Benchmarking Generalizability in Deep Learning-Based White Matter Tract Segmentation

White matter tracts (WMTs) are the brain's structural foundation for information transfer, underlying essential cognitive and behavioral functions. While diffusion MRI and tractography enable non-invasive mapping of these pathways, automated segmentation often lacks generalizability across diverse data sources. We conducted a systematic, cross-dataset evaluation of four state-of-the-art deep learning architectures, benchmarking their performance across independent datasets with varying acquisition protocols and populations. CNN-based models such as TractSeg achieved the highest within-domain accuracy, but performance dropped sharply under domain shift, most severely when we applied adult-trained models to pediatric data. To address this degradation, we introduce Ensemble White Matter Tract Segmentation (EWMTS), which combines complementary models to partially recover accuracy under domain shift, although performance still falls short of within-domain levels. By openly releasing this benchmark and a reproducible processing pipeline, we provide the neuroimaging community with a framework to develop and benchmark segmentation models across the heterogeneity of real-world neuroimaging data.