Norms for Automatic Estimation of White Matter Hyperintensities Burden: LST-AI Service in neuGRID
Background: White Matter Hyperintensities (WMH) are common MRI markers of cerebral small-vessel disease and are associated with cognitive impairment and dementia. Deep Learning (DL) tools have improved WMH segmentation, enabling faster and more reproducible lesion quantification. However, the lack of normative reference frameworks limits the clinical and translational use of WMH volumes. Aims: to develop and validate normative reference curves for automated WMH quantification and to define clinically useful thresholds for rule-out and rule-in interpretation in memory-clinic settings. Methods: We developed age- and sex-adjusted WMH normative models for 2D (n = 788) and 3D (n = 895) FLAIR acquisitions in cognitively normal individuals aged 40–95 years. WMH volumes were segmented using LST-AI and normalized to total intracranial volume. Normative percentiles were derived using Generalized Additive Models for Location, Scale and Shape (GAMLSS) with a Johnson’s SU distribution and externally validated in 458 individuals spanning cognitively normal (CN), mild cognitive impairment (MCI), and dementia groups from two validation cohorts. Normative distributions have been made available through neuGRID, an online platform providing AI-based tools for neuroimaging analysis. Results: WMH burden increased progressively with age in both normative datasets and showed a stepwise increase across the cognitive continuum from CN to MCI and dementia. The optimal balanced thresholds corresponded to the 92nd percentile for the 2D model and the 85th percentile for the 3D model, yielding areas under the receiver operating characteristic curve (ROC-AUCs) of 0.71 (95% CI: 0.63–0.78) and 0.68 (95% CI: 0.61–0.74), respectively. Secondary threshold analyses highlighted complementary operating characteristics, with the 2D model favoring sensitivity and the 3D model favoring specificity, supporting their potential use in sequential diagnostic workflows. Conclusion: This study provides an externally validated normative framework for interpreting LST-AI-derived WMH burden through the neuGRID single-case service. The proposed modality-specific norms enable standardized identification and contextualization of elevated WMH burden and may support clinical stratification, second-opinion assessment, and research applications in cognitive disorders.