This retrospective multicenter study included patients with acute ischemic stroke who underwent PWI within 24 h of symptom onset at two tertiary hospitals in Korea. A total of 216 patients from Seoul National University Bundang Hospital who underwent both PWI and endovascular thrombectomy between January 2019 and April 2024, and 102 patients from Chonnam National University Hospital who underwent PWI within 24 h of symptom onset with or without endovascular thrombectomy (EVT) between January 2015 and December 2015, were initially screened. After pooling the datasets, 318 patients met the inclusion criteria. Of these, patients were excluded due to abnormal arterial input function (n = 6), severe motion artifacts (n = 2), or inadequate images (n = 11). Consequently, 299 patients were included in the final analysis. The study protocol was approved by the institutional review board of Seoul National University Bundang Hospital [IRB# B-1710-429-102], and written informed consent was obtained from all patients or their legal representatives.

Using a standardized protocol (Kim et al., 2014), we prospectively collected demographic data, vascular risk factors (hypertension, diabetes mellitus, hyperlipidemia, coronary artery disease, atrial fibrillation, and smoking history), prior medication use, pre-stroke functional status, and index stroke characteristics, such as initial stroke severity (NIH Stroke Scale, NIHSS) and subtypes. Stroke subtypes were determined by an experienced vascular neurologist, using a validated MRI-based classification system built on the TOAST (Trial of ORG 10172 in Acute Stroke Treatment) criteria (Ko et al., 2014).

All perfusion MRI scans were performed on either 3.0 T (62.3%) or 1.5 T (37.7%) scanners. Regarding the vendors, 34.1% of scans were conducted using GE systems, 60.2% using Philips systems, and 5.7% using Siemens systems, all equipped with an 8-channel head coil. Dynamic susceptibility contrast-enhanced perfusion imaging was performed using a gradient-echo echo-planar imaging (GE-EPI) sequence. The imaging parameters were as follows: repetition time (TR) = 1,000–1,500 ms (6.3%), 1,500–2,000 ms (66.7%), or 2,000–2,500 ms (27.0%); echo time (TE) = 30–40 ms (1.0%), 40–50 ms (91.8%), or 60–70 ms (7.2%); field of view (FOV) = 210 × 210 mm² (5.7%), or 230 × 230 mm² (94.3%); and slice thickness of 5 mm with no interslice gap, covering the entire supratentorial brain with 17–25 slices. Images were reconstructed and exported in DICOM format for subsequent post-processing and quantitative perfusion analysis. To minimize inter-scanner variability, all datasets underwent standardized preprocessing and normalization prior to PWI mapping. All image analyses were done in the central image laboratory operated by Seoul National University Bundang Hospital.

For infarct core estimation, RAPID employed the default threshold of ADC < 620 × 10⁻⁶ mm²/s. JLK PWI utilized a deep learning–based infarct segmentation algorithm applied to the b1000 DWI images, which was developed and validated in previous studies using large manually segmented datasets (Ryu et al., 2023, 2024, 2025).

As illustrated in Figure 1A, JLK PWI performs automated preprocessing and perfusion parameter calculations through a multi-step pipeline. The workflow includes motion correction to acquisition artifacts, brain extraction by skull stripping and vessel masking, and conversion of MR signal. The software automatically selects the arterial input function and venous output function, followed by block-circulant single value deconvolution and calculation of quantitative perfusion maps, including CBF, CBV, MTT, and Tmax.

The infarct core from JLK-DWI is automatically co-registered to the perfusion maps, allowing mismatch computation between diffusion and perfusion lesions. The hypoperfused region was delineated using the threshold of Tmax >6 s. All segmentations and resulting images were visually inspected to ensure technical adequacy before inclusion in the analysis. Figure 1B presents a representative case comparing infarct core and hypoperfusion segmentation between JLK PWI and RAPID.

Descriptive statistics were used to summarize baseline characteristics. Continuous variables were reported as means with standard deviations (SD) or medians with interquartile ranges (IQR), depending on data distribution. Categorical variables were presented as counts with percentages. Agreement between the two platforms in perfusion parameter measurements (ischemic core volume, hypoperfused volume, and mismatch volume) was assessed using concordance correlation coefficients (CCC), Pearson correlation coefficients, and Bland–Altman plots. The magnitude of agreement was classified as: poor (0.0–0.2), fair (0.21–0.40), moderate (0.41–0.60), substantial (0.61–0.80), and excellent (0.81–1.0) (Landis and Koch, 1977).

For EVT eligibility, classification agreement between the RAPID and JLK software was evaluated using Cohen’s kappa coefficient, applied separately for each subgroup defined by the DAWN and DEFUSE-3 trial criteria. The DAWN classification stratified eligible infarct volume based on age and NIHSS into three prespecified categories, while the DEFUSE-3 classification used a mismatch ratio ≥1.8, an infarct core volume <70 mL and an absolute volume of penumbra ≥15 mL. Cases with discordant EVT eligibility classifications were additionally analyzed descriptively.

Subgroup analyses were conducted for patients with anterior circulation large vessel occlusion (including internal carotid artery, middle cerebral artery M1-M2 branches, and anterior cerebral artery) and those with basilar artery occlusion. In each subgroup, agreement metrics and outcome prediction models were separately generated to evaluate software performance across stroke types. Additional analyses stratified by MRI vendor and field strength were conducted to assess the consistency of agreement across acquisition settings.

All statistical analyses were performed using STATA version 16.0 (StataCorp LLC, College Station, TX) and R version 4.2.3 (R Foundation for Statistical Computing, Vienna, Austria). A two-sided p-value < 0.05 was considered statistically significant.

For 299 subjects included, the mean age was 70.9 years (SD 11.6), and 55.9% were male. The median NIHSS score on admission was 11 (IQR: 5–17). The most common stroke subtype was cardioembolism (45.2%), followed by large artery atherosclerosis (29.1%) and undetermined etiology (13.0%). Intravenous thrombolysis was administered in 157 patients (52.5%).

Regarding occlusion sites, 208 (69.6%) subjects had anterior circulation large vessel occlusion, and 31 had basilar artery occlusion (10.4%). Meanwhile, 60 (20.1%) subjects had no large vessel occlusion on MRI. The median time from the last known well to PWI was 360 min (IQR: 216–750) min, and the median time from PWI to groin puncture was 55.5 min (IQR: 40.8–82.3).

Ischemic core volumes showed high agreement between RAPID and JLK PWI, with CCC = 0.87 (95% CI, 0.77–0.94; Figure 2B). The Bland–Altman plot showed a mean difference of −4.05 mL and limits of agreement ranging from −41.62 to 33.53 mL (Figure 2A). Similarly, hypoperfused volumes showed high agreement (CCC = 0.88 [95% CI, 0.80–0.93]; Figure 2D). The mean difference was 2.46 mL, with limits of agreement from −59.37 to 64.30 mL (Figure 2C). Mismatch volumes demonstrated substantial agreement (CCC = 0.78 [95% CI, 0.69–0.84]; Figure 2F), with a mean difference of 6.51 mL and limits of agreement from −68.86 to 81.88 mL (Figure 2E). Overall concordance was good, although relatively large volumetric discrepancies were observed in some subjects, as reflected in the wide limits of agreement (see Table 1).

Subgroup analyses for patients with anterior circulation large vessel occlusion (Supplementary Figure 1) and basilar artery occlusion (Supplementary Figure 2) showed similar trends in agreement across core, hypoperfusion, and mismatch volumes. In the basilar artery occlusion group (n = 31), ischemic core volumes demonstrated high agreement between RAPID and JLK PWI (CCC = 0.95 [95% CI, 0.88–0.97]), whereas hypoperfusion volumes showed moderate agreement with CCC = 0.55 (95% CI, 0.31–0.81).

Additional subgroup analyses by different field strengths and MRI vendors demonstrated consistently high concordance across scanner types, with comparable CCC values for ischemic core, hypoperfused, and mismatch volumes (Supplementary Figures 3–7).

To assess the concordance in determining eligibility for EVT, we applied the strict inclusion criteria from the DAWN and DEFUSE-3 trials to the relevant subgroups within our patient cohort. For the DAWN trial criteria, the analysis included 123 patients with an anterior circulation large vessel occlusion and an initial NIHSS score of 10 or higher. In this subgroup, the agreement between RAPID and JLK PWI was excellent (Cohen’s κ = 0.873; 95% CI, 0.773–0.973; Table 2, Figure 3A).

For the DEFUSE-3 trial criteria, the analysis included 163 patients with an anterior circulation large vessel occlusion and an initial NIHSS score of 6 or higher. The agreement between the platforms was substantial (Cohen’s κ = 0.761; 95% CI, 0.660–0.862). Both platforms concordantly identified 58 patients as eligible and 86 as ineligible. There were 19 discordant cases, with 15 deemed eligible by JLK PWI only and 4 by RAPID only (Table 2, Figure 3B). A detailed breakdown of the discordant classifications for DEFUSE-3 is provided in Supplementary Table 1. For the 15 patients deemed eligible only by JLK PWI, the most common reason for ineligibility by RAPID was an ischemic core volume >70 mL (13 of 15 cases). For the four patients deemed eligible only by RAPID, the primary reasons for ineligibility by JLK PWI were a mismatch volume <15 mL (3 of 4 cases) and a mismatch ratio <1.8 (1 of 4 cases).