All patients participated in the prospective observational 1000Plus study from September 2008 to June 2013 (clinicaltrials.org NCT00715533). The single-center study performed by the Center for Stroke Research Berlin acquired MRI data from 1472 patients presenting to the emergency room with a clinical diagnosis of an acute cerebrovascular event (16). The study design was approved by the institutional review board of the Charité Universitätsmedizin, Berlin (EA4/026/08). The study was carried out in accordance with The Code of Ethics of the World Medical Association (Declaration of Helsinki). Written informed consent was obtained from all patients.

A subset of 126 patients was selected for retrospective analysis. Inclusion criteria included available DWI data, either a vessel occlusion on day 1 with vessel recanalization on day 2 or no vessel occlusion on day 1, as well as the known time of symptom onset. Vessel recanalization on day 2 was required to investigate the degree to which automatically delineated DWI lesions on day 1 reflect final infarct volume (assessed between days 3 and 5 using FLAIR) (17). In the group without vessel occlusion on day 1, we excluded very small strokes (≦3 mL in volume) because we considered that these would not be appropriate for automated threshold-based delineation. Following the exclusion of 18 patients after data processing 108 patients were selected for analysis (Figure 1).

Imaging was performed on a 3T MRI scanner (Tim Trio; Siemens AG, Erlangen, Germany). DWI was performed with a spin-echo echo-planar imaging sequence with a ‘b’ value of 1000 s/mm² (TR/TE = 7600/93 ms, matrix = 192 × 192, field of view = 230 mm, slice thickness = 2.5 mm, slice gap = 0 mm, number of slices = 50, acquisition time = 2 min 11 s). The study protocol also included T2*-weighted images, time-of-flight magnetic resonance angiography (TOF-MRA), and fluid-attenuated inversion recovery (FLAIR) images (16). All participants were scanned relative to the time of stroke symptom onset (TOO) on day 1 (within 24 h) and day 2 (24–48 h). In 78 patients, follow-up imaging data obtained around the fifth day of admission (range: day 3–7, mean 4.8 days) was available.

All images were first converted from DICOM to NIfTI format using dcm2nii software (18). Brain extraction was performed on b0, trace (b = 1000), and ADC map images using BET (Brain Extraction Tool) (19). The brain-extracted b0 images were registered to MNI152 T1-weighted 2 mm isovoxel standard space using three-dimensional diffeomorphic symmetric normalization (using antsRegistrationSyNQuick) (20–24). Registered images were checked visually and unsatisfactory registrations were excluded. Resulting transformation matrices were used for mapping of ADC map and trace images to MNI152 standard space using antsApplyTransforms (20–24). Tissue segmentation of cerebrospinal fluid (CSF) and brain parenchyma was conducted on b0 images using FMRIB’s Automated Segmentation Tool (25). Baseline DWI lesions and follow-up FLAIR lesions were delineated by a group of stroke imaging researchers using MRIcro and supervised by either an experienced radiology resident, a board-certified neuroradiologist, or a radiologist (26). DWI ROIs were mapped nonlinearly to MNI152 standard space using antsApplyTransforms (20–23).

MR imaging data obtained at admission within 24 h of symptom onset were used. After image preprocessing, the brain extracted and registered b0, trace, and ADC map images were read by the algorithm.

A threshold-based lesion segmentation algorithm was developed using the FSL software package (27–29). The approach solely relies on the acquired DWI sequence (the trace and ADC) and outputs a lesion mask that aims to match the expert lesion. It uses artifact reduction techniques known from previous research and uses ADC thresholds already implemented in commercially available software (8, 15). First, the algorithm localizes the infarct and creates a primary lesion mask (see Figure 2). Artifact reduction and lesion contrast enhancement are achieved by dividing the trace image by the ADC map. Resulting relative voxel values inherit greater contrast between healthy and ischemic tissue. To assure comparability between subjects and brain regions, normalization of relative values is performed in every slice separately by dividing the voxel’s signal intensity by the average voxel value of the whole slice. Consecutively, smoothing using an isotropic 5 voxel Gaussian kernel is performed. Mirroring along the x-axis and subtraction of the contralateral hemisphere further reduces artifacts. A lower threshold of 0.75 image intensity is applied to the resulting processed image based on testing lesion coverage on a subset of 5 randomly chosen subjects. Partial dilatation completes the mask creation. Finally, the delineated infarct region includes areas within the mask with ADC values between 200 and 620×10⁻⁶ mm²/s (8). Removal of scattered trace-hypointense voxels is achieved by removing voxels with trace intensities below the 95th percentile (Figure 2).

Statistical analysis of this retrospective, exploratory study was performed using R Studio Version 2022.07.1 + 554 (30). The Dice coefficient

was used to study spatial overlap between the automated, and the manual delineated ROIs (31). Volumetric agreement between manual and automated DWI lesion volumes as well as between initial lesion and follow-up FLAIR lesion volumes was investigated using Bland–Altman analysis (32). The correlation of manual and automated delineated lesion volumes was assessed by calculating Pearson’s correlation coefficient (33). A whole-brain ROC analysis was conducted in every subject using the fslmaths function of the FSL software package to determine the optimal ADC threshold for distinguishing infarcted from healthy tissue (28, 29). Therefore, voxel-wise analysis was performed between ADC values of manually delineated day one DWI lesions and the whole brain parenchyma ADC map. A second analysis used the same ground truth, but a different ADC map centered on the DWI lesion. This was achieved by dilating the DWI lesion using the fslroi function from the FSL toolbox and applying the resulting mask to the ADC map (29). ROC analysis output was pooled across subjects to identify a generalized threshold. The Youden index

was used to determine the best possible threshold (34). Multivariate and univariate linear regression was used to investigate whether initial DWI lesion size, DWI lesion ADC value, mean ADC in the brain parenchyma, age or TOO would be associated with the performance of the segmentation algorithm itself and the diagnostic capability of the ADC threshold found in the patient. As the performance of the algorithm was measured in spatial overlap, the Dice coefficient was the response variable in the first model. As diagnostic capability was assessed with the Youden index (YI) it represented the response variable in the second model. Assumptions for using linear models were checked, assessing skewness, kurtosis, link function and heteroscedasticity with the R packages “lmtest,” “gvlma” and “corrplot” (35–37). Visualization was accomplished using the R package “ggplot2” (38).

The image processing scripts, segmentation algorithm, and statistical analysis scripts are openly accessible.

The baseline characteristics of the 108 patients analyzed are summarized in Table 1.

After applying the segmentation algorithm to imaging data acquired within 24 h of symptom onset (median 247 min, IQR 108–737 min), the resulting delineations were compared to manually delineated ROIs. The median ADC value in the automated ROIs was 504 × 10⁻⁶ mm²/s (IQR 421–533 × 10⁻⁶ mm²/s) compared to 686 × 10⁻⁶ mm²/s (IQR 591–780 × 10⁻⁶ mm²/s) in the manual ROIs (see Figure 3A). Automated ROIs had a median lesion size of 5.5 mL (IQR 1.4–12.7 mL), and the median volume of manual ROIs was 6.9 mL (IQR 1.8–16.5 mL) (Figure 3B).

The Bland Altman analysis showed a mean bias of −4.92 mL with an upper limit of agreement of 19.36 mL and a lower limit of agreement of −29.21 mL (Figure 4A). The volumes of automated and manual ROIs correlated strongly, r(106) = 0.79 (CI = 0.7–0.85), p-value < 0.0001 (Figure 4B). Linear regression had a slope of 0.49 (adjusted R² = 0.62, p-value < 0.0001). In two outlier subjects, the algorithm did not detect an infarction. The median Dice coefficient for the automated segmentation was 0.43 (IQR 0.20–0.64) (Figure 5A). No spatial overlap was seen in 15 lesions with low volume (manually delineated lesion size IQR: 0.2–0.8 mL, median 0.46 mL) (see Figure 5B).

The optimal generalized threshold to match the expert ischemic lesion delineation in the brain parenchyma using pooled ROC analysis was an ADC ≤704 × 10⁻⁶ mm²/s (sensitivity 65% and specificity 77%, AUC 0.76). Narrowing down the search area by a median of 96% (IQR 0.93–0.97) to the vicinity of the infarct resulted in an optimal pooled threshold of ≤693 × 10⁻⁶ mm²/s (sensitivity 63% and specificity 75%, AUC 0.74). In both cases, individual ROC curves differed substantially from those generated using pooled data (see Figure 6). The individual whole-brain ROC analysis determined thresholds with a median of 691 × 10⁻⁶ mm²/s (IQR 660–750 × 10⁻⁶ mm²/s; sensitivity median 75%, IQR 64–86%; specificity median 80%, IQR 73–83%).

Sixteen patients with unknown time from symptom onset to imaging were excluded from this analysis. At first, predictors of the performance of the segmentation algorithm were assessed. The overall regression was statistically significant (adjusted R-squared = 0.54, F(5, 86) = 22.03, p < 0.001). The mean ADC value in the DWI lesion (β = −0.65, [−0.80 – −0.50], p < 0.001) and the DWI lesion volume (β = 0.28, [0.13–0.42], p < 0.001) were significantly associated with the Dice coefficient. The mean ADC values in the brain parenchyma (β = 0.16, [−0.03–0.35], p = 0.09), time from onset to imaging (β = 0.01, [−0.13–0.16], p = 0.84) and patient age (β = −0.17, [−0.36–0.01], p = 0.07) were not.

Secondly, variables possibly impacting the identification of individual ADC thresholds as determined by ROC analysis (compare Figure 6) were analyzed. The dependent variable was the Youden index (YI). The overall regression was statistically significant (adjusted R-squared = 0.87, F(5, 86) = 118.7, p < 0.001). Significant predictors for the Youden index were mean ADC values in the DWI lesion (β = −0.96, [−1.04 – −0.87], p < 0.001), mean ADC values in the brain parenchyma (β = 0.28, [0.18–0.38], p < 0.001), DWI lesion volume (β = −0.11, [−0.18 – −0.03], p = 0.006) and patient age (β = −0.14, [−0.25 – −0.04], p = 0.006). Time from onset to imaging (β = 0.05, [−0.03–0.12], p = 0.20) did not show a significant association. Visualization of the results of the regression analysis and additional univariate models can be accessed in the Supplementary Figures S1–S3.