Automated delineation of enlarged perivascular spaces (EPVS)

Abstract

Background Enlarged Perivascular Spaces (EPVS) are associated with small vessel disease. The objective of this study was to develop semi‐automated techniques for the quantification of EPVS burden. Method An intensity‐based threshold was used to identify EPVS. A threshold of maximum plus two standard deviations of intensities of normal‐appearing white matter in the supratentorial region was used to identify EPVS. To compare the performance of the semi‐automatic approach for EPVS segmentation to visually rated EPVS scores, two regions of interest were outlined in the basal ganglia (BG) region and the centrum semiovale (CS). For both these regions, a score was generated by an experienced neurologist by visually counting the number of EPVS with a score of zero representing the absence of any EPVS, a score of 1 representing 1‐10 EPVS, a score of 2 representing 11‐20 EPVS, a score of 3 representing 21‐40 EPVS, and a score of 4 representing >40 EPVS. The semi‐automatic approach provided a count of the number of EPVS that were converted to grade using the same scaling as the visual approach. Spearman Rank Order Correlation Coefficient was calculated to estimate the correlation between the two methods. Mann‐Whitney U tests of ranks are employed to determine if any group differences are present. Finally, Cohen’s kappa was used to compare the concordance between measures between the two different methods. Result Figure 2 shows the results of the segmentation in subjects with differing EPVS burden. Spearman Rank Order Correlation analysis showed a significant correlation between the EPVS scores determined visually and those obtained semi‐automatically in 60 subjects (Table 1). Results of the Mann‐Whitney U tests showed that there was no statistically significant (p > 0.05) difference between the EPVS scores obtained visually and using the semi‐automatic approach (Table 2). Cohen's κ test showed that there was moderate agreement between the two estimation methods in the BG (κ = .522, p < .0001) and CS regions (κ = .571, p < .0001). Conclusion The developed method showed good concordance with expert visual read and can provide quantitative measures of EPVS burden (counts and volume) for evaluation of vascular risk.

A critical guide to the automated quantification of perivascular spaces in magnetic resonance imaging

“…Multiple algorithms have been developed to facilitate or improve the consistency of PVS grading ( Descombes et al, 2004 ; Zong et al, 2016 ; González-Castro et al, 2017 ; Dubost et al, 2019a , b ; Jung et al, 2019 ; Sepehrband et al, 2019 ; Goryawala et al, 2021 ; Williamson et al, 2022 ). For example, Jung et al (2019) applied a convolutional neural network (CNN) to increase the signal-to-noise ratio (SNR) of PVS, thereby enhancing its appearance and improving detection ( Figure 5 ; Jung et al, 2019 ).…”