Benchmark White Matter Hyperintensity Segmentation Methods Fail on Heterogeneous Clinical MRI: A New Dataset and Deep Learning–Based Solutions

Existing automated methods for white matter hyperintensity (WMH) segmentation often generalize poorly to heterogeneous clinical MRI due to variability in scanner types, field strengths, and protocols. To address this challenge, we introduce a diverse clinical WMH dataset and evaluate two deep learning-based solutions: an nnU-Net model trained directly on the data and a foundation model adapted through fine-tuning. This retrospective study included 195 routine brain MRI scans acquired from 71 scanners between June 2006 and October 2022. Participants ranged in age from 46 to 87 years (median, 70 years; 94 females). WMHs were manually annotated by an experienced rater and reviewed under neuroradiologist supervision. Several benchmark segmentation methods were evaluated against these annotations. We then developed Robust-WMH-UNet by training nnU-Net on the dataset and Robust-WMH-SAM by fine-tuning MedSAM, a vision foundation model. Benchmark methods demonstrated poor generalization, frequently missing small lesions and producing false positives in anatomically complex regions such as the septum pellucidum. Robust-WMH-UNet achieved superior accuracy (median Dice similarity coefficient [DSC], 0.768) with improved specificity, while Robust-WMH-SAM attained competitive performance (median DSC up to 0.750) after only limited training, reaching acceptable accuracy within a single epoch. This new clinically representative dataset provides a strong foundation for developing robust WMH algorithms, enabling fair cross-method comparisons, and supporting the translation of segmentation models into routine clinical practice.