Automated volume estimate of the brain's lateral ventricles reveals a persistent association with cognitive
Bradley J MacIntosh1,2,3, Pål Cw Morberg4,5,6, Elakkyen Murugesu1
1Computational Radiology & Artificial Intelligence Unit, Department of Physics and Computational Radiology, Oslo University Hospital, Oslo, Norway.
Background:
Post-stroke cognitive impairment (PSCI) is a significant barrier to recovery. While computed tomography (CT) is the dominant modality for acute stroke diagnosis, it can also quantify brain anatomy, such as lateral ventricular volume (LVV). This study investigates whether initial LVV is associated with Montreal Cognitive Assessment (MoCA) scores over long-term follow-up.
Methods:
Initial head CT, longitudinal MoCA, and functional assessments were collected at sites from the Norwegian COgnitive impairment After STroke (Nor-COAST) study. LVV was estimated using an established deep learning segmentation tool. Associations between initial LVV and repeat MoCA were tested across four timepoints (up to three years) using a linear mixed effects model. To provide context, we explored the independent influence of LVV in relation to National Institutes of Health Stroke Scale (NIHSS), modified Rankin Scale (mRS), and Global Deterioration Scale (GDS).
Results:
The sample of N = 547 participants were 74 ± 12 years (range: 35-97 years) and 54% female. Log-transformed LVV was inversely associated with longitudinal MoCA scores (t=-3.75, p < 0.001). This relationship was consistent across cross-sectional analyses at individual time points. Secondary analyses revealed significant LVV associations with NIHSS, mRS, and GDS scores (all p < 0.05).
Conclusion:
LVV was independently associated with lower MoCA scores across the three-year follow-up after stroke. Specifically, a patient with twice the initial LVV scored 0.7 points lower on the MoCA across three years of follow-up. These findings reveal that CT obtained at stroke admission may yield prognostic information about long-term cognitive outcomes, supporting the case for automated LVV estimation.
