Relationship Between Venules and Perivascular Spaces in Sporadic Small Vessel Diseases

Supplemental Digital Content is available in the text.

. Overview of the process of the development of the region of interest (ROI). A) First ROI, we would miss out on venules and PVS. B) The small ROI contained anterior, middle and posterior sections. This was applied in 8 patients. With this ROI we would still miss out on venules and perivascular spaces (PVS) further away from the ventricles. Thus we enlarged the ROI. C) We applied a larger ROI in only 1 slice since the counts and overlap were similar between the slices and to decrease time to count.

MSS-3:
We used a 3T Siemens Magnetom Prisma MRI scanner with a 32-channel head coil (Siemens Healthcare, Erlangen, Germany) to acquire isotropic axial 3D T2weighted SPACE, sagittal 3D T1-weighted MPRAGE, isotropic sagittal 3D FLAIR SPACE and axial 3D SWI. We performed phase contrast MRI to measure pulsatility in the superior sagittal sinus (SSS), straight sinus (StS) and transverse sinuses (TS) as described previously 2 . We used a 2D cine phase-contrast pulse sequence with retrospective peripheral pulse gating to obtain 32 velocity images per cardiac cycle. To measure venous flow waveforms, we selected a coronal-oblique slice that intersected the SSS at 2 cm above the torcular and through the midpoint of the StS. Retinal images were acquired with a SPECTRALIS imaging platform (Heidelberg Engineering, Heidelberg, Germany). We employed a circular scan around the optic nerve head, generating 30° infrared (IR) images (8.8 x 8.8 mm) of the fundus via a scanning laser ophthalmoscope.

Image processing and analysis
Quantification of venules To assess venules we used GRE and SWI sequences. For the iSVD study, we had pre-and post-contrast GRE scans. The first MRI scan was immediately followed by intravenous infusion of an ultrasmall superparamagnetic particles of iron oxide (USPIO) contrast agent (ferumoxytol). Eight of the twelve patients had an additional scan immediately after infusion. All twelve patients were scanned at 24-30h and at four weeks after infusion 1 . The postcontrast GRE made it more complicated to differentiate between venules and arterioles as the contrast agent might highlight arterioles and venules while vessels visible on pre-contrast GRE are most likely to be venules. Therefore we only used the pre-contrast GRE (from now on mentioned as GRE), however the post-contrast GRE scans were still available for comparison if needed.
All sequences we used were uploaded into Carestream PACS software version 11.3.2.0220 (Carestream Health, Inc., 2011). Most sequences were already lined up. If images were not lined up this was done manually by using linking and referencing tool available in Carestream. We also used drawing tools available in Carestream to draw the ROIs.
We considered hypointense vessel-like structures on the GRE and SWI sequences to be venular. First we applied the ROI and adjusted it to the scan. Then we counted all visible venules per section, i.e. anterior, middle or posterior, within the ROI in both hemispheres. Combining all the counts in both hemispheres (six sections in total) results in a total venular count per slice. If a venule was visible in two sections, e.g. middle and posterior, venule we included the count of the venule in the section where it originated from. In order to avoid double counting of venules, adjacent slices were reviewed. Maximum intensity projections were available for MSS-3. These were used to verify the presence of a venule when the venule was not very clearly visible on the SWI. However, venules were only counted when they were visible on the SWI.

Quantification of PVS
After application of the ROI we counted PVS using T2-weighted images. In the CSO, PVS are seen as rounded or linear hyperintensities depending on direction of imaging slice 3 . PVS were counted in a similar way to venules, per part of ROI and per hemisphere. Combining these counts results in a total CSO PVS count per slice, calculated similar to the above for venules. Similar to venules, to avoid double counting of linear PVS, adjacent slices were reviewed. Since the ROI might overlap sulci in some patients, reviewing adjacent slices also prevents mistaking CSF in sulci as PVS as they might look similar.

Assessment of overlap
First we explored the use of computationally co-registered images using various in-house and open source software, e.g. FSL (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki) and SPM (https://www.fil.ion.ucl.ac.uk/spm/), to improve the assessment of overlap. GRE and SWI were registered to the T2-weighted images. However, the registration process resulted in loss of image resolution which reduced the visibility of venules. Thus we compared the GRE and SWI images with the T2-weighted images to assess overlap. We determined the location of overlap based on a relating point in Carestream in both scans and by comparing the surrounding structures. The T2-weighted images from MSS-3 were resampled to create a similar slice thickness to SWI. We used this as an extra confirmation of PVS, but overlap was based on the original images. Additionally, we used sagittal T1-weighted images for MSS-3 to trace PVS and confirm overlap. We considered venules and PVS to overlap when they had the same location, shape and direction. However, determining overlap proved to be difficult due to reduced visibility of venules and PVS due to WMHs, stroke lesions or movement artefacts. When venules and PVS were less visible it was harder to determine if they had the same location, shape and direction. We created categories of overlap to describe cases where the venule and PVS did seem to overlap despite not having the exactly same shape and direction (see Table I).
Some venules and PVS seemed to only partially overlap, e.g. a PVS could overlap with a venule based on same location, shape and direction, but only for a part of the venule. This overlap was still scored as definite, probable, or possible overlap with the addition 'partial', e.g. partial possible overlap. When overlap could still not be confidently identified based on the categories it was not counted as overlap.
Pulsatility measures: Using phase contrast images, a trained analyst, blinded to clinical, venular and PVS data, manually drew ROIs around the SSS, StS and left and right TS and in adjacent non-vascular background tissue. We subtracted the average background velocity from those of the vessel ROIs, corrected the background phase error and calculated the pulsatility index (PI) as (Flowmaximum -Flowminimum)/Flowmean 2 .
Retinal vessel analysis: The IR fundus images were processed with the Vascular Assessment and Measurement Platform for Images of the REtina (VAMPIRE; Web version, Universities of Edinburgh and Dundee), a validated software application for semi-automatic quantification of retinal vessel properties 4 . We analyzed the central arterial equivalent (CRAE), central venular equivalent (CRVE) and arteriole-to-venule ratio (AVR), representing the widths of the arterioles and venules, and the ratio of vessels widths.
SVD lesions: Brain SVD lesions were scored (ACCJ) and checked by an experienced neuroradiologist (JMW) blind to clinical data, venule scores, venular-PVS overlap, pulsatility and retinal results. We rated periventricular and deep WMH using the Fazekas scale 5 (each score range 0-3, and a combined Fazekas score, 0-6). PVS were scored in the BG and CSO on a validated scale (range 0-4) 3 . Lesions were defined according to the STRIVE criteria 6 .

Inter-observer reliability analysis
To recreate a practical scenario, two observers with different levels of experience and training visually assessed the perivascular spaces (PVS) and the hypointense vessels in two different visual platforms and following similar, but not identical procedures (Table II) in the 12 datasets from the Inflammation in Small Vessel Disease (iSVD) study. The assessments were done blind to each-other's results. To evaluate the validity of the results presented and limits of agreement, inter-observer reliability was analyzed using Bland-Altman plots. There was excellent inter-observer agreement in the number of PVS counted in the anterior ROI section (6 PVS in the left hemisphere (LH) and 4 in the right (RH), mean difference 0, standard deviation (SD) of the difference between observers 2 in LH and 1 in RH). However, Observer 2 generally overestimated the number of PVS in the middle and posterior sections of the ROI in both hemispheres (mean difference per section 2 or 3 PVS, SD 2). The agreement was very good to excellent in the visual counting of suspected venules and almost perfect in reporting the number of PVS and suspected venules that overlapped in all sections (Table III, Figure III).  In summary, despite inter-observer differences in the counts of PVS, the assessment of the overlap between suspected venules and PVS from both observers was almost identical. Tortuosity was higher in eAD and aMCI than controls *Similar method; †similar method; ‡same ROI; §same method; | | Different subjects, not included in sample size; # Voxels of periventricular venules within bilateral ROI with intensity lower than 10% of the mean intensity of surrounding WMH. Periventricular area (mm 2 ); ** Different subjects, not included in sample size; aMCI: