Artificial intelligence for white matter hyperintensity segmentation: performance, determinants, and clinical implications

Abstract

Objective White matter hyperintensities (WMH) are key neuroimaging markers of cerebral small vessel disease, associated with cognitive decline and stroke risk. Accurate quantification is essential yet manual segmentation is time-consuming and variable. We systematically evaluated AI performance for automated WMH segmentation across disease contexts and identified factors influencing model performance. Methods Following PRISMA guidelines, we searched PubMed, Embase, Web of Science, Scopus, and IEEE Xplore from database inception to February 3, 2026 (PROSPERO: CRD420261355769). Studies employing AI-based WMH segmentation and reporting Dice similarity coefficient (DSC) were included. A random-effects meta-analysis was conducted, along with subgroup analyses, meta-regression, sensitivity analyses, and publication bias assessment. Results 26 studies comprising 4,288 participants contributed to quantitative syntheses: 13 evaluated WMH segmentation in general cohorts, and 13 evaluated AD or CSVD-specific cohorts. In the general cohort meta-analysis, the pooled DSC was 0.78 (95% CI, 0.74–0.82). Disease-specific pooled DSC estimates were 0.75 for AD and 0.78 for CSVD. Exploratory meta-regression suggested an association between scanning parameters and DSC (p = 0.042), whereas magnetic field strength showed a borderline association (p = 0.052). Conclusion AI demonstrates promising performance for automated WMH segmentation, with U-Net–based architectures showing superior accuracy. However, substantial heterogeneity highlights the need for standardized imaging protocols and robust external validation. Future multicenter studies are essential to improve model generalizability and facilitate clinical translation.