Fig 1.
Automated image processing stages used in subsequent analysis.
Data from two representative sequences obtained from the same subject at 400 fps with 593 nm light. Top shows higher quality data obtained 2° nasal and 0.5° inferior to fixation. Bottom shows lower quality data obtained 6.5° nasal and 2° inferior to fixation. Top left: Average of 80 frames. Top right: Motion contrast image (standard deviation). Bottom left: White shows binary segmentation, red shows skeletonized vessel segments. Bottom right: Labelled vessel segments. Corresponding raw and filtered data sequences are shown in S1 Video (high signal: noise) and S2 Video (low signal: noise). Scale bars show 50 μm.
Fig 2.
Illustration of velocity calculation for a single pixel.
a) Shows the motion contrast image during a 100 ms epoch. The reference pixel whose velocity is to be determined is shown in red. All available image pixels are assayed to determine the best match (white boxes show example candidates); b) shows intensity over time for the reference pixel (red), for the pixel producing the best match after shifting forwards by one frame (blue), and for the other example pixels that were indicated in (a) (black); c) shows the similarity image for the reference pixel (red), which is populated by calculation of root-mean-square (RMS) error between the reference signal and all considered image pixels after shifting by one frame. The brightest pixel (blue) indicates the best match; d) as per (c), but image has been “normalized” by subtracting the standard deviation, which enhances the amplitude of the peak relative to the background; e) as per c), but zoomed according to dashed box shown in (c) for improved visualization; f) as per e), but zoom has been applied to d). The calculation of velocity for this example is overlaid on the image. Scale bars show 50 μm.
Table 1.
Key parameters used in image processing and velocimetry.
Fig 3.
Example spatial maps of flow speed for both PIX and PIV.
Panels correspond to the same sequences shown in Fig 1, and in S1 Video and S2 Video. Raw maps are shown at top in each panel, and interpolated maps are shown in the second row. Maps generated by the PIX algorithm are shown left, with comparison maps from PIV shown right. Scale bars show 50 μm.
Fig 4.
Temporal fidelity in tracking in the face of variability from the cardiac cycle.
Sequence was acquired at 300 fps and can be viewed in S3 Video. Two simultaneously imaged segments were tracked manually (left and right panels). Black crosses show manually tracked data, blue shows the output of the proposed PIX method, red shows output for particle image velocimetry (PIV), and green shows output for the spatiotemporal kymograph (STK). The legend shows, from left to right, the average velocity (AV), the pulsatility index (PI, defined as (max-min) / mean), the R2 goodness of fit to manually tracked data, and the root-mean-square (RMS) residual error to manually tracked data.
Fig 5.
Influence of a white blood cell on capillary velocity.
Top: Velocity trace for one segment showing long-term and short-term fluctuations in velocity captured with the PIX algorithm. Arrows indicate a rapid drop from the presumed systolic peak at ~ 3 mm/sec to ~2 mm/sec, which occurs due to passage of a white blood cell. Bottom: PIX velocity maps corresponding to the time points indicated by arrows. White boxes: the vessel segment plotted whose average velocity is plotted in top. Raw and mean-subtracted sequence can be viewed in S3 Video, and evolution of velocity maps in time can be viewed in S4 Video. White boxes indicate the vessel whose average velocity is plotted top. The legend shows, from left to right, the average velocity (AV), the pulsatility index (PI, defined as (max-min) / mean), the R2 goodness of fit to manually tracked data, and the root-mean-square (RMS) residual error to manually tracked data. Scale bars show 50 μm.
Fig 6.
Spatial velocity maps comparing performance for slow and fast flow.
Top: velocity trace used to infer phase of the cardiac cycle. Left column: presumed diastole. Right column: presumed systole. Sequence was acquired at 300 fps and can be viewed in S5 Video. Top plot confirms the phase of the pulse wave. Second row: PIX (raw). Third row: PIX (filled). Bottom row: PIV (filled). Physiological consistency across the network appears preserved in this sequence during systole with PIX, but not with PIV, which is unable to track the faster systolic flow. The legend shows, from left to right, the average velocity (AV), the pulsatility index (PI, defined as (max-min) / mean), the R2 goodness of fit to manually tracked data, and the root-mean-square (RMS) residual error to manually tracked data. Scale bars show 50 μm.
Fig 7.
Correlation of velocities measured in each vessel segment to the field average.
For each algorithm considered (blue = PIX, green = PIV, red = STK), the average velocity for all visible segments in a field was computed over several cardiac cycles to provide a surrogate for cardiac influence at the capillary level. This procedure was repeated over 7 non-overlapping fields acquired at 300 fps in 2 subjects, yielding a total of 280 unique vessel segments. The goodness of fit (R2) to the field average is plotted for each vessel. The higher R2 values obtained for the PIX algorithm indicate that it generally returned more physiologically plausible outputs under the imaging parameters used here.
Fig 8.
Linear dependence of spatial heterogeneity on mean flow.
Sequences correspond to S3 Video and S5 Video. Capillary transit time for individual segments was calculated by dividing segment length by velocity. Network heterogeneity was quantified by the standard deviation (CTTH) and this was plotted against the mean (CTT) for each 100 ms temporal window (symbols). Systolic and diastolic extrema were identified from a representative velocity trace in each field (circles). A strong linear relationship was evident in both subjects, though marked variability in slope exists.
Fig 9.
Co-localization of high reliability of velocimetry with the vascular tree, but not necessarily with motion contrast.
Video sequence acquired at 200 fps is shown in S6 Video. Top left: average image. Top right: standard deviation image (motion contrast or perfusion image). Bottom left: Peak similarity measure (derived as illustrated in Fig 2) obtained for each pixel, plotted on log scale, without any binary masking of the vascular tree. Bottom right: Velocity map produced from this sequence without any binary masking of the vascular tree. Scale bars show 50 μm.