Quantitative assessment of enlarged perivascular spaces via deep-learning in community-based older adults reveals independent associations with vascular neuropathologies, vascular risk factors and cognition

Abstract

Abstract Enlarged perivascular spaces (EPVS) are common in older adults, but their neuropathologic correlates are unclear mainly because most work to date has relied on visual rating scales and/or clinical cohorts. The present study first developed a deep-learning model for automatic segmentation, localization and quantification of EPVS in ex vivo brain MRI, and then used this model to investigate the neuropathologic, clinical and cognitive correlates of EPVS in 817 community-based older adults that underwent autopsy. The new method exhibited high sensitivity in detecting EPVS as small as 3 mm3, good segmentation accuracy and consistency. Most EPVS were located in the frontal lobe, but the highest density was observed in the basal ganglia. EPVS in the cerebrum and specifically in the frontal lobe were associated with infarcts independent of other neuropathologies, while temporal and occipital EPVS were associated with cerebral amyloid angiopathy. EPVS in most brain lobes were also associated with diabetes mellitus independently of neuropathologies, while basal ganglia EPVS were independently associated with hypertension, supporting the notion of independent pathways from diabetes and hypertension to EPVS. Finally, EPVS were associated with lower cognitive performance independently of neuropathologies and clinical variables, suggesting that EPVS represent additional abnormalities contributing to lower cognition.