Data were analyzed from consecutive patients diagnosed with acute posterior circulation infarction enrolled in the International Stroke Perfusion Imaging Registry (INSPIRE). Patients were recruited from nine hospitals across Australia, Canada and China between February 2007 and February 2020. Patients had multimodal CT (non-contrast CT, CT angiography and CT perfusion) at acute presentation with follow up imaging (CT or MRI) performed at 24 to 48 h post stroke. Patients were treated with acute reperfusion therapies (including thrombolysis and mechanical thrombectomy) according to local eligibility protocols. Modified Rankin Scale (mRS) was documented at 90 days following stroke by an accredited assessor. Written informed consent was obtained from all participants enrolled in the registry. The INSPIRE study was approved by the local hospital ethics committees in accordance with Australian National Health and Medical Research Council guidelines.

Baseline and follow up imaging was assessed by a central imaging panel with each case being analyzed by at least two experienced readers. Cases with disagreement were resolved via consensus review with a third reader. Occlusion severity on baseline CT angiography (CTA) was graded using the thrombolysis in myocardial infarction flow grade score (TIMI) in keeping with previously published methods (19). Patients were included if they had evidence of complete occlusion on baseline CTA represented by a TIMI of 0 in a branch of the verterobasilar circulation. Patients were excluded due to the following reasons including; no follow up 24–48 h diffusion weighted MRI, missing CTP data, inadequate CTP scan coverage of the posterior circulation, excessive movement during CTP, absence of complete occlusion (TIMI 0) of a posterior circulation artery (defined as vertebral, basilar or posterior cerebral arteries) on baseline CT angiography, or partial recanalization (TIMI 1–2) on follow up vessel imaging.

Patients were divided into two groups on the basis of occlusion severity on follow up CTA or magnetic resonance angiography (MRA) performed at 24 to 48 h post stroke as adjudicated by the central imaging panel. Group 1 demonstrated evidence of complete occlusion of a posterior circulation artery on follow up imaging, defined by a TIMI of 0. This group was used to represent the acute penumbral lesion. Group 2 demonstrated “major reperfusion” with complete revascularization of the posterior circulation, defined by a TIMI score of 3 on follow up imaging, and was used to represent the acute infarct core. The selection process is summarized in Supplementary Figure 1.

Baseline CT imaging was performed using 64-, 128-, 256-, or 320- detector scanners. Axial z-axis coverage ranged from 40 to 160 mm. Scan acquisition sequence was; a non-contrast helical CT from skull base to vertex followed by CTP and then CTA from aortic arch to vertex. Details of each scanner and CTP protocol are provided in Supplementary Table 1.

Follow up MRI was performed at 24 to 48 h post stroke irrespective of treatment using 1.5 to 3 tesla scanners. The MRI protocol included an axial gradient-echo T2^*-weighted series, isotropic diffusion-weighted imaging (DWI), MR time of flight angiography and fluid-attenuated inversion recovery imaging.

CTP maps were calculated using a commercial image processing package (AutoMiStar, Apollo medical imaging technologies, Melbourne, Australia). This software automatically performs motion correction. It then autonomously derives arterial input and venous output functions by selecting an unaffected major artery (commonly the anterior cerebral artery) and venous sinus (commonly the superior sagittal sinus). Selected inputs were confirmed by an expert analyst (LE) prior to image processing. Areas of gliosis, chronic infarction and cerebrospinal fluid were automatically masked from perfusion maps using a Hounsfield unit threshold. CTP source imaging was processed using delay and dispersion corrected singular value decomposition deconvolution (24). Processed maps included delay time to peak of the residue function (DT), mean transit time (MTT), cerebral blood flow (CBF) and cerebral blood volume (CBV). This method of deconvolution produces delay time maps rather than the standard time to peak of the residue function maps (Tmax).

Each CTP source imaging slab was coregistered a 3d rigid method to the corresponding 24 h DWI using the MiStar imaging fusion tool. Accurate co-registration was confirmed using interactive visual blending (L.E.). Regions of interest representing the posterior circulation vascular territory and specific vascular territories of the thalamoperforating/basilar tip, basilar perforating, calcarine, non-calcarine posterior cerebral and cerebellar arteries were determined using a template (22) (Supplementary Figure 1) which was manually adjusted to the CTP source imaging. The 24 h DWI lesion was semi-automatically delineated using a region of interest tool based on signal intensity thresholds. Regions of interest were transferred to the acute CTP maps. A range of relative CTP thresholds were incrementally investigated (Supplementary Table 2). Relative thresholds were defined respective to the mean perfusion values for tissue in the contralateral hemisphere.

Data processing, analyses and visualization were performed using Python (version 3.9) using the NumPy, SciPy, and Matplotlib packages. Receiver operating curve (ROC) analysis was used to investigate the optimal CTP parameter threshold relative to follow up DWI infarct volume. Optimality was defined as the threshold and parameter which maximized the area under the curve (AUC). The DWI lesion was considered the “true” lesion. Overlapping voxels between the DWI lesion and the CTP lesion were considered “true positive.” DWI voxels outside of the overlapping DWI and CTP lesions were considered “true negative.” Voxels within the CTP lesion but not in the DWI lesion were considered “false positive.” Voxels within the DWI lesion but not within the CTP lesion were labeled as “false negative.” Sensitivity [true positive/(true positive + false negative)] was plotted against 1 - specificity [true negative/(true negative + false positive)] to generate a ROC curve for each perfusion map. An area under the curve (AUC) was determined for each ROC curve. Sensitivity, specificity, positive predictive value and negative predictive value were calculated for each perfusion threshold increment. Area under the curve was calculated for each using the ROC. An AUC of < 0.7 was considered poor for this study.

Only voxels within the posterior circulation vascular territory template were included in the analyses. This prevents a large number of “true negative” voxels relative to “false negative” voxels from disproportionately influencing the calculation of specificity and AUC calculation. Volume analysis was performed by scrutinizing the threshold which produced the smallest mean volume difference between the acute perfusion lesion and follow up imaging. Linear regression was used to assess the relationship between the acute CTP lesion and follow up MRI volumes. Agreement with the following DWI lesion was assessed (Bland-Altman) at the perfusion threshold resulting in the smallest mean volume difference.

Individual patient data from INSPIRE is not publically available. Individual data can be shared with partners based upon individual transfer agreements upon request with the corresponding author.

There were 331 patients with POCI in the INSPIRE across the assessed period. Exclusions are detailed in the STAndards for the Reporting of Diagnostic accuracy studies (STARD) diagram (Supplementary Figure 1). Thirty nine patients were included for analysis in this study. Median age was 71 years (range 34–91 years) and 41% were female. Median National Institute of Health Stroke Scale (NIHSS) score was 11 (interquartile range [IQR]: 6–17).

Median last known well (LKW) time to CTP was 223 (IQR: 109–385) min. Median time to follow up imaging was 1.8 days (IQR: 1.1–3.7 days). On follow up imaging, 12 patients (30.8%) had no reperfusion (Group 1) and 27 patients (69.2%) had major reperfusion (Group 2). Thrombolysis was administered in 23 patients (59.0%). Thrombectomy was performed in 21 patients (53.8%). Eleven patients (28.2%) received both therapies. Compared to patients with no reperfusion (Group 1), those with major reperfusion (Group 2) had higher rates of acute treatment including thrombolysis (18/27 [66.7%] vs. 5/12 [41.7%]), thrombectomy (17/27 [63.0%] vs. 4/12 [33.3%]) or combined therapy (9/27 [33.3%] vs. 2/12 [16.7%]). This difference was not significant (all p > 0.05). Median LKW time to thrombolysis and thrombectomy was 189 (IQR: 137–295) and 415 (IQR: 249–547) min, respectively.

Of the 12 patients in this group; there were two vertebral, two basilar, and eight posterior cerebral artery occlusions. Voxel based ROC analysis of all patients in the group (Figure 1A) demonstrated the perfusion parameter with the maximal AUC was delay time (AUC 0.74). The optimal delay time threshold for predicting penumbra was a DT >1 s (sensitivity 0.60, specificity 0.83) (Table 1). The optimal thresholds for the other perfusion parameters were: MTT >145% (sensitivity 0.35, specificity 0.94), CBF < 75% (sensitivity 0.53, specificity 0.68) and CBV < 80% (sensitivity 0.37, specificity 0.67).

Volumetric analysis indicated MTT >160% was best correlated with 24 h infarct volume, with an R² of 0.71 (Table 2; Figures 2A, B). This led to a mean volumetric difference (MVD) between the acute perfusion and 24 h MRI lesion of 0.31 cm³ (95% CI −11.92 to 12.53). The Bland-Altman 95% limits of agreement were-37.75 and 36.04 cm³ (Figure 2C). This was followed by DT >2.5 s (R² = 0.41), CBV < 50% (R² = 0.17), and CBF < 40% (R² = 0.01).

Analysis of posterior circulation subregions demonstrated that CTP was most accurate in the calcarine territory (AUC = 0.77) (Table 3) with the smallest MVD (0.24 cm³, 95% CI −18.91 to 19.38) (Supplementary Table 3). CTP penumbral estimates were less accurate in the other POCI subregions including the cerebellar (AUC = 0.66, MVD 0.95 cm³), non-calcarine PCA (AUC = 0.61, MVD 1.08 cm³), thalamoperforator and basilar tip (AUC = 0.55, MVD −1.42 cm³) and basilar perforating artery (AUC = 0.53, MVD 1.48 cm³). CTP tended to overestimate the acute penumbral volume within regions other than the thalamoperforator and basilar tip vascular regions.

Of the 27 patients in this group; there were 4 vertebral, 17 basilar, and 6 posterior cerebral artery occlusions. On voxel based analysis of all patients in the group (Figure 1B), both DT and MTT demonstrated the highest AUC (0.73). The optimal threshold for these parameters were a MTT >145% (sensitivity 0.51, specificity 0.86) and DT >1.5 s (sensitivity 0.58, specificity 0.81). The optimal perfusion thresholds for the other parameters were: CBF < 45% (sensitivity 0.28, specificity 0.92) and CBV 70% (sensitivity 0.31, specificity 0.77).

A MTT of > 170% led to the smallest MVD with the follow up infarct volume (0.03 cm³, 95% CI −13.19 to 13.24) (Figures 3A, B) (Table 2). The Bland-Altman 95% limits of agreement were-64.24 and 64.30 cm³ (Figure 3C). All acute perfusion parameters demonstrated low correlation with follow up infarct volume (MTT >170% R² 0.11, CBF < 40% R² 0.13, DT >4.5 s R² 0.03, and CBV < 45% R² 0.01).

On subregion analysis, CTP was most accurate in detection of calcarine and cerebellar lesions with an AUC of 0.79 in both areas (Table 3). This was followed by basilar perforating (AUC 0.73), non-calcarine PCA (0.64) and thalamoperforators and basilar tip regions (0.50). DT and MTT were the optimal CTP parameters in these regions.