In vivo measurement of organelle motility in human retinal pigment epithelial cells.

Retinal pigment epithelial (RPE) cells are well known to play a central role in the progression of numerous retinal diseases. Changes in the structure and function of these cells thus may serve as sensitive biomarkers of disease onset. While in vivo studies have focused on structural changes, functional ones may better capture cell health owing to their more direct connection to cell physiology. In this study, we developed a method based on adaptive optics optical coherence tomography (AO-OCT) and speckle field dynamics for characterizing organelle motility in individual RPE cells. We quantified the dynamics in terms of an exponential decay time constant, the time required for the speckle field to decorrelate. Using seven normal subjects, we found the RPE speckle field to decorrelate in about 5 s. This result has two fundamental implications for future clinical use. First, it establishes a path for generating a normative baseline to which motility of diseased RPE cells can be compared. Second, it predicts an AO-OCT image acquisition time that is 36 times faster than used in our earlier report for individuating RPE cells, thus a major improvement in clinical efficacy.


Introduction
Retinal pigment epithelial (RPE) cells are composed of organelles that are under constant motion as they execute cellular and molecular tasks, which encompass essentially every aspect of RPE cell physiology. Thus, organelle motion is fundamental to RPE function [1]. Because of the key role RPE cells play in the interchange of nutrients, ions, water and waste between the choriocapillaris and the photoreceptors, this monolayer of cells is the initiation point of many retinal degenerative diseases, most notably age-related macular degeneration, but also Best's disease, Stargardt's disease, retinitis pigmentosa, and others [2]. Early manifestation of RPE degeneration occurs at the cellular level and so detection, monitoring, and treatment should be most effective when targeting processes at this level. Consistent with this, animal studies have shown that diseased RPE cells exhibit abnormal organelle motility [3][4][5], thus pointing to motility as a potentially sensitive indicator of cell health. However, how these findings translate to the living human eye remains unknown. The primary obstacles to making use of RPE organelle motility as a biomarker has been the limited resolution of in vivo imaging techniques necessary to delineate these cells and the strong waveguiding property of the overlying photoreceptors that mask the RPE. Adaptive optics optical coherence tomography (AO-OCT) and scanning laser ophthalmoscope (AO-SLO) approaches have overcome these obstacles to provide detailed mapping of the RPE cell mosaic [6][7][8][9][10][11][12][13][14][15][16]. AO-OCT delineates RPE cells by taking advantage of the cells' intrinsic organelle motion to Despite these early demonstrations, measurements have been limited to morphological properties of RPE cells, such as cell density, area, and photoreceptor-to-RPE ratio. Measurement of cell function as expressed by organelle motion remains unexplored in living eyes. In a previous RPE imaging study, we found that AO-OCT volumes acquired at 3-minute intervals exhibited different speckle patterns in the RPE cells, sufficiently so that averaging of these volumes suppressed speckle noise and revealed the RPE cell mosaic [10]. Cancellation of speckle by averaging confirmed that the motility dynamics of these cells occurred on a time scale no greater than 3 minutes. In a subsequent AO-OCT pilot study, we characterized the temporal dynamics of RPE motility in two normal subjects and found them to be significantly shorter, being on the order of seconds rather than minutes. Averaging of images acquired at this shorter time interval also resulted in delineation of the RPE cell mosaic, thus substantiating the presence of faster dynamics [18].
In this study, we extend these first motility measurements: acquiring measurements on seven subjects, optimizing the sampling protocol for more complete characterization of the RPE motility dynamics, and confirming results on two separate AO-OCT imaging systems. We quantify dynamics in terms of an exponential decay time constant (τ), the time for motility to decorrelate the amplitude speckle field across individual RPE cells. This characterization has two fundamental uses. First, it provides a path for obtaining a motility baseline for normal, healthy RPE to which we can compare diseased RPE. Second, it predicts the extent to which our previously reported imaging experiments for individuating RPE cells can be reduced from its 90-minute length, a necessary requirement for clinical translation. We further demonstrate generalization of the clinical utility of RPE motility by obtaining results that are consistent on two AO-OCT platforms. These platforms, while different, share similar performance attributes, most critically the coherence volume of their imaging beam, which determines the number of organelles that contribute to each speckle.

Description of AO-OCT imaging systems
RPE motility measurements were acquired using two AO-OCT imaging systems, one at Indiana University and the other at the U.S. Food and Drug Administration (FDA). The two systems share similar optical designs (see [17,19,20] for description) and acquired RPE images of similar clarity [10,17]. Key system parameters are summarized in Table 1. Critical for this study, the imaging coherence (speckle) volumes of the two AO-OCT systems were essentially the same: 24.2 µm 3 (Indiana) and 23.1 µm 3 (FDA) in retinal tissue (see Table 1). Thus speckle formed by each system encapsulated the same total number of organelles and therefore should yield similar speckle dynamics when imaging in the same retinal tissue. The systems acquired images fast enough to track RPE speckle dynamics up to 2.75 Hz.

Experimental design
Subject: Seven subjects, ranging in age from 21 to 49 yr (S1 = 21, S2 = 26, S3 = 47, S4 = 49, S5 = 27, S6 = 32 and S7 = 34 yr old) and free of ocular disease, were recruited for the study. Four subjects (S1-S4) were imaged with the Indiana AO-OCT system, and three different subjects (S5-S7) were imaged with the FDA AO-OCT system. All subjects had best corrected visual acuity of 20/20 or better. Eye lengths ranged from 23.56 mm to 25.40 mm as measured with IOLMaster (Zeiss, Oberkochen, Germany) and were used to correct for axial length differences in retinal image scale following Bennett et al. [21]. All procedures on the subjects adhered to the tenets of Helsinki declaration and approved by the Institutional Review Board of Indiana University and the FDA, respectively. Written informed consent was obtained after the nature and possible risks of the study were explained. AO-OCT beam intensity was measured at the cornea (Table 1) and below the safe limits established by ANSI [22] for the retinal illumination pattern used (bidirectional point scan over a square area) and length of the experiment. The right eye was cyclopleged and dilated with one drop of Tropicamide 0.5% for imaging. The eye and head were aligned and stabilized using a bite bar (Indiana) or chin and head rest (FDA) attached to a XYZ translation stage.
Imaging protocol: AO-OCT volumes were acquired at 7° temporal to the fovea. Foveal scans were avoided to mitigate scanning beam distraction and additional eye motion by the subjects. For each subject, 30 to 62 AO-OCT videos were collected at 1 min or less interval and time stamped. As we did not know which video acquisition parameters would best capture the dynamics of RPE organelle motility, three experiments were conducted using different imaging parameters (Table 2). We used two main video durations: short videos (1.8 s for Experiment 1 and 1.3 s for Experiment 2) were acquired in all seven subjects, and long videos (8.6 s and 9 s video durations: Experiment 2 and Experiment 3) were acquired in four of the seven subjects (S2, S5-S7). Experiment 1 and 3 images were acquired at high speed (500 kHz A-line rate; 5.5 Hz volume rate), and Experiment 2 images were acquired at lower speed (210 kHz A-line rate; 2.3 Hz volume rate). These speeds enabled temporal dynamics of RPE cell motility to be captured up to 2.75 Hz and 1.2 Hz, respectively. During the long video acquisitions, subjects were instructed to blink once every 5 s to maintain good tear film quality while minimizing the number of volumes lost due to blinking. Time interval between consecutive videos ranged from 25 s to 80 s.
Prior to collection of the AO-OCT volumes, system focus was adjusted to optimize cone image quality, determined by visual inspection of cones in en face images that were projected axially through the portion of the AO-OCT volume that contained the cone inner/outer segment junction (IS/OS) and cone outer segment tip (COST) reflectance bands [23].

3-D image registration and data analysis
3-D registration was applied to the AO-OCT volumes, followed by layer segmentation of the cone IS/OS, COST, and RPE layers. The B-scan based 3-D registration algorithm [24] was processed on a Graphical Processing Unit (GPU) via the Compute Unified Device Architecture (CUDA; NVIDIA, Santa Clara CA) parallel programming platform. The segmentation and data analysis steps were processed with custom algorithms developed in MATLAB (Mathworks; Nattick, MA, U.S.A.). Registration and segmentation were based on further advancements of previously published algorithms [24,25], and data analysis algorithms were new for this study.
Volumes with excessive motion artifacts were removed. Remaining volumes were combined into a single, time-stamped video and then registered in all three dimensions with subcellular accuracy to correct for motion artifacts. Registration to this level was necessary to prevent masking of organelle motility within individual RPE cells, which themselves were manually identified in the averaged, registered RPE en face image. Two additional layers were selected as motility controls (see Fig. 1): reflections from the cone layer (IS/OS + COST) and outer nuclear layer (ONL). Without visible light stimulation, the cone reflectance (IS/OS + COST) is known to be largely stable over minutes and exceedingly so over several seconds [26], the temporal bandwidth of the RPE motility dynamics we measured in this study. Reflectance changes associated with disc shedding are infrequent and occur over much longer periods [27][28][29][30][31]. Therefore, we expected the cone reflections to have a relatively long decay time constant (τ), which we used for normalizing the RPE signal. In contrast, the ONL reflection is considerably weaker, so much so in our images that white noise dominates and results in a delta-like τ.
To quantify the temporal dynamics of RPE motility, a correlation function (CF) was calculated from the dynamic speckle pattern of individual RPE cells. Pixels within each RPE cell, defined by a Voronoi region [10] (see Fig. 1(E and F)), were used to compute the correlation coefficient (CC) between a reference (at time t 1 ) and each consecutive image (at time t 1 , t 2 , t 3 , …) in the image series of that RPE cell: where rc R is the intensity of the rth row and cth column pixel in the reference image, R is the mean intensity of all pixels in the cell of the reference image, is the mean intensity of all pixels in the cell of the nth image at time point t n-1 .  r  c is summation over all pixels within the RPE cell as defined by its Voronoi region. CC equals one if no changes occur between the two images (i.e., absence of motility). CC was computed for each RPE cell (see Fig. 1(F)) then averaged across all of the identified RPE cells at the same time point and plotted against time as CF in the local area. The same analysis was also applied to the corresponding regions of the two control layers: cone and ONL for comparison and normalization purposes. On four of the subjects (S2, S5-S7), additional CF values were computed at different depths across the RPE-Bruch's membrane (BM) complex, thus testing for motility differences that might occur as organelle types and concentrations change across the complex depth, e.g., melanosomes, phagosome, and lipofuscin [32]. The five tested depths were apical, middle and basal areas of RPE, rod outer segment tip (ROST), and BM layer. The CF is predicted to follow an exponential decay model [33], which is described as The correlation function is defined by two independent terms. The first term, ( ) A t , captures residual motion contributions from the eye and system, e.g., uncorrected scanning artifacts, which are reduced in magnitude to the sub-cellular level after image registration is applied.
, captures organelle motility. τ is the de-correlation time of the speckle pattern and the metric we used to quantify RPE motility. Unfortunately, separating the two terms is problematic as both are unknown. Here we took advantage of the fact that all pixels of a given A-scan are simultaneously acquired, and therefore A(t) must contain the same system and uncorrected eye motion artifacts regardless of depth in the AO-OCT  . (1). A averaged ) are also aster than ming our cone and asured for the first sample point at 0.2 s, again confirming expectation. Similar behavior of the cone and ONL signals were found in the other six subjects.

AO-OCT
To determine τ for RPE, system and residual motion errors were removed by normalizing to the cone reflectance as shown in the motility functions plotted in Fig. 3(B). The plot shows that CC is high at the beginning of the image series, indicating an initially stable speckle pattern. But with the lapse of time, CC decreases monotonically until after a few seconds it reaches a plateau determined by features common to all images in the series. Figure 3(C) shows the normalized CC measurements and corresponding CF fit (Eq. (3) for the seven subjects. From Experiment 1, the CF fit gives an average τ of 2.7 s for RPEs based on 363 to 475 RPE cells per subject. However, the Experiment 1 video duration of 1.8 s was too short to sample across the drop of the motility curve, including near the time constant value (2.7 s).
To assess whether this undersampling may have caused an error in our CF fit, Experiment 2 and 3 used longer video durations (~9 s) in order to more completely capture the RPE motility exponential decay. These experiments yielded an average τ = 4.9 s based on 475 to 651 cells (see Fig. 3(c)). Note that the three subjects in Experiment 2 were imaged with the FDA system, and the averaged time constant (τ = 5.1 s) from those three subjects was almost twice long as the averaged time constant (τ = 2.7 s) from Experiment 1, but close to the time constant (τ = 4.2 s) from Experiment 3 obtained with the Indiana system.  Table 2 for acquisition parameters). The gap between ~10 s and ~100 s for this data set is due to the time interval between two consecutive videos. (see Section 2.2). Fitted time constants are given in the key for each subject in each experiment.
To determine whether these differences in time constant were due to the use of different imaging systems or to an undersampling of CC, we used the data from Experiment 3 and refitted the exponential decay after removing the data points between 1.8 s and 9 s to mimic the short video scenario. The model yielded a similar time constant (τ = 2.8 s) as the one (τ = 2.9 s) achieved from the short video in Experiment 1. A shorter time constant (average τ = 2.7 s) was also produced when the same strategy (remove data points between 1.8 s and 9 s) was applied to the Experiment 2 data, thus confirming that undersampling was the root cause of the underestimation in Experiment 1. Regardless, results from any of the three experiments point to a motility time constant that is orders of magnitude less than the 3-minute interval we used in our first RPE imaging study [10].
A repeated measurement on RPE motility, performed using two long videos from the same subject (S6) on the same patch of retina, resulted in time constants of 4.74 s and 4.61 s. The small dif possible varia

Optimizin
In this study, minutes to les assessed visua duration of 3 clinical use. T retaining clari of the previo measurements images. This imaging sessi  With an a visually in th interval of 7. more clearly (Fig. 4(C). Im show better d compared to Fig. 4 butions of RPE ntrast, we calc ls (with TI = 0 mber of image OCT file timest ns that we did mposition of th at would be in ur measuremen ither location lts are presente al RPE cells a 4(A). In con (Fig. 4(B) Fig. 4. are difficult to ntrast, the aver th individual R the longer D power spect d decreased no n Fig. 4 Fig. 4. The slight SNR increase with TI<τ is partially from the RPE speckle pattern variation, given the total image duration is 8 s (0.23 (TI) × 35 (# of images)), about twice longer than the predicted time constant ( 4.2 τ = s on S2). However, the SNR improvement is primarily attributed to averaging, which reduces the variation of OCT image noise (mainly photon noise) that sets the noise floor in the 2-D power spectra of our images. We observed the speckle patterns change little between Fig. 5(B) (N = 1) and Fig. 5(E) (N = 5), and somewhat alter their appearance in Fig. 5(C) (N = 35) in which the RPE clarity is clearly inferior to the en face image in Fig. 5(D) (N = 35). The comparison between Fig. 5(C) and Fig. 5(D) suggests the speckles remain largely correlated for TI = 0.23 s, and longer TI is required to enable speckle decorrelation to further improve RPE clarity. This is further illustrated by comparison of the en face images in Fig. 5(C) and 5(G), which appear qualitatively similar even though one is produced with 35 images but TI<τ and the other is produced with 5 images but TI τ > , confirming the relative averages that produce equivalent SNR (horizontal crossing point of dash line in Fig. 5(A)). Similar results were obtained for the same analysis in other subjects as well. Note, eye motion may result in black pixels and/or lines in single en face images, which results in zero contribution to those pixels and/or lines in the image average. Thus variation in the effective number of images that are averaged at any pixel is expected. This is demonstrated by the false-colored images of Figs. 5(F) and 5(H) that show the number of volumes averaged varies from 0 to 5 depending on pixel location in the image.

Discussion
We developed a novel AO-OCT-based method for measuring organelle motility in individual RPE cells. Our method detects the time-dependent fluctuations in scattered light intensity (speckle) from the cells and relates these fluctuations to physical properties of the same cells (organelle motility in this study). This dynamic speckle technique has been successfully used in prior studies for capturing intra-cellular motion [35], enhancing subcellular details [36,37] and characterizing motility of other retinal tissue layers [38,39]. Using motion-evoked speckle changes here, we successfully delineated RPE cells and measured the RPE motility decay time constant (τ) in seven healthy subjects. Because our technique relies on the 3-D resolution provided by AO-OCT and the speckle decorrelation dynamics inherent in RPE organelle motility, we expect the same approach can be applied at any retinal location, including the fovea. In support of this, the RPE mosaic has previously been observed with AO-OCT at numerous locations across the macula with different densities of cones and rods [10,40].

Important parameters of the AO-OCT design
Our measurements of τ are consistent between the two imaging systems that were used in this study (Fig. 3(C)), a finding perhaps not unexpected given their similar designs. Consistency provides evidence of generalizability and repeatability of our method, but also raises questions about the role that differences between the two systems might play, as for example our systems differed in physical size and design strategies for the sample arm, and used different wavefront correctors, wavefront sensors, linescan detectors, light sources, and acquisition software. In AO-OCT and OCT images, speckle arises from the coherent interaction of the imaging beam with localized random scatterers in the tissue (in our case RPE organelles) that result in constructive and destructive interference. Because speckle is the fundamental physical phenomena that underlies our RPE motility measurements, the optical properties that control speckle size and shape (i.e., coherence volume) must drive system design and be fundamental for achieving consistency across AO-OCT platforms. Theoretically, these properties reduce to imaging wavelength and numerical aperture (for beam entering the human eye). For our two systems (Table 1), we used similar imaging wavelengths (785 nm and 830 nm) and same beam diameters at the eye (6.7 mm), resulting in almost identical coherence volumes (24.2 μm 3 and 23.1 μm 3 ) in tissue.

Possible organelle sources of the motility signal
The RPE cell interior is in constant motion with intracellular transport of organelles along three-dimensional scaffolding of actin filaments and microtubules. These protein-enabled thoroughfares allow organelles to traverse the cell in a second or so. Surprisingly, the literature on organelle content in human RPE is limited [32], and thus we can only speculate as to which types may contribute to the motility signal in our AO-OCT images. Prominent organelles in the RPE include melanosomes, phagosomes, and lipofuscin. The first two are well known for their motility behavior [1]. Both have high refractive indices (n = 1.7 of melanin [41] and 1.43 of photoreceptor outer segment remnants [42]) relative to their cytoplasmic surround (n = 1.37 [41]) and thus are likely sources of reflected light. Consistent with this, several studies point to melanin, which is synthesized and compartmentalized in melanosome organelles [43], as the primary source of the OCT signal in the RPE band [10,44,45]. Animal model studies show that the melanosome movement in RPE cells can be faster than 1 µm/s, enabling these organelles to quickly traverse the coherence volume of our AO-OCT systems (~2.5 × 2.5 × 4 μm 3 ). If a primary contributing source in our images, melanin motility will have direct clinical significance as interruption of its transport across the cell has been shown to cause blindness in humans [4].
Phagosomes also exhibit motility behavior. They form from the membranous discs that are regularly pruned from the tips of photoreceptor outer segments. These pruned tips descend into the RPE cell where they are engulfed and phagocytized. Regulating phagosome movement is thought to involve the same molecular machinery that regulates melanosome movement [1,3,5]. We know the initial pruning event (disc shedding) causes a pronounced, transient disturbance in the reflection at the photoreceptor/RPE interface, as has been recently observed using AO-OCT [30,31]. Thus it is reasonable to expect the downstream phagocytosis process to generate similar reflectance disturbances within the RPE cell as the engulfed contents are digested, but this awaits testing.
Finally, lipofuscin is the undigested byproduct of the phagocytosis event, composed of a heterogeneous mixture of lipids and proteins that accumulate in the RPE cell with age. The optical properties of this byproduct must vary with mixture concentrations, which makes characterizing its reflectance contributions difficult. Nevertheless, if it were to strongly contribute, we would expect RPE image clarity to also vary with age (either increase or decrease). Our imaging results to date have not demonstrated an age dependence, but the number of subjects we have imaged is limited and the trend could be masked by other factors, as for example inter-subject variation and age-dependent properties of other organelles.

Depth-dependent motility in RPE layer
Some of the primary RPE organelles (e.g., melanosomes and lipofuscin) have been reported to stratify to different depths in the RPE [46]. Given that their different physical size and physiological role may lead to different motility dynamics, we hypothesized a possible variation in CF with depth in the RPE. To test, we calculated CF at three depths inside the RPE layer, a fourth one at BM, and a fifth at ROST. The three RPE depths (marked as (1)(2)(3) in Fig. 6(A)) were selected at the apical (expected high melanosome density [46]), middle (the strongest RPE reflectance in our AO-OCT images), and basal (expected high lipofuscin density [46]) portions of the RPE layer. The apical and basal locations were chosen to be ~ 4 µm above and below the peak intensity, a separation large enough to generate different speckle patterns in the corresponding en face images (Fig. 6(D-F)), but small enough to avoid signal contamination from the adjacent ROST and BM layers. Although the averaged en face views at the th different spec   Fig. 6(G,K)) of 4.9 s is closest to that of the middle RPE band (τ = 4.3 s). These measurements demonstrate a clear path for establishing motility baselines across the RPE complex in normals and with which comparisons can be made against pathologic retinas.

Influence of motion on speckle decorrelation
We removed motion contributions from the eye and imaging system by registering the acquired images with subcellular accuracy (confirmed by the resolution of photoreceptor and RPE cells in the averaged images) and normalizing the correlation function with the cone signal (Eq. (2)). While this approach appeared effective, we were unable to directly determine what residual motion artifacts might have remained and what their influence, if any, might have been on the observed speckle decorrelation, i.e., τ ~5 s for RPE cells. Unable to quantitatively assess, we resorted to a qualitative examination of the possible contributions of motion. To do so, we took advantage of the notable difference in RPE clarity in images acquired with TI = 0.23 s and 7.8 s (see Fig. 5). We examined whether this clarity difference could have been caused by angle compounding (a pupil effect caused by eye rotation, thus a consequence of eye motion) [47] or minute levels of image blur induced by residual eye motion.
To start, we extracted the three-dimensional eye motion from a representative OCT data set and plotted as displacements in the x, y, and z directions in Fig. 7. The displacements were computed from the output of the 3-D registration. B-scans with registration coefficients larger than the selected coefficient threshold were used for this calculation (the mean displacement was calculated as the displacement of the volume, which is plotted in the figure. As evident from the figure, considerable eye motion occurred during both video acquisitions with motion traversals larger than the resolution element of the two imaging systems (2.4 × 2.4 × 4.2 µm 3 and 2.5 × 2.5 × 3.7 µm 3 ). The overall 3-D displacements were comparable, 29.2 ± 15.5 μm and 43.3 ± 19.8 μm (mean ± std) for TI of 0.23 s and 7.8 s, respectively. Similar results were obtained for the same analysis in other subjects as well. Comparable retinal motion implies comparable range of angular illumination of the retina and therefore comparable levels of contrast enhancement of the RPE.
Next, we examined the possibility of motion-induced image blur. Based on the algorithmic design of our image registration method, we expect registration accuracy to worsen with the magnitude of the eye motion, both for motion parallel to and perpendicular to the fast B-scan direction. This trend coupled with the comparable motion displacements in the two imaging scenarios (TI = 0.23 s and 7.8 s) points to comparable registration accuracy for both scenarios and thus comparable motion-induced image blur and contrast improvement. The fact that we observed notably different RPE contrast suggests residual eye motion is an unlikely contributor. In summary, we have found no evidence to link our measured time constant and RPE contrast improvement to sources other than the physiological dynamics internal to the cells imaged. Consistent with this conclusion, the RPE time constants we have measured (~5 s) fall well within the range reported for other biological tissues (0.4-34 s) that were measured under highly controlled and stable environments [36,38,48,49], albeit this range is relatively large.

Clinical i
Our method h by its time c interactions o complex light for τ will allo Inhibited tran [5] sources to separate the RPE layer or other means of RPE cell contrast), and AO seems to be the most straight-forward technique to achieve the required lateral resolution. What we have demonstrated in comparable measurements on two different AO-OCT systems is that once the necessary coherence volume is achieved, other factors in system performance appear less important. This increases the potential of validation as the measurement is not tied to a highly specialized technique. The second condition is proper validation in normal, healthy eyes. Repeatability and reproducibility measurements must be performed on a sizeable cohort of subjects (perhaps hundreds) to establish a normative baseline. The relatively recent development of OCT-measured NFL thickness as a surrogate for GCL loss and glaucoma disease stage points to a similar validation pathway for RPE motility in degenerative diseases. Finally, motility differences in RPE-related diseases must be measured, along with studies that definitively determine the organelle source of motility speeds (and optical decorrelation). The path is arduous, while the reward in terms of AO clinical translation may be significant.