Comparative Analysis of Artificial Intelligence-Based Quantification versus Visual Rating of Enlarged Perivascular Spaces in the Multi-Ethnic Study of Atherosclerosis Cohort
Clinical Snapshot
PICO Framework
| P — Population | 235 participants from the Multi-Ethnic Study of Atherosclerosis cohort who underwent brain MRI (mean age 72.1 years, 40% male, 54% Black) |
| I — Intervention | AI-based automated quantification of enlarged perivascular spaces (PVS) |
| C — Comparator | Visual rating by expert neuroradiologist using semiquantitative scale |
| O — Outcomes | Association with vascular risk factors (age, race, systolic BP, diabetes) and cognitive performance (global cognition, delayed memory, attention/processing speed) |
Bottom Line
This cross-sectional study of 235 older adults from the Multi-Ethnic Study of Atherosclerosis demonstrates that automated deep learning quantification of enlarged perivascular spaces is more sensitive than expert visual rating in detecting associations with vascular risk factors and cognitive performance. The AI method identified significant relationships with systolic blood pressure, diabetes, and multiple cognitive domains that were missed by traditional visual assessment. While these findings suggest AI-based quantification could advance large-scale neuroimaging research, the clinical utility remains to be established. The cross-sectional design limits causal inference, and the technology requires sophisticated infrastructure not readily available in routine practice. For Australian clinicians, this represents an emerging research tool that may eventually enhance our understanding of small vessel disease and its cognitive consequences, but current clinical decision-making should continue to rely on established neuroimaging markers and comprehensive clinical assessment.
Key Findings
P Value: Not explicitly reported but confidence intervals exclude null for key associations
Effect Size: For basal ganglia PVS: systolic BP β=0.20, diabetes β=11.51, global cognition β=-0.012
Primary Outcome: AI-derived PVS counts showed stronger associations with vascular risk factors and cognitive outcomes compared to visual ratings
Nnt Or Sensitivity: AI method detected additional significant associations not found with visual rating method
Confidence Interval: Systolic BP (0.05-0.36), diabetes (3.48-19.55), global cognition (-0.023 to -0.0004)
Clinical Application
Requires access to advanced MRI and deep learning infrastructure, may be challenging for routine clinical practice Could enhance research capabilities in Australian dementia and stroke prevention programs, though implementation would require significant technological investment and validation in local populations Older adults (≥65 years) undergoing brain MRI for vascular or cognitive assessment
Abstract
BACKGROUND AND PURPOSE: Enlarged perivascular spaces (PVS) are fluid-filled spaces surrounding small cerebral vessels. Current PVS research has been limited by the absence of efficient and scalable quantification tools. We compared visual ratings versus artificial intelligence (AI)-based quantification in identifying associations of PVS with vascular risk factors and cognitive performance. MATERIALS AND METHODS: This cross-sectional study included 235 participants from the Multi-Ethnic Study of Atherosclerosis who had undergone brain MRI and had available both visual ratings and AI-derived PVS quantification. Visual ratings were performed by an expert neuroradiologist (K.D.H.) using a semiquantitative scale. AI-derived counts were based on a fully automated deep learning algorithm. Visual and AI-derived counts were grouped into 4 a priori-defined anatomic locations: basal ganglia, frontoparietal cerebrum, midbrain, and cerebellum. The relationships of PVS counts with demographic characteristics, vascular risk factors, and global and domain-specific cognitive scores were examined using ordinal logistic regression (for ordinal categoric outcomes) and linear regression (for continuous outcomes) models. RESULTS: Mean age (SD) was 72.1 (6.8) years; 95 (40%) participants were men; and 126 (54%) self-reported as black. Means (SD) of AI-derived regional PVS counts were 63.7 (24.6) for basal ganglia, 414.9 (167.5) for frontoparietal cerebrum, and 9.8 (4.4) for midbrain. On visual ratings, the most prevalent count category for each region was 11-20 for basal ganglia (40%), 21-40 (31%) for frontoparietal cerebrum, and 1-5 (83%) for midbrain. For basal ganglia PVS, while both methods were associated with older age and white race/ethnicity, AI-derived counts exhibited additional associations with higher systolic blood pressure (β, 0.20; 95% CI, 0.05-0.36) and diabetes (unstandardized β coefficient [β], 11.51; 95% CI, 3.48-19.55), as well as poorer global cognition (β, -0.012; 95% CI, -0.023 to -0.0004), delayed memory (β, -0.005; 95% CI, -0.010 to -0.0005), and attention/processing speed (β, -0.005; 95% CI, -0.009 to -0.001) cognitive performance. CONCLUSIONS: In this cross-sectional study, AI-derived PVS quantification was more sensitive in detecting associations with vascular risk factors and cognitive outcomes than traditional visual ratings. AI-based quantification may aid in the analysis of large-scale epidemiologic data, advancing PVS research.
References
- 1.Torres, M. F., Charisis, S., Rashid, T., Brandigampala, S. R., Hiatt, K. D., Ware, J. B., Whitlow, C. T., Nasrallah, I. M., Romero, J. R., Tanley, J. E., Seshadri, S., Hohman, T. J., Heckbert, S. R., Davatzikos, C., Hughes, T. M., & Habes, M. (2026). Comparative Analysis of Artificial Intelligence-Based Quantification versus Visual Rating of Enlarged Perivascular Spaces in the Multi-Ethnic Study of Atherosclerosis Cohort. AJNR. American Journal of Neuroradiology. https://doi.org/10.3389/fnagi.2021.624732
Related Research
Sleep medicine reviews
Machine and deep learning in REM sleep behavior disorder: a scoping review and analysis of reporting quality
3 Aug 2026
Neurological research
Federated deep learning model for epilepsy seizure detection using electroencephalogram (EEG) signal
2 Aug 2026
Journal of psychiatric research
Systematic review of machine learning and deep learning models for EEG-based detection of depression.
2 Aug 2026
This content is for educational purposes for healthcare professionals only and does not constitute clinical advice. Clinical decisions should be based on individual patient assessment, current guidelines, and appropriate specialist consultation. Editorial Standards · Privacy Policy · Terms of Service