Oxygen exchange occurs within capillaries of the order of 5–10 μm in diameter [1]. Within neural tissue, capillary diameter is typically closer to the narrower end of this range. Given a cell-free layer 0.5–1.0 μm thick [2], the available channel for cellular passage may be as small as 3 μm. As this is much narrower than the free diameter of a red blood cell at ∼8 μm, a significant change in cell shape is required to squeeze through capillary networks. Current understanding of this process is based largely on computer modeling and on in vitro preparations in which blood cells are propelled through synthetic channels [3]. Assessing cellular transit in humans, especially within neural tissue, is of particular interest due to the low diameter of these capillaries and high metabolic activity of the tissue.

The required investigations can now be achieved in the living human retina, using adaptive optics to permit noninvasive cellular resolution in vivo. This allows observation of the blood column in its natural biological milieu and without washing, spinning, or labeling of cells which may impart unknown physiological change.

Prior developments have enabled visualization of capillary flow channels [4] and quantification of flow velocity [5–13] and flux (cell counts) [14–16]. The blood column has also been visualized by freezing a confocal scanning raster to sample a vascular cross section at high speed [13]; the steady passage of the blood column allows it to be rendered in high detail via confocal sectioning at kilohertz sampling rates [10,16]. However, this method creates ambiguity between spatial and temporal information, and typically assays only a single vessel. Accordingly, a quantitative description of red blood cell shape in the living retina has not yet been given.

In this Letter, we describe a new method to measure the shape of red blood cells as they traverse narrow capillary tubes. High-speed flood-illuminated adaptive optics was used to acquire noninvasive video data of single-file cellular flow in the retinal capillary network. A new analysis approach was developed to render the average shape of red blood cells within capillary tubes of varying size, permitting quantification of cell shape and orientation.

High-resolution image data were collected from three healthy human subjects aged 23, 24, and 36 years and with no ocular pathology. Fixation was directed on a calibrated grid to locations 1.1° to 2.2° from the foveal center. At each location, a flood-illuminated adaptive optics ophthalmoscope imaged a 1.25° diameter field. Wavefront information was measured at 835 nm with a superluminescent diode (Hamamatsu, Japan) and a custom Hartmann–Shack aberrometer. Aberrations were corrected with a deformable mirror (HiSpeed DM97-15, Alpao, France) at 20 Hz. When root-mean-square wavefront error reached ∼0.06 µm or better over a 7.0 mm pupil, a 3 s video was acquired with a scientific CMOS camera (NEO, Andor Technology, Belfast, UK) at 300–400 fps, with pixels corresponding to 0.5 µm on the retina for an emmetropic eye. The light source was a supercontinuum laser passed through a tunable transmission filter (SC480-8 and Superchrome, Fianium, Southampton, UK) to provide an imaging band at 750 ± 25 nm. Light power at the cornea was 0.3 mW to 1.3 mW, ≥10 times below ANSI maximum exposure [17].

Raw video data were background-subtracted and flat-fielded to correct for stray light and nonuniformities in illumination. Each frame was corrected for ocular motion by translation with cross correlation. Ocular torsion was not corrected, but generally negligible within sequences [5,18]. For each registered video, motion contrast images [Fig. 1(A)] were generated from the variance in pixel intensity [4]. A skeletonized vascular network was then traced manually in Photoshop (Adobe, USA). Vessel centerlines were used to construct a kymograph [Fig. 1(B)] from the video data for each vessel, as commonly described [6,7,9,10,12,13,16,19]. Here, contrast has been adjusted so that bright bands correspond to cells and dark bands to plasma (as described below, the sign of contrast can be reversed depending on the plane of focus; we determined cellular identity for each vessel by manual inspection of video sequences, and excluded ambiguous vessels).

 

Fig. 1. Image processing pipeline. (A) Motion contrast image showing a vessel of interest highlighted in green. (B) Kymograph showing the variation in intensity over time along the vessel centerline. Cells appear bright, separated by dark plasma gaps and inclined at an angle commensurate with their flow velocity. Scale matches that shown in (C). (C) Motion-corrected version of (B), with cells now appearing stationary in time (i.e., as vertical bars). Spatial scale shown also applies to (D) and (E). (D) Blood column image, showing average intensity across time for paths parallel to the vessel centerline. Data from (C) correspond to the central row here. (E) As for (D), with thresholded pixels overlaid in red and cell centroids in green.

Download Full Size | PDF

Kymographs were corrected for motion of the blood column by rigid translation of each frame in the flow direction [19]. This presumes that the blood column, as a fluid, is incompressible. The translation was optimized by trial-and-error to maximize contrast of the resulting blood column image, seeded by an initial velocity estimate from pixel intensity cross correlation [7]. This procedure was repeated in 100 ms windows across 3 s of data, with interpolation used to estimate the correction at each frame. The output is shown in Fig. 1(C), which is a row-shifted version of Fig. 1(B).

Assessment of cell shape requires kymograph data to be generated across the full blood column, not just the centerline. To achieve this we fit a 2D spline to the centerline, permitting computation of an orthogonal intensity cross section at each point along the vessel. The cross section spanned 25 pixels, or approx. 12.5 µm (±6.25 μm). This allowed us to render the blood column and surrounds in a 3D data array (with one time dimension, and two spatial dimensions parallel and perpendicular to flow). In other words, the data formed a stack of kymographs, each parallel to the flow axis. The motion correction described above for the central kymograph [converting Fig. 1(B) to Fig. 1(C)] was applied to all kymographs, i.e., to all paths parallel to the flow axis. This assumes constant speed across the width of the blood column. Of course, it is known that the boundary plasma layer travels at different velocity [3]. Our method relies on contrast from the alternating passage of cells and plasma and so is blind to this phenomenon; however, this is not required to measure cell shape. Similarly, our method is blind to tank-treading of the cell membrane [3], and cannot delineate cells within aggregates (vessels containing more than sporadic aggregates were excluded as described below).

The above procedure produced a stack of motion-corrected kymographs, each parallel to the flow axis. The resulting 3D data array was averaged in the time dimension to produce a spatial image of the blood column [Fig. 1(D)]. Each bright band in Fig. 1(C) is aligned with a cell in Fig. 1(D); the central row of the latter is equal to the time average of the former. This procedure was applied across 3 s of data, producing very long and thin images; a selected 500 ms period is shown here for illustrative purposes and data are wrapped for visualization over longer periods in the subsequent figures.

Blood column images were automatically thresholded [20], demarcating cells from plasma [Fig. 1(E), red]. To robustly identify cell centroids we first used principal component analysis to categorize each position along the blood column as either “cell” or “plasma.” Contiguous blocks of “cell” at least 2 μm long were collated to create a proto-image of the average cell [Fig. 2(A)]. This image was used as a rolling 2D correlation window to compute a similarity score for each point along the blood column. This gave a convex representation of cell position, with each local maximum taken as the centroid of a cell [Fig. 1(E), crosses]. Cell centroids were used to rearrange the binary data [Fig. 1(E), red] in a 3D stack, with each cell a different slice. The stack mean was computed to map the probability that a pixel was part of a cell, as a function of distance from its centroid [Fig. 2(B)]. Pixels with probability >50% were taken to comprise part of the “average cell,” whose profile was smoothed for shape analysis [Fig. 2(C)]. A mean ± standard deviation of 151 ± 112 cells contributed to each average cell.

 

Fig. 2. Example representation of the “average cell” in a single capillary. (A) Average intensity as a function of distance from the cell centroid located at the image center. (B) Probability map, showing the likelihood of observing a cell as a function of distance from its centroid. (C) Binary cell profile, taken as a smooth contour of 50% probability from (B).

Download Full Size | PDF

Two key features of the average cell image [Fig. 2(A)] are noteworthy: reflectance profiles appear mirrored between cells and plasma (i.e., the profile appears the same but with reversed sign), and there was a multimodal cross section whereby intensity variations are apparent well beyond the likely bounds of the vessel lumen. Combined, these features explain the often striped appearance of small blood vessels in adaptive optics images. This can be better understood by considering the average intensity profile perpendicular to the flow direction for the same example vessel (Fig. 3). Where the blood column contains a cell (red), a multilobed pattern is noted which extends some distance beyond the vessel lumen (solid black line). An equivalent but mirrored pattern is evident for plasma (blue). This “mirroring” of the intensity profile is consistent with the notion that cellular contrast is analogous to that in defocusing phase contrast microscopy, which depends upon the axial displacement from the plane of focus, the refractive index difference compared to surrounding tissue, and the curvature of the cell [6,21]. As the axial position and vessel curvature should not vary between cells and plasma, the sign of contrast is dictated by refractive index differences, with the refractive index of red blood cells and plasma straddling that of the retinal tissue to produce the observed polarity in contrast [6,21].

 

Fig. 3. Capillary intensity profile perpendicular to flow. The intensity profile for cells (red) mirrors that for plasma (blue), with a multimodal pattern extending well beyond the vessel lumen (black, solid). The alternating passage of cells and plasma creates variance over time (magenta) which explains the striped appearance of capillaries in motion contrast images. The flow axis is indicated (black, dotted).

Download Full Size | PDF

The multimodal intensity distribution of Fig. 3 could result from scatter (e.g., by hemoglobin); however, it is clear that plasma has a similar reach outside the vessel (e.g., in Fig. 3, first maximum of the blue curve coincides with first minimum of the red curve). A more likely explanation is optical interference, as the thicknesses of capillary wall structures such as the basement membrane [22] and the plasma boundary layer [2] are of the order of 0.5–1.0 µm and hence comparable to the wavelength of our imaging light. For consistency, we segmented the bright part of the average cell [e.g., white pixels in Fig. 2(A)] as this corresponds to the major source of contrast and hence the most likely bounds of the lumen.

To assess cell shape, an ellipse was fit to the cell outline [e.g., to the binary image in Fig. 2(C)]. This allowed estimation of cell length, width, and orientation (degrees of inclination of the long axis of the ellipse relative to the flow axis). Cellular ellipticity was defined as the ratio of length to width. To compare measures across different sized vessels we quantified the size of the vascular lumen, or more properly here the channel in the plasma layer through which cells flow. This was defined as the widest extent, i.e., perpendicular to the flow direction, required to accommodate the entire cell profile.

Using the above method, we estimated cell shape in 123 unique vessels (31 to 49 per subject). Vessels were selected where flow was single-file and cell aggregates were rare, biasing our sample toward narrower vessels and lower hematocrits. Blood column images were required to be of sufficient quality to see cells, further biasing our sample toward longer vessels where cells were visible for more time (e.g., mean vessel length was 53 µm, cf. our sample at 67 µm). We also required vessels to be visible for at least 1 s of continuous imaging. Some 24% of vessels met these inclusion criteria. Lumen diameter for these vessels ranged from 3.2 to 8.4 µm (mean 5.1 µm); samples were broadly equivalent across subjects, with mean diameter of 4.9 to 5.8 µm and mean length of 64 to 71 µm.

Examples are shown in Fig. 4, representing marked differences in cell shape across our dataset. Mean ± standard deviation in cell length was 8.5 ± 1.6 µm, with cell width of 4.7 ± 0.7 µm. Cell ellipticity was 1.85 ± 0.47. Along the flow centerline, cell length was 7.6 ± 1.7 µm and width was 5.1 ± 1.1 µm, with the latter measure being defined as the lumen diameter. These latter measures of length and width differ from the above figures due to nonaxial orientations; the long axis of a cell was inclined to the flow axis by 17.4 ± 19.4° (range = 0.3° to 85.4°). Hematocrit was 0.41 ± 0.10 (range 0.15 to 0.64).

 

Fig. 4. Examples of the capillary blood column and appearance of the average cell. Panels are arranged in descending order of capillary diameter: from (A) 7.0 µm to (H) 3.6 µm. Upper panel parts show the probability map for each cell (see text) on the left, and a binarized representation on the right. Lower panel parts show the blood column, wrapped for visualization. Flow direction is rightwards. Scale in (A) applies to all panels; in the upper part of panels it equates to 5 µm and in the lower part to 25 µm (i.e., upper panel parts are zoomed ×5).

Download Full Size | PDF

As lumen diameters narrowed, cells transitioned from their longest direction being perpendicular to the flow, to being aligned with it (Pearson’s r = 0.67, p < 10⁻¹⁷). Cell shape transitioned from being rounder to more elliptical (r = −0.47, p < 10⁻⁷), and hematocrit was reduced (r = 0.40, p < 10⁻⁵). Figure 4 confirms these associations, with cells aligned perpendicular to the flow in wider vessels [Figs. 4(A) and 4(B)] transitioning to oblique orientations [Figs. 4(C) and 4(D)], to more axially aligned cells [Figs. 4(E) and 4(F)], and finally to the most elongated shapes in the thinnest capillaries [Figs. 4(G) and 4(H)].

The average cell was inclined to the axis of flow in many vessels; this was surprising, and does not appear to have been demonstrated previously for capillaries in vivo. However, numerical simulations indeed predict oblique cell orientations, stable over time, particularly in larger capillaries as found here [23]. In general, stable cell orientations are more likely as vessels narrow or as flow becomes faster [24,25]. In this regime of stable orientation, cells are more obliquely oriented in wider and in slower vessels [23–25]. Outside of this regime, cells tumble without a well-defined orientation. Notably, the oblique cells we observed showed reduced image quality [e.g., Fig. 4(C)]; this also holds true for the occasional oblique cells noted in some vessels with otherwise high image quality [e.g., Fig. 4(H), upper rows]. A likely explanation is that the orientation of these cells was unsteady, creating blur when averaging cell appearance over time (signal-to-noise ratio for single frames was not sufficient to confirm this). Such changes in orientation could occur due to continual oscillation in flow [26], which occurs in the vast majority of retinal capillaries as they directly manifest pulsatile variations in the cardiac output [7–10,16]. Cell geometry may also evolve with distance from a branch, becoming more stable over distances of the order of 60 µm [27] which exceeds the length of many vessels analyzed here. Hence it may be expected that a stable geometry may never, or only partially, be reached within some vessels of the type analyzed.

The cell shapes observed included both axisymmetric [e.g., Figs. 4(A) and 4(F)] and non-axisymmetric [e.g., Figs. 4(D) and 4(H)] forms. Modeling and in vitro studies demonstrate that red blood cells generally transition from symmetric, centered, “parachute” shapes to nonsymmetric, lateralized, and more efficient “slipper” shapes under increasing shear stress, i.e., as vessels narrow and as flow becomes more rapid [3,24,28–30]. Accordingly, it may be possible to use our measurements of cell geometry to infer shear stress, a primary driver of endothelial damage in vascular disease [31].

One might expect vascular resistance to be elevated in narrower vessels, and hence that flow should be overall slower. However, we did not find any correlation between measures of flow velocity and lumen thickness (nor between velocity and measures of cell shape or orientation; p > 0.05). This may be because elevated resistance to cellular passage causes narrower vessels to receive fewer red blood cells and more plasma [32], consistent with the association mentioned above between lumen diameter and hematocrit. Hence, narrower vessels need not be associated with slower velocities.

There are several limitations to this study. In vivo imaging of the retina is limited by resolution of 1.5–2 µm [33], and light levels are limited by safety considerations [17]. This makes it hard to assay individual cell shape, although the average does strongly resemble shapes reported in other studies [27]. As with in vitro studies, we cannot easily study cell shape in 3D, and so cells could be larger than they appear or obliquely oriented out of the plane. Retinal vessels need not have their flow axes oriented entirely within our imaging plane, which could shorten the apparent length of cells; they also stratify at different depths which creates the potential for blur to confound shape measurements (though ratio-based measures such as ellipticity should be robust to this). Our method did not assess the anatomical lumen due to the low contrast of the vessel wall; this means that lateralization within the lumen predicted for non-axisymmetric shapes [27,30] cannot readily be studied. In our analysis we considered the average cell, precluding analysis of variability between successive cells in terms of size, orientation, or lateralization in the vessel. Similarly, averaging over time for a given cell precludes the opportunity to consider its variation over time. However, such information could be extracted from our data in many vessels [e.g., Fig. 4(H), top, shows a subset of oblique cells whilst others are axially aligned; presumably some cells begin obliquely and then rotate to align with the flow [27] ].