Abstract
Artificial intelligence (AI) is increasingly used in cerebral small vessel disease (CSVD) imaging, but the extent of validation and clinical applicability across the literature remains uncertain. We performed a study-level cross-sectional analysis using a frozen Web of Science Core Collection (WoSCC) cohort supplemented by PubMed and IEEE Xplore searches to assess database coverage and the stability of findings from the WoSCC cohort. After study-level reconciliation, 463 independent studies were included. White matter hyperintensity segmentation/quantification was the most common task (196/463, 42.3%), followed by cerebral microbleed detection/classification (82/463, 17.7%), perivascular space/lacune assessment (67/463, 14.5%), CSVD burden/risk modeling (62/463, 13.4%), and clinical outcome prediction (56/463, 12.1%). Deep learning was the largest model family (185/463, 40.0%). Higher validation levels (L4-L5) were reached by 146/462 studies (31.6%), external or multicenter validation was reported in 147/463 (31.7%), and higher clinical applicability evidence was identified in 159/462 (34.4%). Calibration (51/438, 11.6%), decision curve or net-benefit analysis (40/436, 9.2%), and reader-, workflow-, or implementation-based evaluation (34/436, 7.8%) remained uncommon. Broader database coverage materially changed the task and model composition of the evidence base, particularly by increasing the representation of white matter hyperintensity and image-processing studies, but did not alter the central finding that high-level validation and clinically oriented evaluation remain limited.