Super-resolution (SR) with supervised learning has emerged as a viable tool to enhance 4D flow magnetic resonance imaging (4DMR) data post hoc, but its impact on carotid hemodynamic measurements remains unclear. In this study, we quantified uncertainties in a phantom carotid artery model by comparing native 4DMR with and without SR against computational fluid dynamics (CFD) results. A phantom carotid artery model with steady flow was scanned at 3 Tesla. Magnitude images were segmented, and phase images were converted to velocities (3T). A pretrained 4DFlowNet was implemented to produce SR data (3T SR). A CFD simulation was performed on the same set up, serving as a reference. Uncertainties of geometrical processing were assessed by surface and Hausdorff distances. Hemodynamic metrics included intraluminal velocity and vorticity statistics, flow rate, and wall shear stress (WSS). A paired two-sided Wilcoxon signed-rank test was used to test the comparisons. Segmentations yielded a mean ± SD surface distance of 0.31 ± 0.06 mm (38 ± 7.5% of voxel size) and a Hausdorff distance of 1.2 mm. SR processing improved flow patterns and statistics of intraluminal hemodynamic parameters (p = 0.033). For flow rate, standard deviation decreased from 14.7% (3T) to 2.8% (3T SR), but mean values showed larger deviation after SR processing. Regarding WSS, qualitative improvements were observed, but quantitative comparisons did not show significant difference (p = 0.20). Our results suggest that SR processing stabilizes flow rate estimates and improves intraluminal hemodynamic statistics. But applying SR on data with inadequate quality may lead to larger errors. Further efforts are needed to improve generalizability.
