1. Introduction

Retinal blood flow is an important area of research as it contributes to the understanding of pathophysiology of major sight-threatening diseases such as diabetic retinopathy, glaucoma, and vision loss due to stroke. Blood flow in capillaries in particular is special as they are the major site of exchange for gas and other metabolites and manifest the earliest signs of damage in vascular diseases [1–4].

Although capillaries can be studied in other parts of the body using intravital microscopy and nailfold capillary pressure measurements [5–7], those in the retina have received much attention because of their optical accessibility, and the fact that they are the portion of the microvascular system of the central nervous system that can be studied non-invasively [8]. Currently there is no gold standard to track the flow of cells through retinal capillaries (or indeed, of capillaries in general) rapidly and automatically. Many studies have therefore relied on manual approaches [9–11]. Semi-manual methods based on tracking small subsets of individual erythrocytes have been developed, including use of kymographs formed by the use of repeated 1-dimensional (D) line scans across the flow axis to capture the passage of cells at a retinal location: centroid positions are manually marked, and the velocity is calculated for each cell as a ratio of the distance travelled to the time [11]. The distance travelled by single erythrocytes across time can also be estimated using a temporal offset (of 5 ms) between two separate scanning channels, which are repeated multiple times at each retinal location to obtain estimates of flow. However, the accuracy of velocity measurements using these methods could be affected by precision of locating individual cell centroids [12]. These semi-manual direct methods are also limited by the time taken even to study a small number of vessels.

Increasingly, automated methods have been developed based on the use of spatio-temporal (ST) plots, or kymographs, to visualize and quantify capillary flow in a two-dimensional format. In short, this works by evaluating the movement of cells in a particular direction (in space) over time [13–15]. In the ST plot approach the cells and plasma appear as dark and light bands whose intensity along the centerline of a vessel is plotted over time, noting that the contrast of cells and plasma depends critically on the axial positioning of adaptive optics (AO) focus [16]. ST plots are useful in a number of ways, for example, the distribution of intensities can be used to estimating cell flux, density, and hematocrit, while the slope of the bands represent the flow speeds [11,14,17,18]. However, all these findings can be limited/unreliable when the flow speeds are higher such that imaging speeds are insufficient to capture the quick successive movements of individual cells from one frame to the next. This results in an aliased ST plot where there is an ambiguous pattern of bands making the distinct identification of cells and plasma a difficult task [19,20].

More recent developments have moved beyond traditional ST plot approaches, employing correlation based methods which relate information through time across more areas of the image [16,20,21], or on correcting the aliasing which can corrupt ST plot in the presence of fast flow relative to the sampling rate [19,22].

Given the rise of automated approaches in capillary velocimetry, it is important to validate such measures against manual tracking approaches. If manual tracking proves superior, depending on the goals of a study it may be worthwhile to spend the extra researcher time on this task. On the other hand, if automated tracking proves superior under certain conditions, it is instructive to know when this can be expected so that subjective errors can be avoided. Whichever method is superior should also be considered as the “gold standard” against which new analysis approaches should be compared.

To achieve these goals, here we measured retinal capillary blood flow using both manual cell tracking and automated ST analysis, in the same acquired video data from 3 healthy subjects, across 3 + seconds to capture natural variations across the cardiac cycle.

2. Methods

The study methods, participants, and the imaging procedures are similar to our previous studies [23,24]. The study was approved by the institutional Human Research Ethics Committee (ethics approval number 1137234) and adhered to the tenets of the Declaration of Helsinki. All participants had signed a written informed consent before the start of the study.

The study participants included 3 healthy individuals (1 male and 2 females) aged 22-23 years. They were free from any known ocular or systemic conditions and had a clear media and a refractive error of less than 4 diopters spherical and 2 diopters astigmatism. Retinal vessel imaging was performed in the left eye of all participants using optical coherence tomography angiography (Spectralis, Heidelberg system) and AO (custom “flood” illuminated system). Eyes were dilated with 0.5% tropicamide (Alcon, USA) 20 minutes preceding AO imaging [23,24].

2.1 AO imaging

The study subjects were imaged using “flood” AO, the imaging setup and the procedure has been detailed in previous studies [21,25–27] and briefed here.

In order to track the motion of erythrocytes in the smallest capillary vessels, we used image acquisition rates of about 200-300 frames per second (fps). The image capturing was achieved using a 2560 × 2160-pixel Andor NEO 5.5 sCMOS camera (Andor Technology PLC, Belfast, UK). The camera’s global shutter mode was used to expose all the pixels simultaneously, this resulted in images free of intra-frame distortion that might arise due to involuntary eye movements. Wavefront aberrations arising from the ocular structures were measured using a Hartmann-Shack wavefront sensor (Adaptive optics associates, Cambridge, MA) and the errors in the measured wavefront slopes were corrected by a 97-channel deformable mirror (Alpao, France) using custom MATLAB software. Image recording begins when root-mean-square wavefront error measured over a 7 mm pupil is in the range 0.04 to 0.08 µm (cf. Marechal criterion for diffraction-limited imaging of approx. 0.05 µm at 750 nm). This implies a point-spread function close to the diffraction limit at this wavelength. The best focus criterion was set to maximize the number of capillaries visible in a retinal field, and that single file flow could be subjectively resolved such that all cells appeared trackable. This was achieved by altering the adaptive optics defocus component in 0.05D steps [23,24].

2.1.1 Imaging procedure

The foveal microvascular network of all the study participants was imaged in areas of 1.25° diameters in a darkened laboratory using a fixation target that was diffusely back-illuminated with a fixation grid printed in black on white. The participant was guided to fixate at specific locations on the target in order to bring a retinal field into view. An area ranging 1° to 2.5° from fixation was imaged around the foveal avascular zone (FAZ). We used a supercontinuum laser (a 750-nm light) passed through a tunable transmission filter (Superchrome SC480-8, Fianium, Southampton, UK) that resulted in a 0.36 mW power at the cornea. The imaging protocol adhered to the ANSI guidelines of maximum permissible exposure standards for continuous illumination over 3.4 sec [28]. From these recorded movies, cellular velocities were measured using two distinct methods (manual and automated) as described further.

2.1.2 Image processing

Using the AO ophthalmoscope described above, blood flow imaged over a duration of 3 seconds is acquired in a proprietary movie file format. A series of image processing steps are involved to extract the individual frames and the averaged images for each of the sequences using a custom software written in MATLAB (Mathworks, Natick, MA). Movies were flat-fielded, registered, and mean subtracted before creating a motion-contrast (or a “division”) image [16,29].

The manual tracking using raw movie data could be limited by factors such as poor contrast, and ambiguous identification of cells due to unwanted reflections from the underlying photoreceptors. In order to overcome the difficulties arising due to poor contrast, all movies were enhanced by subtracting the mean intensity over a rolling window of 200 ms to minimize structural information. This significantly improved contrast and visualization of individual erythrocytes [20,21].

2.2 Manual method of velocity estimations

Each imaging run resulted in a video file containing a sequence of 600-1000 individual frames. Consecutive frames were scanned through to manually identify and track the movement of individual erythrocytes (red blood cells) using custom software written algorithm in MATLAB. The process of selecting vessels for manual tracking, the process of red blood cell identification and tracking, and the analysis of the resulting cell velocity data are detailed below:

2.2.1 Vessel selection & cell identification

In order to study the flow of erythrocytes using the manual and automated methods, “vessels of interest” (VOIs) were carefully selected, which were capillaries defined by the presence of single file flow of erythrocytes separated clearly by plasma gaps [21]. Vessels containing cell aggregates and/or displaying more rapid flow were included as long as flow could be unambiguously tracked. Our AO imaging setup ensures that we identify vessels including capillaries (and the flowing erythrocytes) with good resolution (2 µm). We selected all vessels in the imaged field for which single-file cellular flow could be resolved (that is, that the positions of individual cells could be determined on single image frames which is a requirement for manual tracking) and it requires that the vessel contrast was good. The single file flow of corpuscles in a vessel were ensured by observing the flow of all capillaries within approximately 2° of the foveal centre (i.e. close to the foveal avascular zone). To identify red cells we assessed the size, shape, and separation between components of the blood column. Erythrocytes were identified as cells smaller in size with well-defined edges. Leukocytes, on the other hand, were identified as cells relatively bigger in size and that appeared only intermittently in a vessel. No vessels demonstrating leukocyte flow were included in the analysis. A total of 13 VOIs were selected across 3 subjects for detailed study. Whilst this is a relatively small sample size in terms of vessels, it represents many cells in labelling as described further below.

2.2.2 Red blood cell tracking

The paths taken by erythrocytes though capillaries were tracked by manual labelling of individual cells along the length of a movie for each VOI. There were 967 erythrocyte cells tracked across 13 VOI. Each cell was tracked for an average of 10 consecutive frames. In this way at least 60-80 cells (for about 600-1000 frames) for each vessel were identified and tracked. The total number of cell labels across all the frames made were about 8,600. The average time required to track all erythrocyte paths for a single vessel was about 5 hours. So, manual tracking in this way totaled approximately 65 hours (for 13 vessels).

The manual labelling of individual cells was overseen by a single grader (the first author). For 6 of the vessels, manual tracking information was seeded by manual labels made by optometry students undertaking a research project in the laboratory. The first author reviewed all grades and corrected them on a frame-by-frame basis in the case of any clear error.

Cell labels were made with help of a user interface created in MATLAB that had features to look at two successive frames in quick succession. In this way the “tip” (start) and “tail” (end) of each cell passing through the capillary could be identified and marked. The start and the end of the cells were confirmed as per the flow direction visualized in the movie. Each cell was given a unique label to identify it as it was tracked across several frames [ Fig. 1]. For example, Fig. 1 (A and B) represent examples of a single frame where the cell tracking was done on three different vessels labelled “a”, “c” and “d” using raw data and mean subtracted data, with each tracked cell assigned a number indicating its order in the sequence of cells passing through the vessel. The visualization of cells was easier using the mean subtracted as compared to raw movie data.

 

Fig. 1. Identification of individual erythrocytes using raw and mean subtracted data. A:D Examples of cell labelling and tracking on consecutive frames (DIV images from a field 0.5°inferior and 1.5°temporal to foveal avascular zone) in subject 2. A-B raw data, C-D mean subtracted data. Green indicates the tip of a cell and its tail in red. The alphabetical IDs in the figure (such as a, c and d) represent a unique vessel, and the numbers represent the sequence of cells within that vessel. For example, for a vessel “d” there are 3 cells tracked in frame A and 4 cells tracked in frame B. Arrowhead indicates blood flow direction. Side panel displays the erythrocytes (indicated by an arrow). Scale bar is 20 µm.

Download Full Size | PDF

Cells were tracked along vessel lengths that were uninterrupted by crossings or junctions. A short region of the vessel segment where the visibility of cells was better appreciated was chosen to manually track the cells, a shorter region also allowed us to reliably track the cells within the time constraints of this study unlike the large regions such as branch to branch vessel length which would need larger amounts of time and human skill.

A total of 600 frames were tracked for movies imaged at 200 frames/sec (with 3 seconds of data acquisition) and 1020 frames at 300 frames/sec (3.4 seconds data acquisition).

2.2.3 Measuring erythrocyte velocity

To measure cell velocity, the centroid of each cell was calculated as the average of the tip and tail points. The shift in the cell centroid between consecutive frames was then calculated. Cell velocities were then plotted as a function of time (which were referred to as “velocity plots”) [18]. It was observed that individual velocity estimates from a single frame pair are typically noisy [ Fig. 2 (B)]; hence the measures corresponding to each individual cell were averaged to produce a single robust estimate of that cell’s velocity [Fig. 2 (C)]. This precludes the analysis of acceleration for a given cell in this work (e.g. due to the systolic transition) [18,22], but was necessary to create more robust manual velocimetry measures.

 

Fig. 2. Erythrocyte flow pulsatility in a single vessel. A – Adaptive optics image highlighting the vessel segment with tracked cells. B– individual cell velocity estimates and C - average erythrocyte cell velocities plotted across 3 seconds data (600 frames), circles indicate mean and error bars indicate standard error of the mean.

Download Full Size | PDF

2.2.4 Reliability checks for manual tracking

All manual data were re-checked by the first author of this study for the accuracy of tracking, any errors noted in cell labelling / cell markings were corrected. Errors in cell tracking included misidentification due to poor contrast which could result in large standard error of the mean (SEM) at a given time point, for example where the grader had mistakenly used the wrong identifier for a cell. This results in incorrect velocity, inflating the SEM. In cases where the cell in question could not be reliably labelled in consecutive frames, it resulted in missing data. This was seen where the grader was not sure of the cell paths in certain frames due to poor contrast and/or possible flow reversal. False identifications of erythrocytes were occasionally encountered where the apparently labelled points on a vessel were clearly not the actual erythrocytes. This could happen due to unwanted reflections from the cone photoreceptors that confused graders while labelling cells. These markings were also seen at the vessel border and outside the vessel wall. Such points were manually removed from the data set on re-examining the cell tracking. For the data oversight, all the frames across all vessels were rechecked for above specified errors, the cells were retracked where needed. The data were then re-plotted after all possible manual tracking errors have been rectified.

2.3 Automated ST plot approach

2.3.1 Extraction of “vessel traces” for velocity measurements

Unlike the manual method, where the raw images were used to track cell flow, the automated method began by using a division (DIV) image to first define cell paths. The division image specific to each retinal field was used to mark out the single line of pixels that best represented the centerline of each VOI – the resultant skeleton image is described as a “vessel trace” This was performed using the image processing software package Adobe Photoshop (CS2, version 9.0). The trace drawn for the ST plot was similar in length to the extent of manual cell tracking done on a vessel. The length of the trace for different vessels ranged from 15.4 µm to 102.9 µm. The trace length drawn was dependent on factors such as the location of a vessel between branch points and on their velocity relative to the length of the vessel since rapidly moving cells do not remain in the same vessel for long. In order to compare the same video data between manual and automated methods we traced a length of the vessel that corresponded to the areas that have been marked with manual grading.

Drawing a trace can be limited by factors such as poor vessel contrast, a common limitation noticed is seen in Fig. 3(A) (indicated by arrow) where one can appreciate the very common multi-peaked vessel cross-section (i.e. a dark stripe down the middle). This phenomenon is not fully understood, however in such cases the brightest of the two “lobes” was taken to maximize the ST plot contrast.

 

Fig. 3. Process of extracting vessel traces. A: Example of a DIV image with traces overlaid, arrow indicates presence of two vessel peaks. B: extracted skeletonized vessel traces from the DIV image.

Download Full Size | PDF

2.3.2 ST plot approach and velocity estimation

The pixel co-ordinates from the vessel trace were used to create a spatio-temporal plot using Matlab. A spatio-temporal plot indicates corpuscular flow for a given vessel segment by depicting the evolution of pixel intensity patterns over time.

In addition to analysis of the slope of the ST plot itself, 2D spatial and temporal correlograms were employed as described in an earlier study [16], to improve robustness to errors; the representation giving the highest signal-to-noise ratio was chosen.

A 100 ms epoch was used to sample the data at 300 fps and a 150 ms epoch at 200 fps, corresponding to 30 contiguous frames in either case, with each window separated from one another by 15 frames. The selection of the epoch (100-150 ms) may seem somewhat arbitrary. It's essential to have an adequate number of frames for a precise velocity estimate. However, using too many frames can hinder the ability to track sudden changes in velocity. To strike a balance, an epoch of 100 ms (300 fps) and 150 ms (200 fps) was chosen, serving as a compromise between these considerations. This choice is comparable to epochs selected in earlier studies [11,18,20,23]. It would be possible to improve the temporal sampling of flow velocity by employing analysis epochs separated in time by less than half the window size (e.g. separated by one frame). However, this is computationally expensive and diminishing returns are expected given the low-pass nature of using an extended window; the higher temporal fidelity that will result would be difficult to distinguish from noise.

Plotting up the velocities from individual windows as a function of time resulted in a “Velocity-time plot” for each vessel segment [ Fig. 4 (C)]. The slope gives velocity as distance over time. The method of finding the slope has been described earlier by Bedggood and Metha 2014 [16]. The approach is based on the radon projection, i.e. the ST plot image is averaged in different angular directions to determine the best overall angle (when the direction being averaged along matches the direction of the bands in the ST plot, the resulting projection will have maximal variance). Each vessel had an average of 45 epochs where the velocities were estimated using the ST plot analysis.

 

Fig. 4. Overview of extraction of velocity from ST plot. A: Pixel intensity patterns along the central axis of one vessel segment (horizontal direction) evolve over time (vertical direction), indicating single-file flow of corpuscles separated by plasma gaps. B: A 100 ms (30-frames) sample window (highlighted in A) provides an estimate of average velocity (0.58 mm/sec) within this epoch given by the blue line slope. Scale bar vertical is 50 ms and horizonal is 20 microns. C: Velocity-time plot for the complete spatial and frame range; data point (indicated by an arrow) corresponds to the velocity described in B.

Download Full Size | PDF

2.3.3 Identifying potential errors in velocity data

The velocity plots were evaluated for the presence of data points that appear to deviate from the expected smoothly pulsatile pattern. Such points (within individual windows) were verified for the reliability of the velocities reported (by examining the movies and ST plots) and were excluded from the analysis if they met the below mentioned criteria.

Low velocity data points: These are the data points that have a velocity reported less than a minimum velocity criterion of 0.2 mm/sec. It is not expected to observe flow slower than this based on published velocity estimates for the retinal capillaries [10–12,18,21]. Whilst the low velocities can be due to the presence of nearby leukocytes, cell aggregates, or general nearby traffic [20], deviations needed to be checked to see if they indeed arise from valid data.

In case of aliasing, where the true movement is mis-represented due to the regular size and separation of cells and their velocity compared to the frame rate, the algorithm tends to report the velocity as near zero, this was confirmed by checking the sign reversal on flow. An example is shown in Fig. 5 where the flow is reaching a reversal state from positive flow sign. This suggests the possibility of flow reaching a reversal state (i.e., aliasing). These examples were confirmed by observing the movies where the flow can appear very fast and in the opposite direction.

 

Fig. 5. Automated velocity plot showing example of sign inversion. Evidence of aliasing shown as reversals of data points from arbitrary positive to negative flow signs. Flow velocity variability as a function of time is shown here. Circles indicate raw data, and the lines indicate interpolated data.

Download Full Size | PDF

High velocity data points: The high velocity points can be seen in aliased flow where the proper identification of cells is hindered due to under sampling of imaging rates [18,19]. As per Gu et al [18], using 400 fps a maximum velocity between 2 and 3 mm/sec can be estimated with less effects of aliasing as compared to a 200 fps imaging. For example, the possibility of estimating reliable velocities (i.e., > 3 and 3.5 mm/sec) without effects of aliasing are limited for the imaging frame rates (200 and 300 fps respectively) considered in this study. Due to misidentification of cells, there can be a deviation from the expected velocity patterns and hence such points (individual epochs) were excluded from the analysis.

We investigated each epoch removed to confirm as best we could (from manual inspection of the video and from manual inspection of the automated plots) that the flow was legitimately pushing the bounds of being too slow or too fast as described above and hence needed to be discarded (as opposed to being an error in either manual grades or the ST plot).

2.4 Data interpolation

When cell velocities are estimated using the manual tracking method, each cell was labelled over a series of frames. As a result, the instantaneous velocities for each cell have been plotted as a function of time with data points separated by the inter-frame interval, which was 3.3 ms at 300 fps and 5 ms at 200 fps. However, when using the automated approach, the velocities are extracted from each time window, or ‘epoch’, by averaging the estimates from all cells that were visible in the epoch. These epochs are separated by 75 ms (this is for data collected at 200 frames per second, where epochs were of length 150 ms and each one is spaced half the epoch duration apart). A common scale is required so that the velocities from both the methods can be compared and correlated. This is done by data interpolation where both the manual and automated measures have been scaled to reach 5 ms intervals between consecutive data points. After data interpolation, about 7,935 velocity points were available from 13 vessels, for the purpose of correlating the interpolated data between methods, to put the plotted points on a less redundant time scale, every 25^th data point was taken from each vessel, this resulted in a data set of 321 points each for manual and automated methods.

2.5 Statistical analysis

The statistical analysis of the data was performed using IBM SPSS statistics for windows, version 21.0 (IBM Corp., Armonk, N.Y., USA). Pearson’s correlation was used to examine the strength of correlations between variables at a significance level of 0.05. Bland-Altman analysis (using GraphPad Prism version 9.0) was performed to assess the agreement between the automated and manual methods of velocity extractions. Velocity plots for different methods were generated using MATLAB software.

3. Results

This results section presents the average velocities (± standard deviation) obtained for a total of 13 capillary vessel segments from three subjects using both automated and manual methods. This represents the full set of manually tracked data, including 967 manually tracked unique cells and 561 automated analyzed epochs in total. Average velocities from each VOI are reported for the raw data whereas individual velocities from many epochs in each vessel are reported for the interpolated data.

3.1 Comparison of manual and automated velocities

Cell velocities from both manual and automated methods were plotted against time for both raw data and interpolated data.

Raw data: Average velocity was taken from n = 13 vessel segments using both the methods, the average ± standard deviation velocity using the manual tracking method was 1.07 ± 0.37 mm/sec, whereas for the automated method it was 1.10 ± 0.42 mm/sec. Standard error of the mean (calculated as standard deviation/$\sqrt {13} $) for the manual method and the automated method was 0.10 and 0.12 respectively. There was a very high positive correlation noted between the two methods (r = 0.93, p = 0.000, Pearson’s correlation) except for a vessel segment [ Fig. 6 (A)].

 

Fig. 6. Depicting correlations between manual and automated methods. A: Correlation of average raw velocities between manual and automated methods. B: Correlation of interpolated velocities between methods. Blue dashed line indicates line of best fit. C: A Bland-Altman plot representing agreement between Manual and Automated methods. Solid line indicates mean and dashed lines indicates upper and lower bounds of the plotted data.

Download Full Size | PDF

Interpolated data: A total of n = 321 velocity estimates (from n = 13 vessels) were taken from the interpolated data using both methods. A very high positive correlation between methods was observed (r = 0.92, p = 0.000, Pearson’s correlation) [Fig. 6 (B)]. When the correlations were run across individual vessels on the interpolated data, ‘r’ values from 13 vessels ranged: 0.57 to 0.98 (p < 0.001, Pearson’s correlation) with an average ‘r’ of 0.84 (95% CI: 0.76 to 0.91). This is comparable to the overall ‘r’ of 0.92 as mentioned above.

A Bland-Altman analysis (using GraphPad Prism version 9.0) has been performed to assess the agreement between manual and automated methods of erythrocyte velocity tracking. The results show a bias of -0.03 (SD 0.14) and 95% limits of agreement ranging from -0.31 to 0.25. Although a good agreement has been found between the methods for average velocities for most of the data, a significant discrepancy between methods can be observed by a single data point in Fig. 6 (C). The reason for this is explained ahead [ Fig. 8 (A), (A1) and (A3)].

Both raw and interpolated data show a very high correlation between methods. The validity of the automated method was further confirmed by comparing the sequence of data obtained over multiple cardiac cycles with that of the manual method. Out of 13 vessel segments, good agreement of velocity estimates was seen in 9 vessels, out of which a few examples are shown in Fig. 7 which indicate a good manual tracking with small standard error bars, and smooth waveforms that appear to be pulsatile. For these segments, agreement for velocity calculated using manual and automated methods is high, with differences in average velocities being less than 0.1 mm/sec.

 

Fig. 7. Velocity plots depicting high correlation between methods. A:D – Velocity estimates plotted using raw data (in blue – manually tracked data) and automated data (in green), x - mean velocity, error bars indicate standard error of the mean velocity of each manually tracked cell. A1:D1 – Corresponding interpolated data (in blue – manual data, in pink – automated data). A median filter was used to smoothen out the manual data.

Download Full Size | PDF

The other 4 vessels showed poor concordance, a few examples are shown in Fig. 8. The variation in the strength of relationship seemed to depend upon the quality of the manual raw measures; vessels with good manual tracking as demonstrated by tight error bars seemed to be well correlated with the automated measures and vice versa.

Two examples are displayed in Fig. 8 where the velocity estimates generated by the two methods do not seem to agree well. The differences in calculated average cell velocities between methods for these vessels were 0.45 mm/sec [Fig. 8 (A)] and 0.15 mm/sec [Fig. 8 (B)] respectively. It is observed that the data derived from manual tracking in these plots is of concern due to the high variability seen as wider error bars. Indeed, Fig. 8 (A) shows an example where manual tracking of cells through this vessel segment was made difficult for graders as the flow was faster compared to other vessels. This vessel was imaged using the slower frame rate of 200 frames per second, resulting in some epochs of aliasing as clearly evidenced on the ST-plot [Fig. 8 (A2)]. Despite the availability of any other redundant cues, these aliased sections make it very difficult for human graders to disambiguate real from aliased motion and leads to misidentification of cells. This results in variability seen as data points straying from the expected pattern and increased standard error of the mean [Fig. 8 (A)]. However, these limitations can be overcome using the automated approach by additionally supplying the algorithm with the best estimate of direction of flow. This enabled a means to avoid aliasing artifacts and extended correct determination of cell velocities up to 3.3 mm/sec [Fig. 8 (A)]. This was confirmed on the interpolated data too, which [Fig. 8 (A1)] appears comparable to the flow pattern noted in other vessels.

 

Fig. 8. Plots depicting poor correlation between methods. A:B: Velocity estimates plotted using raw data (in blue – manually tracked data) and automated data (in green), x - mean velocity, error bars indicate standard error of the mean velocity of each manually tracked cell. A1:B1 – Interpolated data (in blue – manual data, in pink – automated data). A2, B2: Showing ST plot for A and B, respectively. Aliased epochs where clear direction of flow could not be appreciated are seen more in A2 than B2. A3, B3: Inspecting grades for A and B: vessel trace (in red) used for ST plot overlaid on the centerline of the vessel, in green are the manual data points.

Download Full Size | PDF

Another example [Fig. 8 (panel B)] shows that the manual and automated ST plot measures appear slightly off phase with each other. The possible reason for this is that some of the manually tracked data appears out of the vessel wall (see comparatively wide “cloud” of points plotted in Fig. 8 (B3). This may occur due to real optical effects such as scatter [30], due to difficulty tracking in low contrast raw data, or due to some of the aliased sections seen on ST plot [Fig. 8 (B2)]. This might have resulted in increased standard error bars evident on Fig. 8 (B), whereas automated data showed no artifacts. These two examples show that in some vessels the automatic method can be more robust than the manual “gold standard”. Importantly, we did not encounter any examples where the automated algorithm failed but the manual approach had succeeded (despite having selected vessels subjectively, which should tend to impose bias towards manual tracking).

4. Discussion

This study discusses the challenges and some potential solutions to the problem of validating capillary flow information captured using two standard methods: manual tracking and automated approaches. Use of manual tracking to unravel the flow characteristics at the cellular level requires analysis of many thousands of cells travelling through hundreds of vessels, which is difficult to implement in a time-bound study with a large sample. This study shows how we went about verifying the automatic method against manual measures so that we could use this to perform the large number of analyses needed to discover flow characteristics across a small piece of the microvascular network.

4.1 Average velocities using reported methods

The average red blood cell velocities reported here using the manual method are 1.07 ± 0.37 mm/sec (ranging 0.54 to 1.78 mm/sec). Previously using SLO imaging (with fluorescence) and digital frame-to-frame to cell tracking, the average velocities reported in human retinal capillaries were as high as 3.28 ± 0.45 mm/sec [9] and using AOSLO imaging in mice and semi-manual cell tracking in the capillary vessels showed a range of average velocity as low as 0.45-0.99 mm/sec [11]. The range of velocities reported here are smaller as compared to the previous stated studies in humans. These are values from subset of vessels that were relatively easy to track as compared to vessels where the flow was slippery and fast where single cell passage was difficult to discern. Due to the sampling effect which has inclusion of limited vessels the reported average velocities are on the low end (i.e., fastest average velocity was ∼1.8 mm/sec). However, our range of velocities across all the vessels i.e., the minimum and maximum using both manual and automated methods ranged 0.2 and 3.5 mm/sec and are similar to the previously reported values using other automated measures such as PIV [21].

The utility of ST plots in estimating the cell velocities from larger vessels (of diameters ranging 28-100 µm) has been shown previously [13,14,31,32]. Using the automated approach in our subjects, the mean velocities were similar to the manual method (average ± SD: Automated - 1.10 ± 0.42, average velocity range: 0.48 to 1.79 mm/sec, Manual – 1.07 ± 0.37, ranges: 0.54 to 1.78 mm/sec). These agree to the average erythrocyte velocity reported using the PIV method which was 1.3 mm/sec (ranged 0.3 to 3.3 mm/sec) [21] and the average velocities (ranged 1.19 to 2.37 mm/sec) reported using the PIX methods [20].

4.2 Efficiencies of manual vs automated methods

The raw number of cells tracked manually in this study were about 8,600 (cells) which is lower but comparable to the number of cells (13,000) tracked in an earlier study where manual tracking was done to analyze cell velocities and hematocrit in capillary vessels [11]. Having manually tracked data was important, to establish a sense of expected pattern of flow with which we could rely on the automated measures.

Compared to the automated method, the manual tracking involved a higher amount of time (nearly 5x) and grader’s skill in precise markings of the tip and tail of the erythrocytes. We have observed the accuracy of manually tracked data in a few vessels was limited by reasons such as poor contrast on raw data, background noise due to cone photoreceptors that impeded the exact markings of erythrocytes (despite the vessels having been chosen due to the anticipated ease of tracking cellular flow through them). Although the flow estimation using manual tracking may be considered by many as a “gold standard”, faster flow particularly using lower frame rates may give rise to aliased appearance of cells [19]. Automated methods are also vulnerable to this effect but may be more robust.

As shown in results the velocity estimates with manual tracking had marked variability relative to the automated data due to difficult tracking in aliased appearances of the cells with increased flow speeds. The automated method was still able to give reliable velocity estimates that appear to match the velocity patterns noted in the other network vessels, suggesting that automated approaches may be preferable to manual tracking approaches regarding data fidelity, in addition to the large savings in researcher time.

Using these methods (both manual and automated) reported here, estimating higher speeds (> 4 mm/sec) may be limited by the aliasing errors seen with increased flow speeds. More recent approaches from our laboratory such as pixel intensity cross-correlation and motion correction of kymographs offer alternative approaches to return valid estimates in some vessels even when flow is aliased both on manual inspection and by traditional automated approaches [19,20]. However, these methods are not yet in common use and were also not available at commencement and at the execution of research methodology of this study.

Using both the manual and automated methods, pulsatile pattern of capillary flow was confirmed as described by several other authors [11,12,17,18,20,33]. This flow characteristic was very important to explore flow variability among many vessels in a network as detailed in our previous work [23].

4.3 Conclusions

This study aimed to validate the utility of automated spatio-temporal analysis in comparison to the manual tracking of cells. Although the former is an established method in extracting the blood flow velocities, its reliability in extreme/challenging scenarios such as aliasing is not known. This article presents the reliability of ST analysis in measuring flow velocities even in flow-reversals which is otherwise difficult using a manual tracking method (as shown in Fig. 8).

The automated method described here is comparatively much faster, and is comparable to if not superior in accuracy, to the manual way of tracking erythrocyte velocities. The spatio-temporal analysis described here will therefore assist us in quantifying the flow in many segments across the microvascular network.