Automated Segmentation of MRI White Matter Hyperintensities in 8421 Patients with Acute Ischemic Stroke

Segmentation performance of the SE-UNet, with stratification by infarct volume, white matter hyperintensity volume, and the uncertainty index in external test data sets. A, Violin plot for the performance of WMH segmentation, with stratification by infarct volume (low, <1.7 mL, n = 2792; medium, 1.7–14 mL, n = 1785; and large, >14 mL, n = 1436). B, Violin plot for the segmentation performance, with stratification by WMH volume (low ,<7.6 mL, n = 1984; medium, 7.6–19.6 mL, n = 1984; and large, >19.7 mL, n = 2045). C, Segmentation performance stratified by the infarct volume and the uncertainty index. D, Segmentation performance, stratified by the WMH volume and the uncertainty index. Due to the association between the volume of WMHs and the uncertainty index, none of the patients in the large-volume group had an uncertainty index exceeding 0.8. E, The mean and 95% CI of the uncertainty index after stratification of DSCs and WMH volume (in quartiles). In the violin plots, lines and dotted lines indicate median and interquartile range, respectively. In the line graphs, dots and error bars indicate mean (SD), respectively.

Representative cases of automatic white matter hyperintensity segmentation and their uncertainty indices. Upper row: raw images. Middle row: red and blue areas, respectively, show areas of agreement and disagreement between human segmentation and deep learning algorithm-based segmentation. Lower row: adjudicated final results, with uncertainties resolved.

Relationship between the uncertainty index and the DSC. A, Density plot between the uncertainty index and the DSC. As expected, there is a negative correlation between the 2 values. B, Mean DSC, distribution of subjects, and the uncertainty index. The blue line represents the cumulative percentile of patients, while the green line indicates the mean DSC of cumulative patients. The blue dotted line represents the cumulative percentage of patients when the uncertainty index reaches 0.35 from 0 (red dotted line). The green dotted line shows interhuman DSC. At an uncertainty index of 0.35, accumulated patients comprise approximately 86% of the total population, with a mean DSC of 0.75.

Scatterplot showing a strong correlation between the volumes of WMHs segmented automatically by SE-UNet and manually in external validation. Gray dots indicate each subject. The red line indicates a 45° line of identity. Outlier subjects (blue dots) are reviewed in the Online Supplemental Data. The given equation is y = 2.1693 + 0.8311x, where y represents the automatically segmented WMH volume and x represents the manually segmented WMH volume. After eliminating 6 outlier cases from the scatterplot, the slope showed a modest increase, resulting in the equation y = 2.0671 + 0.8376 x. Coef indicates coefficient.