On the detectability and accuracy of computational measurements of enlarged perivascular spaces from magnetic resonance images

Magnetic Resonance Imaging (MRI) visible perivascular spaces (PVS) have been associated with age, decline in cognitive abilities, interrupted sleep, and markers of small vessel disease. Therefore, several computational methods have been developed for their assessment from brain MRI. But the limits of validity of these methods under various spatial resolutions, and the accuracy in detecting and measuring the dimensions of these structures have not been established. We use a digital reference object (DRO) previously developed for this purpose, to construct an in-silico phantom for answering these questions; and validate it using a physical phantom. Our in-silico and physical phantoms use cylinders of different sizes as models for PVS. Using both phantoms, we also evaluate the influence of the "PVS" orientation on the accuracy of the diameter measured, different sets of parameters for two vesselness filters that have been used for enhancing tubular structures, namely Frangi and RORPO filters, and the influence of the vesselness filter per-se in the accuracy of the measurements. Our experiments indicate that PVS measurements in MRI are only a proxy of their true dimensions, as the boundaries of their representation are consistently overestimated. The success in the use of the Frangi filter for this task relies on a careful tuning of several parameters. The combination of parameters =0.5, {beta}=0.5 and c=500 proved to yield the best results. RORPO, on the contrary, does not have these requirements, and allows detecting smaller cylinders in their entirety more consistently in the ideal scenarios tested. The segmentation of the cylinders using the Frangi filter seems to be best suited for voxel sizes equal or larger than 0.4 mm-isotropic and cylinders larger than 1 mm diameter and 2 mm length. "PVS" orientation did not influence their measures for image data with isotropic voxel size. Further evaluation of the emerging deep-learning methods is still required, and these results should be tested in "real" world data across several diseases.