Figure 1

Model performance for infarct detection (a, ROC curve) and delineation (b, scatterplot; c, Bland–Altman plot; d, confusion matrices) in the “MCA-territory only” test set (see Table 1), based on DWI ground truth, compared to three human experts. (a) Model AUC was 0.95; sensitivity/specificity were 0.96/0.72 at a 0 mL-threshold operating point for infarct detection, 0.82/0.92 at a 1 mL-threshold, and 0.78/0.98 at a 5 mL-threshold for infarct detection, compared to mean reader sensitivity/specificity of 0.64/0.91. (b) Model infarct volume estimates strongly correlated with those of DWI ground truth (r² > 0.98). As per the Bland–Altman plot (c), the model had excellent performance for estimating infarcts smaller versus larger than 50 mL (95%CI <  ± 17 mL), the volume threshold used for patient selection in major late window stroke treatment trials. Expert interrater Cohen’s kappa values ranged from 0.42 to 0.48, suggesting significant variability compared to the model, confirmed by the confusion matrices for volume segmentation (d, mean study-counts-per-category and ranges shown for the 3-experts; calculated at the model’s 0 mL-threshold for infarct detection).