Multi-center evaluation of an automated pipeline to segment perivascular spaces from routine MRI in patients with small vessel disease

Abstract

Introduction: Enlarged perivascular spaces (EPVS) that are visible on conventional MRI sequences are considered a marker of small vessel disease (SVD)[1]. To date, the evaluation of PVS has mostly relied on semi-quantitative evaluation from expert raters, which is time consuming and suffers of intra-rater variability. In this study, we evaluate a previously validated automated PVS segmentation pipeline [2] in a multi-center SVD cohort. Methods: We considered data from the TRACE-VCI cohort[3] (n=356; 251/105 from Amsterdam/Utrecht). T1w, T2w and FLAIR MRI scans were visually assessed by an expert in midbrain, basal ganglia (BG) and centrum semiovale (CSO). The visual rating was according to the Potter scale[4]. Automated EPVS segmentations were performed on T1w images only, given the insufficient resolution of T2w and FLAIR scans. EPVS segmentation optimization consisted of tuning: filter type (Frangi vs RORPO[5]); segmentation threshold(δ). The default value (0.5) of the hyperparameters α (tube vs plate-like structures) and β (tube vs blob-like structures) was used for the Frangi filter. First, we optimized the hyperparameters separately in the two cohorts in a random subset of 50 subjects, and evaluated the spearman correlation between automated EPVS count and the corresponding expert ratings. Second, the pipeline was run with the optimized hyperparameters on all subjects. Results: Fig. 1 shows the results of the threshold optimization for the FRANGI filter. For the midbrain, only modest correlations were observed (<0.35). For the BG, a maximal correlation between 0.6 and 0.7 was observed in both cohorts. Conversely, correlations up to 0.7 were observed for CSO in the Utrecht cohort, as compared to about 0.35 for the Amsterdam cohort. Similar results were observed for RORPO. A qualitative observation of the optimized segmentations (Fig. 2) suggested a plausible agreement between automated EPVS count and visual ratings. Fig. 3 reports a moderate correlation between visual ratings and automated EPVS count, with better results in the Utrecht cohort. Discussion: We have shown that automated EPVS segmentations are moderately correlated to expert evaluations. Hyperparameters optimization has a large effect on the segmentation performance and seems to be site dependent, suggesting the need to optimize hyperparameters