO (ICA+MCA [M1‐segment])	59 (50)
• MCA (M2‐segment or distal)	29 (24)
• ACA	4 (3)
• PCA or vertebrobasilar	4 (3)
• No occlusion	23 (19)
Isolated cervical ICA occlusion, n (%)	1 (1)
Tandem occlusion, n (%)	2 (2)
Collateral grading, n (%)
• Good	72 (61)
• Intermediate	29 (4)
• Poor	6 (5)
ASPECTS, median (IQR)	7 (6–8)

ACA indicates anterior cerebral artery; ASPECTS, Alberta Stroke Program Early CT Score; ICA, internal carotid artery; IQR, interquartile range; MCA, medial cerebral artery; NIHSS, National Institutes of Health Stroke Scale; and PCA, posterior cerebral artery.

Results

Of 121 scans, 2 were not interpretable (Figure 1: patient inclusion flow chart) because of technical issues with the files (n=1 for mCTAp reading arm and n=1 for CTP reading arm). Of the included 119 patients, the median age was 71 (IQR, 61–79) years; 62% was female and the median baseline NIHSS score was 13 (IQR, 8–23). The median onset to initial (CT)‐imaging time (minutes) was 134 (IQR 81–266) minutes. Ninety‐six out of 119 patients (81%) had an occlusion, 53 (45%) was in the left hemisphere. Most of the occlusions were LVOs (ICA or proximal M1 segment). One case (1%) had an isolated cervical ICA occlusion, and 2 cases (2%) showed a tandem occlusion. The median Alberta Stroke Program Early CT Score at baseline was 7 (IQR, 6–8) and most of the cases (61%) had good collaterals, Table 1. Collateral gradings were conducted, according to previously published literature. ⁸

Performances of mCTAp Compared to CTP

The AUC for detecting ischemia and selecting the correct side was 0.85 (95% CI, 0.80–0.91) and 0.94 (95% CI, 0.92–0.97), respectively, for the mCTAp arm and was 0.84 (95% CI, 0.78–0.89) and 0.96 (95% CI, 0.94–0.98), respectively, for the CTP arm (P=0.43 and 0.69, respectively). The AUC for the presence or absence of an LVO and M2/distal MCA occlusion was 0.83 (0.79–0.87) and 0.79 (0.73–0.84), respectively, for the mCTAp arm and was 0.87 (0.83–0.91) and 0.88 (0.83–0.92), respectively, for the CTP arm (P=0.2 and 0.26, respectively). The AUC for the presence or absence of posterior occlusions (PCA and/or vertebrobasilar) and anterior occlusions (ACA) were 0.99 (0.98–1) and 0.82 (0.66–0.98), respectively, for the mCTAp arm and 0.99 (0.98–1) and 0.93 (0.82–1), respectively, for the CTP arm (P=0.52 and 0.54, respectively), Table 2. The sensitivity and specificity for detecting ischemia was 96.5% and 62.3%, respectively, for the mCTAp arm and 96.2% and 71%, respectively, for the CTP arm, Table 3. Inter rater agreements, assessed using Fleiss` Kappa values, for ischemia detection were 0.54 (95% CI, 0.38–0.70) and 0.75 (95% CI, 0.62–0.88) for mCTAp and for CTP, respectively, Table 4. The median (IQR) reader averaged time for image interpretation was 62 seconds (IQR, 46–78 seconds) and 59 seconds (IQR, 42–69 seconds) for mCTAp and for CTP, respectively, (P=0.14), Table 5.

Table 2.

Reader Performances: Reader Averaged AUC

AUC (95% CI)	mCTAp	SE	CTP	SE	P value for difference
Ischemia detection (n=96)	0.85 (0.80–91)	0.03	0.84 (0.78–0.89)	0.03	0.43
Left side (n=53)	0.94 (0.92–0.97)	0.01	0.96 (0.94–0.98)	0.01	0.69
LVO detection (n=59)	0.83 (0.79–0.87)	0.02	0.87 (0.83–0.91)	0.02	0.20
M2 or distal MCA occlusion (n=29)	0.79 (0.73–0.84)	0.03	0.88 (0.83–0.92)	0.02	0.26
PCA or vertebrobasilar occlusion (n=4)	0.99 (0.98–1.0)	0.005	0.99 (0.98–1.0)	0.004	0.52
ACA occlusion (n=4)	0.82 (0.66–0.98)	0.08	0.93 (0.82–1.00)	0.06	0.54

Reader averaged AUCs are based on mixed effect regression models, where the reader is considered as the random effect variable. P<0.05 is considered statistically significant.

ACA indicates anterior cerebral artery; AUC, area under the curve; CTP, computed tomography perfusion; LVO, large vessel occlusion; MCA, medial cerebral artery; mCTAp, multiphase CTA perfusion; and PCA, posterior cerebral artery.

Table 3.

Reader Performances: Reader Averaged Sensitivity and Specificity

	mCTAp		CTP
	Sensitivity	Specificity	Sensitivity	Specificity
Ischemia detection (n=96)	96.5%	62.3%	96.2%	71.0%
Left side (n=53)	98.1%	90.6%	98.7%	93.8%
LVO detection (n=59)	78.4%	87.2%	90.1%	84.0%
M2 or distal occlusion (n=29)	81.2%	75.8%	81.0%	90.0%
PCA or vertebrobasilar occlusion (n=4)	100%	97.3%	100%	97.9%
ACA occlusion (n=4)	66.7%	97.3%	88.9%	97.2%

Reader averaged sensitivity and specificity, based on mixed effect regression models. The reader is considered as the random effect variable.

ACA indicates anterior cerebral artery; CTP, computed tomography perfusion; LVO, large vessel occlusion; mCTAp, multiphase CTA perfusion; and PCA, posterior cerebral artery.

Table 4.

Inter‐Rater Agreements

KAPPA* (95% CI)	mCTAp	CTP
Ischemia detected	0.54 (0.38–0.70)	0.75 (0.62–0.88)
Left*	0.97 (0.89–1.00)	0.97 (0.91–1.00)
LVO (ICA + M1)	0.71 (0.59–0.83)	0.75 (0.64–0.86)
M2 or distal	0.71 (0.59–0.84)	0.73 (0.61–0.85)
PCA + vertebrobasilar	0.76 (0.54–0.97)	0.88 (0.70–1.00)
ACA	0.48 (0.19–0.78)	0.84 (0.61–1.00)

Fleiss’ kappa values (95% CI) for both modalities.

ACA indicates anterior cerebral artery; CTP, computed tomography perfusion; IQR, interquartile range; LVO, large vessel occlusion; mCTAp, multiphase CTA perfusion; and PCA, posterior cerebral artery.

^* Patients with no ischemia were excluded.

Table 5.

Time for Image Interpretation

	mCTAp	CTP	P value
Reader 1, median (IQR)	49 (38–58)	49 (37–56)	0.75
Reader 2, median (IQR)	89 (67–116)	81 (67–106)	0.41
Reader 3, median (IQR)	56 (48–68)	48.5 (39–60)	<0.01
Average of readers*, median (IQR)	62 (46–78)	59 (42–69)	0.15

CTP indicates computed tomography perfusion; IQR, interquartile range; and mCTAp, multiphase CTA perfusion.

^* The average time of all readers, presented for the 2 modalities as medians (IQR) for each modality. Times are displayed in seconds.

Discussion

We compared a novel mCTA based perfusion map (mCTAp) with 4D GE CTP in patients with suspected acute stroke. We show that radiologists are able to assess the presence or absence of brain ischemia, its side, and its location equally well using mCTAp as with 4D GE CTP. We also show that the time taken to assess these images for ischemia using mCTAp is similar to that taken using CTP.

Reader averaged AUCs for detecting the presence of any ischemic deficit (ie, the ability of a perfusion map to distinguish a true ischemia from noise) as well as detecting an LVO occlusion were equally high and comparable between both mCTAp and CTP. The ability of readers to use mCTAp to detect more distal occlusions like MCA, PCA, and ACA occlusions however were 0.79; 0.99 and 0.82 when compared with an AUC of 0.88; 0.99 and 0.93 for CTP, respectively. In addition, we show slightly lower but overall similar sensitivities and specificities when assessing mCTAp compared with CTP, except for sensitivity of ACA occlusions (>10% difference). However, since ACA and PCA occlusions contributed <5% to the total sample, this study is likely underpowered to show any performance deterioration with mCTAp versus CTP for these distal occlusion sites.

Overall inter‐rater variability when using mCTAp was higher compared with when assessing CTP images. Of note however, the average time for interpretation for mCTAp was comparable to 4D CTP (only 3 seconds difference). In our opinion, with more experience using and interpreting mCTAp, the speed and reliability of these reads when assessing brain perfusion in patients with suspected AIS is likely to increase.

The visual assessment of post‐processed perfusion maps is useful for radiologists in detecting ischemia presence, its side, and its location of an arterial occlusion. Physicians use such information to then make treatment decisions. Several articles have investigated the role of the hypoperfusion intensity ratio (HIR), calculated on thresholded Tmax maps, and its association with (i) collateral grading, (ii) infarction growth, and (iii) outcome. ¹¹ , ¹² Others have investigated the role of black holes (areas with severely disturbed perfusion, where no values have been calculated) on time to peak maps and its correlation with functional outcome. ¹³ CTP, however, involves additional contrast and radiation beyond what is used for a conventional CT angiography. Many small hospitals do not have the technicians with expertise or the capacity to acquire CTP 24/7. mCTAp on the other hand is derived from simple CT angiography, which is now done in most stroke‐ready hospitals. It only requires the addition of 2 low‐radiation CTA head acquisitions, timed to be immediately after the first CTA head and neck acquisition. ¹⁴ It will therefore reduce radiation exposure as well as the use of iodine contrast use without impeding the readers ability to visually assess brain perfusion images just like CTP. Other advantages of mCTA derived perfusion maps over CTP are full brain coverage, less image acquisition time, and therefore less susceptibility to motion artefacts, and potentially lower costs. Although the limited brain coverage can be resolved with so‐called 2‐slab scanning, this remains an issue with CTP, especially when a majority of scanners worldwide continue to be 64‐slice or lower. In the supplemental materials (Table S1), we show a summary of what each of CTP and the additional 2 phases of mCTA adds on to standard NCCT and single‐phase CTA for stroke work‐up, regarding abovementioned important considerations such as radiation exposure, acquisition time, contrast dose, etc.

Limitations

This study has some limitations. The number of readers (n=3) and between reader variance therefore is a potential limitation. We used a mixed effects hierarchical model, including the reader as a random effects variable, to account for this. Secondly, the prevalence of LVOs in this dataset is higher than in an unselected cohort of patients with AIS. This means that the performance of mCTAp versus CTP in patients with more distal occlusions may have been underpowered. Thirdly, the de‐noising process of the AI model is not always perfect resulting in false translations into the produced output predicted TMAX/relative cerebral blood flow. For example, Figure 3B shows an artefact in the contralateral hemisphere. In this region, the input non‐deconvolved time to peak map had longer delays which were then falsely translated into an artefact in the output TMAX map. This could be one of the reasons for a lower specificity in detecting distal MCA occlusions and needs more focus in further improvements / developments of the StrokeSENS mCTAp tool. Lastly, unlike in the real world, this study was designed such that the readers did not have the ability to directly assess the anatomy of the vessels or to use clinical information when assessing images. Real world prospective studies that compare mCTAp to CTP may be needed to better understand the comparative utility of these tools. On that note, a valid experiment would be to also include readers from other backgrounds (eg, emergency doctors), in order to further extrapolate our findings to the clinical field.

Conclusion

mCTAp shows similar performance and interpretation times compared to CTP in assisting readers to detect brain ischemia, affected side, and occlusion location, but mainly as it relates to proximal vessel occlusions. The proposed tool still needs further refinement for distal vessel occlusions. Nonetheless, mCTAp is a promising tool as it allows for acquisition of brain perfusion maps with lower radiation exposure, acquisition time, and contrast dose compared with additional CTP.

Source of Funding

The Prove‐IT study is funded by grants from the Canadian Institutes of Health Research.

Conflicts of Interest

Dr Menon holds patents on systems of triage in acute stroke, for LVO detection and for mCTAp, and stock ownership in Circle Cardiovascular Inc. Dr Bala has nothing to declare. Dr Duszynski is an employee of, and holds stock options for Circle Cardiovascular Imaging Inc. Dr Nayem Pinky, Dr Golan, and Luis A. Souto Maior Neto are employees of Circle Cardiovascular Imaging Inc. All other co‐authors have nothing to disclose.

Supporting information

Supplementary Information

Supplementary Information.