A Comparative Study of CT Perfusion Postprocessing Tools in Medium/Distal Vessel Occlusion Stroke

Patient Cohort

The local PACS was searched retrospectively for patients with acute disabling ischemic stroke by using the following inclusion criteria: patient age ≥18 years, acute stroke with symptom onset within the last 8 hours, MDVO with an isolated vessel occlusion, availability of NCCT, CTA, and CTP. Inclusion period was January 1, 2018 to December 31, 2023. Patients were not included if artifacts were found on CTP, such as patient movement indicated by automatic patient movement detection, or if CTP quality was reduced, eg, due to truncation, indicated by an incomplete venous outflow curve. Patients with a high-grade (>70%) stenosis of a cervical vessel according to NASCET criteria and patients with known intracranial arteriosclerotic disease with an intracranial stenosis >50% according to Comparison of Warfarin and Aspirin for Symptomatic Intracranial Arterial Stenosis criteria were not included to avoid confounding of perfusion results due to stenotic lesions proximal to the location of vessel occlusion.

Disabling stroke was defined to encompass NIHSS at admission ≥4 or clearly disabling symptoms such as aphasia or hemianopia. Inclusion was limited to patients with disabling stroke to avoid a cohort with heterogeneous treatment indication status and unclear clinical relevance of perfusion results. This study follows the anatomic definitions used in the DISTAL trial¹⁵: MDVO was defined to encompass occlusions of nondominant or codominant M2 segments that represent the first vertical segment of the MCA, or M3–M5 segments of the middle cerebral artery, A1–A5 segments of the anterior cerebral artery, P1–P4 segments of the posterior cerebral artery, or occlusions of the anterior or posterior communicating arteries. Figure 1 shows a patient flow chart to illustrate how patients were selected.

• Download figure
• Open in new tab
• Download powerpoint

FIG 1.

Patient flow chart. *The full-text search was conducted in the local PACS and included the descriptors “acute stroke,” “cerebrovascular insult,” and “National Institutes of Health Stroke Scale” as well as corresponding abbreviations.

Ethics committee approval and patient consent were not required according to current local legislation because all data were anonymized before analysis and the project involved assessing safety and quality of routine acute ischemic stroke management in the participating institutions. This manuscript follows the methodology proposed by the Strengthening the Reporting of Observational Studies in Epidemiology consortium¹⁶ and the corresponding checklist can be found in the Supplemental Data.

Image Data Acquisition and Postprocessing

Images were acquired by using conventional multislice CT scanners (Somatom Force, Siemens Healthineers). The CTP imaging adhered to a standard protocol as recommended by the manufacturer. Using a dual-head power injector, 60 mL of the contrast agent (Iopamiro 300, Bracco) was delivered at a rate of 5 mL/s through an 18-gauge venous line, followed by a 30 mL saline flush. The acquisition parameters for most images included: 180 mAs, 70 kV, 570 ms, a matrix size of 512 × 512, a field of view of 20 cm, a spiral pitch factor of 0.5, a single collimation width of 1.2 mm, and H36f kernel. A total of 28 contrast phases (20 scans with cycle time of 1.5 seconds and 8 cycles of 3 seconds) were captured in 53.03 seconds.

The following commercially available postprocessing tools were used: RapidAI CTP automated postprocessing (Version 5.0.4, iSchema View) and Syngo.via CT Neuro Perfusion semiautomated postprocessing (Siemens Healthineers).

Perfusion Lesion Segmentation

Semiautomated postprocessing results were manually segmented by using the open-source software package 3D Slicer (https://github.com/Slicer/Slicer)¹⁷ by a radiologist (N.N.) and critically reviewed by a senior diagnostic and interventional neuroradiologist with >10 years of experience (M.-N.P.). Regions with elongated time-to-maximum (Tmax) were segmented to identify tissue at risk, while regions with elongated time-to-drain, but normal Tmax were not included. Regions with reduced CBV were segmented to identify infarct core. For both Tmax and CBV, each single lesion was segmented on all affected slices, encompassing the complete volume of the perceived perfusion abnormality. Areas of visually perceived perfusion abnormality not corresponding to acute ischemia such as chronic postischemic defects were not included in lesion segmentation.

Segmentation of automated postprocessing results was performed automatically by using red, green, blue (RGB) channels with the following thresholds as defined by the vendor: Tmax > 4 seconds and Tmax > 6 seconds to identify tissue at risk, CBF <30% and Tmax > 10 seconds to identify infarct core. Tmax > 10 seconds has previously been used as a surrogate of infarct core and was shown to predict final infarct volume in patients with MDVO with unsuccessful reperfusion¹⁴ and the ratio of Tmax > 10 seconds to Tmax > 6 seconds, also known as hypoperfusion intensity ratio, was shown to correlate to patient outcomes.^18,19

The RGB images were reformatted as binary images. Both the images and the labels from automated postprocessing were resampled to match the number of slices in the maps derived from semiautomated postprocessing. The binary label maps were resampled in the signed distance field domain. Coregistration was performed by using the Advanced Normalization Tools symmetric normalization registration method (affine and deformable transformations),²⁰ with mutual information as the criterion (antspyx 0.4.2). The fixed images were the images from semiautomatic and the moving images were the images from automated postprocessing. The transformations were then applied to the labels of the moving images.

Definition of Comparative Measures

For comparison and high-quality registration, the pixels were resampled to have the same pixel size in both images. Pixels were not made isotropic to avoid excessive resampling. The volume of each perfusion lesion was measured as the sum of the pixels of a label. To obtain an estimate of the actual volume (estimated volume), we multiplied the measured volume by the size of the pixel in each direction (0.4 × 0.4 × 5 mm).

The relative volume difference (RVD) was defined as with U being the volume of the first label and V the volume of the second label. It represents quantitative size difference in relation to the absolute lesion volume.

The Dice score (Sørensen Dice coefficient [DS]) and Hausdorff distance (HD)²¹ were used as quantitative measures of distribution. DS is defined as where X is the first label and Y the second label. It represents how well the 2 labels overlap, with the value 1 representing perfect overlap and the value 0 representing no overlap. HD was defined to be , where D is the distance, A is the ensemble of pixels belonging to the first label and B is the set of pixels belonging to the second label. HD represents the largest distance between the 2 labels and is complementary to the Dice score, with the value 0 representing identical label margins. Figure 2 illustrates the workflow of the study including illustrations of DS and HD.

• Download figure
• Open in new tab
• Download powerpoint

FIG 2.

Study workflow.

Perfusion Lesion Comparison and Statistics

Regarding surrogates of tissue at risk and infarct core, we compared the Tmax lesion segmented from semiautomated postprocessing maps with lesions from 2 different thresholds of Tmax maps derived from automated postprocessing (thresholds: 4 seconds and 6 seconds). We compared the CBV lesion segmented from semiautomated postprocessing with core lesions derived from automated postprocessing (thresholds: < 30% CBF and Tmax > 10 seconds).

For each patient, the result of each comparison was classified to represent agreement or one of several types of disagreement according to the following criteria: First, the RVD was classified as volume agreement if below 0.5 (<50% volume difference) and volume disagreement if above 0.5. Patients in the volume agreement category were then classified in distribution agreement if their DS was above or equal to 0.5 (>50% lesion overlap) and in distribution disagreement if their DS was lower than 0.5. The patients from the category distribution agreement were classified to represent agreement if their HD was below or equal to 50 mm or to represent agreement with outliers if their HD was above 50 mm. This indicates that the distribution of the major component of the lesion was similar in both maps, but a minor fraction of the lesion was unequally distributed. The choice of the above-mentioned thresholds is arbitrary and represents the authors’ consensus of the lowest acceptable agreement between 2 tools if the tools are to be used in clinical decision-making. Furthermore, this consensus represents a reasonable compromise between technically feasible and clinically meaningful options. Figure 3 illustrates the comparison decision tree.

• Download figure
• Open in new tab
• Download powerpoint

FIG 3.

Classification scheme.

Patients in the category volume disagreement were further subclassified into 3 predefined groups. The definition of these groups is derived from observation of typical and relevant patterns of disagreement during routine clinical practice. The 3 groups were: 1) no lesion in automated postprocessing Tmax > 6 seconds map, 2) a substantially larger Tmax > 4 seconds lesion as compared with semiautomated postprocessing results of Tmax (defined as DS <0.5, HD >50, and volume ratio >0.5), and 3) patients with no CBV or CBF lesion in any postprocessing. Figure 4 shows examples of these patterns.

• Download figure
• Open in new tab
• Download powerpoint

FIG 4.

Examples of different categories of agreement or disagreement. Figure shows Tmax maps for semiautomated (upper row) and automated postprocessing. A, Agreement. B, Agreement with outliers. Please note contralateral areas of Tmax > 4 seconds indicated in blue in the RapidAI map. C, Tmax > 4 seconds overestimation. Please note how the Tmax >4 seconds lesion of RapidAI extends beyond the lesion of perfusion delay in the Syngo.via Tmax map. D, No Tmax > 6 seconds lesion. E, Volume mismatch. F, Color reference for perfusion maps derived from Syngo.via. G, Color reference for perfusion maps derived from RapidAI.

The median values for the DS, HD, and RVD were computed. If a category included at least 20% of patients, it was described as a pattern. Patterns were defined to be infrequent, if occurring in 20%–40% of patients, frequent (41%–60%), or dominant (>60%).

Calculation and Comparison of Mismatch Volumes

Regarding the penumbra, ie, potentially salvageable tissue, that is a mismatch between the area of perfusion delay and the infarct core: For semiautomated postprocessing, mismatch volume was defined as Tmax lesion volume – CBV lesion volume. For automated postprocessing, 2 mismatch volumes were calculated, defined as Tmax > 4 seconds volume – CBF <30% volume and Tmax > 6 seconds volume – CBF <30% volume. Mismatch volumes were compared by using t test for paired samples including Cohen d to calculate effect size and visualized by using Bland-Altman plots. SPSS Version 28.0.1.0 (IBM) was used for plotting and statistical testing. We deemed P values <.05 to be significant.