Current adaptive optics flood-illumination retina cameras operate at low frame rates, acquiring retinal images below seven Hz, which restricts their research and clinical utility. Here we investigate a novel bench top flood-illumination camera that achieves significantly higher frame rates using strobing fiber-coupled superluminescent and laser diodes in conjunction with a scientific-grade CCD. Source strength was sufficient to obviate frame averaging, even for exposures as short as 1/3 msec. Continuous frame rates of 10, 30, and 60 Hz were achieved for imaging 1.8, 0.8, and 0.4 deg retinal patches, respectively. Short-burst imaging up to 500 Hz was also achieved by temporarily storing sequences of images on the CCD. High frame rates, short exposure durations (1 msec), and correction of the most significant aberrations of the eye were found necessary for individuating retinal blood cells and directly measuring cellular flow in capillaries. Cone videos of dark adapted eyes showed a surprisingly rapid fluctuation (~1 Hz) in the reflectance of single cones. As further demonstration of the value of the camera, we evaluated the tradeoff between exposure duration and image blur associated with retina motion.
©2006 Optical Society of America
The optical resolution of retina cameras is significantly increased by correcting the eye’s wave aberrations across a large pupil, using for example adaptive optics (AO) . This increase permits the observation of retinal structures at the cellular level, which could not otherwise be seen in the living eye. AO has been successfully applied to a variety of retina camera modalities including the conventional fundus (flood illumination) camera [2–8], confocal scanning laser ophthalmoscope (cSLO) [9,10] and optical coherence tomography (OCT) [11–15]. The latter two (cSLO and OCT) are typically realized by raster scanning a focused point source across the retina. Scanning readily lends these cameras to high frame rates, e.g., video rate imaging. Conventional fundus imaging, on the other hand, consists of flood-illuminating a patch of retina and then recording, with a 2D detector, the reflected light that exits the eye. While flood illumination is the most common platform for AO retinal imaging, it has not achieved the imaging rates routinely used with AO-SLO and AO-OCT instruments. This is partially because lower rates have been sufficient for many of the experiments intended for these cameras. Higher rates, however, would certainly benefit applications that require quick and extensive surveying of the retina landscape, and temporally resolving fast retinal dynamics (such as blood flow). In general higher rates should improve effectiveness of the AO flood illumination modality as a research and clinical tool. Another desirable camera feature is flexible control of the exposure duration. Current AO flood-illumination cameras are often hardwired to a fixed value or exceedingly long (up to 100 msec). The former prevents the use of optimal exposure durations for specific experiments, and the latter may expose the retinal image to unacceptable retina motion blur . Interestingly, some high resolution flood-illumination cameras have reported even longer exposures, e.g. 300 msec .
The camera components that largely limit frame rate and exposure duration are the illumination light source and 2D detector array (CCD) that captures the retinal image. General light source requirements include low spatial coherence (reduces speckle noise), a narrow spectral band (avoids ocular chromatic aberrations), uniform illumination, and high optical energy deliverable during a short exposure (overcomes the ~10-4 loss in the eye and retina motion blur). In addition, the Lagrange invariant fundamentally limits the efficiency of the illumination channel as dictated by the size and divergence of the light source, the numerical aperture of the eye, and the desired illumination patch (field of view). To meet these requirements, AO flood illumination cameras have employed Krypton and Xenon flash lamps or Mercury and Xenon arc lamps to illuminate the retina. For detection, they have relied on CCD cameras of high sensitivity, but with relatively slow readout rates.
As an alternative, we investigate here a novel bench top AO flood-illumination camera that incorporates, for illuminating the retina, superluminescent and laser diodes cascaded with lengths of multimode optical fiber. Both sources are evaluated separately and used in conjunction with a high-speed scientific-grade CCD. In addition to faster modulation (higher frame rates), higher efficiency (shorter exposure durations) and greater software control, the fiber-coupled light sources are less expensive and more compact than flash and arc lamps. One distinct disadvantage is that SLDs and laser diodes are commercially available at only specific wavelengths. This is unlike arc and flash lamps that radiate broadly over the visible and near-infrared spectrum, and are tunable with choice of an appropriate interference filter.
We demonstrate the utility of our camera by imaging the in vivo retina at various locations and at various frame rates up to 60 Hz (continuous) and 500 Hz (short burst). Additionally, we demonstrate the flexibility of the camera to image over a broad range of exposure durations (1/3 to 100 msec) by using this feature to directly assess the impact of retina motion blur on cone image quality. Accounts of this work were presented at the 2003 and 2005 ARVO, 2004 OSA, and 2004 SPIE meetings [18–21]. We previously employed a near-infrared prototype version of the fiber-coupled superluminescent diode (SLD) to evaluate an AO-OCT camera .
2.1 Adaptive optics retina camera
A flood-illumination retina camera was developed for collecting aerial images of microscopic structures in the living human retina. The camera consisted of three sub-systems: (1) adaptive optics for compensation of the eye’s wave aberrations, (2) pupil retro-illumination and fixation channel for alignment of the subject’s eye to the camera, and (3) retinal imaging using a novel fiber-based light source and scientific-grade CCD. A schematic of the camera is shown in Fig. 1.
The AO system is described in detail in Zhang, et. al.. In short, the system consists of a Shack-Hartmann wavefront sensor (SHWS) and 37 actuator deformable mirror (Xinetics, Inc.) that are controlled in closed loop via a desktop computer. The sensor employs a 0.75 mW pigtailed single mode SLD operating at 788 nm (Δλ=20 nm). A 17×17 lenslet array (f=24mm; lenslet dia.=0.4mm), placed conjugate to the eye’s pupil, samples the exiting wavefront across a 6.8 mm pupil. The deformable mirror is positioned upstream of the lenslet array at a plane conjugate to the eye’s pupil. The middle horizontal row of mirror actuators traverses 6.8 mm at the eye’s pupil. The AO operates at up to 22 wavefront measurements and corrections per second, although reduced rates down to 10 Hz were used to permit additional diagnostic tools to run in the background. While these slower rates are not optimal, it has been established that the dominant temporal components in the human eye’s wave aberrations occur at low frequencies and the correction of these is typically sufficient for yielding sharp images of the cone mosaic and retinal capillary bed [3,10,13]. Exposure level at the cornea for the 788 nm SLD was 5 µW, more than 117 times below the maximum permissible exposure for continuous intrabeam viewing recommended by ANSI .
The illumination channel for retinal imaging consists of two light sources that were evaluated separately. The first is a 10 mW pigtailed single mode SLD (λ=679 nm, Δλ=10.8 nm) coupled to 25 meters of multimode step index optical fiber (Lucent Technologies, Inc.). The fiber has a numerical aperture (NA) of 0.22, core diameter of 105 µm, and core refractive index (ncore) of 1.457 at λ=0.633 µm. As pictorially shown in the Fig. 1 inset, light that is directed into the fiber is distributed among the fiber modes that propagate at different axial velocities (via modal dispersion). The 25 meters was of sufficient length to cause the time delay between exiting modes to be larger than the temporal coherence length of the SLD so as to effectively mitigate the source’s spatial coherence and reduce speckle contrast. The exit face of the fiber therefore acts as a spatially incoherent, quasimonochromatic light source [23–25]. Calculations  for estimating the necessary fiber length are detailed further below.
The second source is a stock 200 mW multi-mode laser diode (λ=670 nm). The spectral bandwidth of the diode was not provided by the manufacturer, but its temporal coherence length, after the multimode fiber, was measured at about 250 µm in air. This corresponds to a spectral bandwidth of approximately 0.8 nm. Owing to the significantly narrower spectral bandwidth, the laser diode was coupled to a much longer (300 meters) and higher NA multimode step index optical fiber (Lucent Technologies). The fiber has a numerical aperture (NA) of 0.39, core diameter of 200 µm, and core refractive index (ncore) of 1.457. Light exiting the two fibers were directed into the subject’s eye where the SLD flood illuminated a 1° patch of retina and the laser diode a 1.8° patch. The tips of the fibers were conjugate to the subject’s retina. To achieve improved SLD beam uniformity, a fiber mode scrambler was attached to the multimode fiber for some of the eyes examined. Exposure duration, light intensity, and delay between consecutive images were computer controlled.
A back-illuminated scientific-grade 12 bit CCD (Quantix 57, Ropier Scientific, Inc.) captured aerial images of the retina whose acquisition was synchronized to the strobing SLD and laser diode. The CCD array was 1056x530 pixels and consisted of a light-sensitive region of 512×530 pixels plus a similar storage area underneath the frame transfer mask. Custom dielectric beamsplitters were designed to reflect and transmit the corresponding flood illumination (670 and 679 nm) and SHWS (788 nm) wavelengths. This allowed simultaneously wavefront correction and retinal imaging without loss or mixing of light. The chromatic aberration of the eye caused a shift in focus between the two wavelengths that was offset by axially translating the Quantix CCD camera.
The role of the multimode optical fibers is to reduce spatial coherence of the SLD and laser diode. While the SLD has a much shorter temporal coherence length than that of the laser diode, its spatial coherence is similarly high, and thus can produce high contrast speckle noise. Effectiveness of the fiber to reduce spatial coherence depends on the number of modes it supports, the relative axial velocity of the individual modes, the fiber length, and the temporal coherence of the entering light. The 25 meter fiber employed in our retina camera generates almost six thousand modes (# of modes=0.5 (π*NA*core diameter/λ)2). The time difference between the leading and trailing modes (i.e., maximum modal dispersion) was calculated to be 1.38 nsec using L*NA 2/(2*ncore*c), where L is the fiber length and c is the speed of light in a vacuum. This difference corresponds to a maximum modal separation in air of 415 mm and a mode spacing of 69 µm. The full width half height (FWHH) temporal coherence length of 18.8 µm (2*ln2*λ 2/π*Δλ) in air for the 679 nm SLD is 3.7 times narrower than the mode spacing and assures that the 6,000 modes incoherently interfere. Assuming uncorrelated modes of equal energy, speckle contrast should reduce by about 77 (i.e., √6,000) times. This estimate does not account for other forms of dispersion or fiber imperfections and bending that would further reduce speckle contrast. Using the same calculations, the 300 meter fiber generated almost 67,000 modes with the time difference between the leading and trailing modes to be 52 nsec. This time difference corresponds to a maximum modal separation in air of 15.7 m and a mode spacing of 234 µm. The FWHH temporal coherence length of 250 µm in air for the 670 nm laser diode is approximately the same as the mode spacing. Again assuming uncorrelated modes of equal energy, speckle contrast should reduce by about 259 (i.e., √67,000) times. These calculations suggest the laser diode with the 300 m fiber will yield lower spatial coherence than the SLD with the 25 m fiber.
2.2 Human subjects
Retinal images were collected on five subjects (19 to 39 years of age) that were free of ocular disease and had normal corrected vision. Spectacle sphere and cylinder, which were obtained by a professional subjective refraction, ranged from -2.0 to -2.25 and 0 to -1.5 diopters, respectively.
The subjects’ line of sight was centered along the optical axis of the retina camera with the aid of a fixation target, bite bar stage, and video camera that monitored the subjects’ pupil in retro-illumination. The fixation target was located at the subjects’ far point and consisted of high contrast cross hairs positioned on a rectilinear grid 0.5 deg apart. The target was back illuminated with uniform red light. As shown in Fig. 1, the target is positioned in the illumination channel and is therefore not viewed through the AO path. A dental impression attached to a sturdy xyz bite bar translation stage stabilized the head and provided accurate pupil positioning. Retro-illumination of the pupil was realized with the 788 nm SLD.
The subjects were cyclopleged and their pupils dilated using two drops of Tropicamide 1% that were administered prior to measurements and one drop every hour thereafter. A single drop of Phenylephrine Hydrochloride 2.5% was also applied at the beginning of the measurements if additional dilation was needed. Data collection on each subject typically lasted one to two hours.
Sphere (defocus) and cylinder (astigmatism) were minimized in terms of the measured wavefront RMS by inserting appropriate trial lenses at the spectacle plane. Residual defocus and astigmatism associated with quantization of spectacle lens power (0.25 D) and subjective criterion for optimum focus in the presence of higher-order aberrations  were corrected with the AO system.
Subjects were requested to refrain from blinking during the video recordings so as not to create instabilities in the wavefront measurement and correction. As an additional safeguard from blinks that did occur, we developed a robust software filter that ignores such wavefront measurements, such that measurements are acquired and stored, but corrections are not performed . The filter was tailored to the specific experimental conditions. Figure 2 illustrates the effectiveness of the filter to suppress instabilities in the RMS wavefront error immediately following an eye blink.
2.3 Retinal imaging at continuous frame rates
Videos were collected of the subjects’ retinas at continuous rates of 10, 30, and 60 Hz. The higher rates were achieved by sacrificing field of view, which was necessary as camera speed was limited by the Quantix readout rate. Specifically 10 Hz (full frame) provided a 1.8 degree field of view; 30 Hz a 0.8 degree; and 60 Hz a 0.4 degree. For all three patch sizes, sampling of the CCD pixels at the retina (1.1 µm/pixel) and illumination patch size (1.8 degree) were held fixed. Exposure times were 1 or 2 msec. To assess the utility of the retina camera for quick focusing and surveying the microscopic retina, videos were collected of the cone mosaic, the retinal vasculature, and through focus of the entire retina thickness. Videos were captured at 1.4 to 2.5 degrees retinal eccentricities. For the through focus videos, focusing was achieved by translating the Quantix CCD parallel to the retina camera’s optical axis.
The number of images per video clip ranged from 100 to 120 and corresponded to a total elapsed time of 2 to 10 sec. Exposure level at the cornea was 3.0 mW during each pulse, and maximum average exposure level at the cornea over the course of a video clip was 0.36 mW for a 1.8 degree field of view. Both levels are more than 22 times below the limits recommended by ANSI .
2.4 Retinal imaging at short-burst frame rates
Acquisition of a finite number of images at very high rates was explored using the kinetics mode of the Quantix CCD, a method which we will refer to as short-burst imaging. In this mode, the CCD array was used as a temporary storage of four (256×530 pixels) and eight (128x530 pixels) images. The illuminated area of the CCD was controlled by a micrometer-adjusted mask (razor blade) that was positioned immediately upstream of the CCD in a conjugate plane. The timing diagram in Fig. 3 illustrates the temporal sequence of the short-burst imaging. Control software was developed that permitted image acquisition rates up to 500 Hz. At this rate, the exposure and delay durations were each 1 msec with four and eight images acquired in 7 and 15 msec, respectively.
To evaluate short-burst imaging, we focused our attention on the dynamic behavior of the capillary bed that defines the edge of the foveal avascular zone. As the exposure and time delay for the image bursts (synched to the strobing light source) are software controlled, it was straightforward to adjust these parameters to optimize viewing of cellular flow in capillaries. To this end, image bursts were collected with exposures ranging from 1 to 4 msec and delays from 1 to 40 msec.
Burst images were obtained without adaptive compensation by flattening the deformable mirror and translating the science camera to the plane that yielded the sharpest image of the capillaries. Burst images were also obtained with adaptive compensation and best focusing of the science camera. To account for fluctuations in accommodation not corrected by the AO, 10 separate image bursts were acquired with one second pauses between bursts (see Fig. 3). From the two sets (with and without AO) of 10 bursts, the sharpest images of the capillaries were selected for the two cases. Blood flow velocity was measured in several of the capillaries by directly mapping the movement of individual cells in the four or eight images.
Contrast of the burst images was slightly adjusted to improve visualization of retinal structure. In addition burst images were translationally registered using an automated cross-correlation method, usually by less than a couple of pixels; this translation stabilized the retina in the short-burst videos and enhanced the visibility of capillary flow. Exposure level at the cornea for the 679 nm SLD was at most 1.4 mW, which is more than 22 and 15 times below the maximum permissible exposure for individual 4 msec pulses and bursts of 4 msec pulses (8 in total) at 200Hz, respectively, as recommended by ANSI .
2.5 Determining optimal exposure duration
Flood illuminated videos of the same patch of cone photoreceptors centered at either 1 or 1.4 degree eccentricity in two subjects were acquired with adaptive compensation and with exposure durations ranging from 1/3 msec to 100 msec. At the start of the experiment, one of the subjects was well acclimated to the instrument and the fixating task. For the other, we observed his accuracy to fixate during imaging improved and eventually plateaued after several trial runs. These initial runs were discarded. The ordering sequence of the eight trials (each with a different exposure duration) was randomized without the subjects’ knowledge. The power level of the 670 nm laser diode was adjusted so that the total energy per exposure duration was held approximately constant, except for the shortest two exposures (1/3 and 1 msec) in which the SLD reached its maximum. Power spectra were computed from the first 20 frames of each cone mosaic video and then averaged to increase signal to noise. Comparison of the normalized average power spectra permitted a straightforward means to quantify the impact of motion blur as a function of exposure duration.
3.1 Wavefront correction
Figure 4 shows specific RMS wavefront errors measured across a 6.8-mm pupil during the retinal imaging experiments, averaged over three subjects, with and without AO correction. The two subjects not included had comparable RMS errors. The error is displayed in terms of the total error (2nd through 10th order aberrations), Zernike defocus (c4), two Zernike astigmatism terms (c3 and c5), and the higher order aberrations (3rd through 10th order aberrations). As indicated in the figure, AO decreased the total RMS error by more than an average factor of 7, with an average residual corrected error of 0.13 µm. It is interesting to note that the majority of residual wavefront error, after correction, is due to contributions of higher order (3rd through 10th) aberrations. For the experiments, the total RMS wavefront error during dynamic correction varied between 0.07 and 0.21 µm.
Figure 4 illustrates the overall effectiveness of the AO system to correct ocular aberrations, but it does not reflect some of the software advancements that go beyond basic AO control and that were utilized for the experiments in this paper. We found these additions helpful for diagnosing AO problems, optimizing system performance, and efficiently operating the system for the eye. These include image sharpening, temporal power spectra analysis, power rejection curves of the closed-loop AO system, time stamping of SHWS measurements, extensive logging, and improving corrector stability. A detailed description of these can be found elsewhere . As described in this citation, the cutoff frequency of the AO system (the frequency at which the rejection curve attains a value of 1) was found to be 0.87 Hz, indicating that aberrations with temporal frequencies below 0.87 Hz are reduced by the system, while those above 0.87 Hz are amplified.
3.2 Effectiveness of the multimode fiber for reducing speckle noise
Figure 5 illustrates the extent to which the multimode fiber attenuates speckle noise in the SLD illumination. Shown are flood-illuminated images of essentially the same patch of cone photoreceptors in one subject’s eye with and without the fiber modification. Both images were collected with adaptive compensation and followed the same experimental protocol. The exit pupil was 6 mm. Images are representative of what we routinely observe in the laboratory. Note the mottled appearance in the left image that is characteristic of speckle and which completely masks photoreceptor information. The right image was obtained with the fiber and reveals a regular pattern of bright spots, which represent light exiting individual cone photoreceptors. Speckle noise is not visually evident indicating a strong reduction of the SLD spatial coherence by the fiber.
As further supportive evidence of the reduction of speckle noise, Fig. 6 shows single one msec snapshots with the camera focused on the cone mosaic for one subject with and without AO correction. Sharpness of the bright spots was optimized by axially translating the science CCD camera. Note the substantial gain in contrast and clarity afforded by correcting the most significant wave aberrations of the eye. Row spacing of the bright spots (=center-to-center spacing * cos30°) ranges from 4.2 to 5.55 µm along a meridian that intercepts the foveal center (which is located below and to the right of the image). This range is consistent with anatomical and psychophysical estimates of cone spacing for this location [29,30]. Note that our conversion from angular to linear retinal size is based on 300 µm/degree. For this particular subject (Fig. 6), a conversion of 299 µm/degree was determined from axial length measurements using ultrasonography . This approach has an expected error of ±10 microns (±3.4%). Speckle is an unlikely source of these bright spots as the spots are of uniform size and arranged in a regular pattern. Furthermore, the theoretical average speckle size (1.22λf/d) for a 6 mm pupil is 2.3 µm, which is noticeably smaller than the spacing of the observed bright spots (4.2 to 5.55 µm). The fact that granular structure of about 2.3 µm is not visually evident is additional evidence of the effectiveness of the multimode fiber to reduce speckle contrast (see fiber description in Section 2). Both uncorrected and corrected images are typical of cone images obtained using conventional incoherent light sources (flash and arc lamps) [2–5,7,8].
3.3 Retinal imaging at continuous frame rates
Figure 7 shows a 10 Hz video acquired before and during adaptive compensation. The video is of a fixating eye with the science camera focused to visually maximize sharpness of cone photoreceptor cells. After AO compensation begins, which takes about 4 frames to reach full compensation, punctuated bright spots a few microns in diameter are observed across the entire 1.8° field of view and correspond to individual cone cells. A number of the cones appear noticeably brighter and are relatively easy to follow through the video. Darkened hazy sub-regions (usually a few cones in diameter) are likely the shadow patterns of small retinal vessels projected onto the underlying mosaic. The video qualitatively illustrates the improvement in resolution and contrast that accrues when the AO system compensates for the aberrations of the eye including residual defocus and astigmatism. Note however that even at 10 Hz the video appears somewhat choppy owing to involuntary lateral movement of the retina between consecutive frames even though the subject was fixating at a single location.
The three AO compensated 30 Hz videos in Fig. 8 show the cone mosaic, retinal vasculature, and a through focus of the entire retina thickness. Illumination for these was provided by the 670 nm laser diode after passing through the 300 m multimode fiber. In the left video, individual cone cells fill the entire 0.8° field of view. The higher imaging rate (30 Hz compared to 10 Hz) clearly improves visual continuity of the cone mosaic, which is moving between frames. Note the retinal capillaries that traverse diagonally through the top one third of the frame. The capillaries are out of focus and largely transparent, but are made apparent by the cellular motion within them. This motion is not captured by individual frames and illustrates a clear advantage of video rate imaging.
This advantage is better illustrated by the center video in Fig. 8 that shows the camera focused on a network of retinal vessels and capillaries located in the inner retina. Two large vessels roughly 14.5 µm in diameter extend downward and branch into a complex web of small capillaries that are ~5 µm in diameter and typical of the smallest vessels in the retina. Many of the capillaries are well defined, albeit of low contrast due to their high transparency at the 670 nm illumination wavelength. In general, the capillaries are much easier to detect in the video than in individual still frames. Instrument resolution and sensitivity with AO is found sufficient to reveal the dynamic behavior of highly structured reflections from within capillaries, being approximately on the same scale size as individual blood cells (leukocytes and erythrocytes) and other microscopic blood constituents (platelets, plasma). Here we have assumed that the detected reflection is from within the vessel rather than at the exterior vessel surface. If our interpretation is correct, this may permit a direct means of non-invasively identifying blood composition at the cellular level in the smallest retinal vessels.
The right video in Fig. 8 depicts a through focus of the retina that was acquired at increasing depths, starting at the retinal vasculature and ending at the photoreceptor layer. In an attempt to remove the motion artifacts (which were comparable in magnitude to that present in the first two videos of Fig. 8), adjacent frames in the through focus video were registered using a cross correlation algorithm. Because of the significant variation in retinal information across the video, registration did not reach pixel accuracy, however, as can be seen in the video most retina motion was effectively removed. This registration was found to permit easier visual analysis of the video. Well defined capillaries are visible at the beginning of the video. The underlying cone photoreceptors are well out of focus. As the video progresses, the capillaries transform into faint shadows that project onto the underlying cone mosaic in the last few frames. The ability of these largely transparent vessels to generate shadows suggests a form of scintillation caused by a change in refractive index either within the vessel or between the vessel and surrounding neural tissue. Most of the vasculature observed in the video are small capillaries approximately 5–7 µm in diameter.
It is interesting to note the dynamic behavior of the bright single cone in the middle of the frame that lies directly behind a large capillary. The intensity of this cone appears to fluctuate rapidly with the capillaries in focus, presumably caused by cellular flow in the capillary, but then becomes reasonably stable with the cone in focus. This suggests that for cone imaging experiments that rely on accurate reflectance measurements special care should be taken to account for the optical impact of overlying capillaries. Also note that capillaries are observed at different depths in the retina with some vessels at best focus in the first few frames and others at best focus later in the video. This general observation of a stratified vasculature parallels histology  and recent results with optical coherence tomography [33,14]. AO flood illumination might represent a potentially straightforward and non-invasive approach to quickly map the retinal capillary network in three dimensions.
The AO compensated 60 Hz videos in Fig. 9 show small 0.4° patches of the cone mosaic and retinal vasculature. Illumination was again provided by the 670 nm laser diode and multimode fiber. Signal to noise in the 30 (Fig. 8) and 60 Hz videos is essentially identical as the retinal illumination, exposure duration, and CCD read noise remained unchanged for the two cases. In the left video, individual cone cells fill the entire 0.4° field of view. In comparison to the corresponding 30 Hz video in Fig. 8, cone clarity in this faster video is actually higher, though this must be attributed to better AO compensation as all other aspects of the two imaging experiments were essentially identical. The middle and right videos are identical with the right running four times slower to permit better visualization of the cellular blood flow. The video shows a small portion of a large vessel lying adjacent to a small one, both of which traverse diagonally through the frame. Even at the 60 Hz frame rate, the direction of flow in the vessels is not readily obvious even though the flow itself is quite apparent. We do not believe this confusion is due to temporal aliasing. While the frame size is small, being only 120 µm wide, a blood cell traveling at 1.5 mm/sec (a typical velocity) will move only 25 µm per frame and therefore require almost 5 frames to traverse the full frame width. For reasons we do not fully understand, punctuated bright reflections within the vessels (which should correspond to individual cells) could not be tracked across more than two or three frames. Perhaps the reflective properties of the cells change as the cells propagate and re-orient themselves within the confines of the vessels. In support of this, some frames contain highly punctuated and bright reflections (e.g., #50 and #81) that essentially vanish in the adjacent frames, i.e., within ±16.7 msec.
In the previous cone videos (particularly the 60 Hz video in Fig. 9), the reflectance of some isolated cones fluctuate, in some cases disappearing and then reappearing within a second or two. We have observed that the rapidity and acuteness of the fluctuations and number of cones that fluctuate in a given patch of retina depend on the temporal coherence of the light source and the bleached state of the retina, with fully regenerated cones (dark adapted eye) providing the most variation.
Figure 10 is an example of the latter. The video shows a patch of cone photoreceptors in a dark adapted eye. Intensity of the 670 nm flashes during the video is sufficient to fully bleach both the medium and long wavelength cones in roughly the first second. Fluctuations are clearly much stronger and widespread than in the previous videos. Interestingly, the cones fluctuate at distinctly different temporal periods, with some highly stable throughout the video (no fluctuations) and others oscillating quickly at the beginning, and then gradually slowing and stabilizing as the video progresses. To more easily visualize these temporal differences, the right side of Fig. 10 shows three extracted cones that are representative of the three general groups of oscillations in the video. The cone in the red box oscillates quickly, going through 1.5 oscillations in the first second (30 frames) with no appreciable oscillations afterwards. The cone in the green box oscillates more slowly; the cone in the blue box does not oscillate at all, i.e., it appears stable. There is suggestive evidence that these three groups correspond to the three spectral subtypes of cone photoreceptors, which bleach at different rates when illuminated with 670 nm light.
While dynamic fluctuations in the reflectance of single cones have already been reported, these have occurred over periods of minutes to many hours in the living eye [34,35]. The fluctuations observed here are on a surprisingly much shorter time scale (seconds), which were easily revealed by the high speed acquisition of our retina camera. One possible explanation, which we are exploring , is that the light exiting individual cone photoreceptors originates from reflections both in front and behind the outer segments. Interference of these reflections coupled with a change in the optical path length of the outer segment (due to photopigment bleaching by the imaging light source) could in principle generate temporally fast oscillations. We are developing an optical model of the photoreceptors that will explain these observations.
3.4 Retinal imaging at short- burst frame rates
A range of exposure durations (1 to 4 msec) was examined with the shortest duration (1 msec) found to provide the visually sharpest images of individual blood cells flowing through retinal capillaries. In fact a noticeable reduction in cell sharpness was observed for the next longer exposure of two msec. Delay between consecutive exposures (1 to 40 msec) was also examined with larger delays found to increase the difficulty in visually identifying and tracking the same blood cells. The shortest delay of one msec provided the most reliable tracking.
Figure 11 shows two representative four-burst videos acquired at 500 Hz of the same retinal capillaries in one subject with and without adaptive compensation. The videos represent the best from a set of ten each, although many of the compensated videos were of similar quality. The videos are best viewed in loop mode in which the video automatically cycles. In the uncompensated video (Fig. 11, top), the capillaries appear highly blurred due to the relatively poor image quality, however, gross blood motion (flow is from left to right in the video) can still be detected. Correction of the most significant aberrations in the eye (Fig. 11 bottom) reveals a complex capillary network with two prominent capillaries (the lower of the two has a diameter of 6.6 µm) stretching from the bottom left to the right and a large dark vessel located in the upper right corner. Notice that within any single (still) frame, it is difficult to discern without ambiguity the presence of additional vasculature structures. However, when the video executes, the extent of the vasculature network is immediately apparent. Note the complex web of small capillaries in the left half of the video and the lack thereof in the right half. The vessels are characterized by a flowing stream of light and dark reflections that are suggestive of both punctuated and extended micro-structures in the vessels. The extended structures may represent chains of erythrocytes or plasma gaps. Unlike the previous 30 and 60 Hz videos, the 500 Hz video clearly reveals the direction of blood flow for each capillary as well as the much larger vessel in the upper right. Blood velocity in the capillaries was directly measured at 1.5 mm/sec, which falls within the normal range as measured by other techniques, for example acridine orange staining of leukocytes in cynomolgus monkeys  and adaptive optics scanning laser ophthalmoscopy in humans . The overall frame brightness increases with each successive frame and is likely due to bleaching of the underlying photoreceptor pigment. In general, the short burst images contain sufficient temporal bandwidth, sensitivity and optical resolution to clearly capture high temporal dynamics of the microscopic retina, such as the motion of single blood cells in individual capillaries.
Figure 12 demonstrates an eight-burst 500 Hz video on another subject. The AO correction is not as good as in the other videos, but illustrates that longer burst trains are readily possible at the expense of a reduced vertical field of view, in this case by a factor of two relative to the four-burst videos shown in Fig. 11.
Leukocyte flow in parafoveal capillaries has already been reported using a 30 Hz SLO equipped with AO and a 660 nm light source . While the flood illumination camera described here also uses AO and operates at a similar wavelength, the appearance of the videos from the two instruments are strikingly different. Capillary clarity and cellular contents appear better defined, and the underlying bright cone mosaic less distracting with flood illumination. Flood illumination also reveals cellular structure across the full extent of the capillaries (e.g. Fig. 11), thus permitting velocity measurements of many different cells in the same capillary.
3.5 Determining optimal exposure duration
Flood illuminated videos of the same patch of cone photoreceptors were acquired on two subjects with adaptive compensation and with exposure durations ranging from 1/3 msec to 100 msec. Illumination for these was provided by the 670 nm laser diode after passing through the 300 m multimode fiber. As representative examples of the acquired raw data, Fig. 13 displays a collage of four cone videos with exposure durations of 1, 10, 33, and 100 msec. As indicated in the figure, the frame rate varied between videos owing to the different exposure durations. From visual inspection, 1 and 10 msec exposures (left two videos) capture the fine granular structure of the cone mosaic in almost every frame. This suggests that blur from involuntary retina motion is minimal across these short exposures. For the longer 50 µm exposures of 33 msec (middle right) and 100 msec (rightmost), cone quality is significantly worse and in many frames is essentially lost suggesting that retina motion blur is the limiting factor. Interestingly however even at 100 msec, some cone structure can be observed in frames that coincide with what appears to be the end of a retinal movement, at which point the retina momentarily comes to rest. Even in these frames though, cone quality does not reach that found in the 1/3 msec and 10 msec videos.
As a means to quantify the impact of motion blur on image quality, Fig. 14 (left) shows the computed average power spectra for one of the two subjects at each of the eight exposures durations examined. Results from the other subject were similar. Power spectra are normalized to one at zero frequency and plotted using a log ordinate. While cone photoreceptors are visible in the Fig. 13 videos, no corresponding cusp in the power spectra (Fig. 14) is readily evident. This is likely attributable to the sampling of different retinal patches due to eye motion during the video. As shown in the figure, spectra for exposure durations from 1/3 to 10 msec contain approximately similar energy at frequencies up to about 100 cycles/deg. Above this the shorter 1/3 and 1 msec curves deviate towards higher noise floors that are attributed to their lower optical energy per flash. In principle the 1/3 msec exposure should be least corrupted by motion blur and therefore provide the most energy at any given spatial frequency. As shown in Fig. 14, however, other exposures (1, 4, and 10 msec) sometimes yield slightly more energy. We attribute this to differences in the quality of the AO correction, the refractive state of the eye, and the subjects’ ability to fixate. These sources of noise reflect the accuracy of the experiment. Nevertheless, Fig. 14 (left) reveals a clear trend in which longer exposures (>10 msec) lead to lower energy, in particular between 20 and 90 cycles/deg. Power is noticeably lowest for the longest exposure of 100 msec. Exposures less than 10 msec produce similar power curves suggesting movement of the retina is either negligible for these short times or at least within the error of the experiment.
To further quantify the impact of exposure duration, Fig. 14 (right) shows the ratio of the 4 msec power curve to each of the others. The ratios are averaged across both subjects. The 4 msec curve was chosen as it represents the shortest exposure at which the total energy per flash was not compromised by flash duration. As before, exposures up through 10 msec produce similar ratios, which again suggest that eye motion is negligible over these exposures. Exposures greater than 10 msec show a monotonic increase in the ratio and illustrate a clear benefit of short exposures (<10 msec). Power of the 4 msec exposure is more than a factor of three higher across a broad range of frequencies compared to that of the longest exposure (100 msec). As expected, the 4 msec exposure shows smaller gains when compared to the other exposures (20, 33, and 66 msec) with the gains localized to higher frequencies (between 70 and 80 cycles/deg). For comparison, the fundamental spatial frequency of the cone photoreceptors in the patches of retina imaged ranged from 60 to 75 cycles/degree.
These preliminary findings on two normal subjects suggest that AO flood-illumination retina cameras should operate with an exposure duration less than about 10 msec so as to minimize motion blur. This should be considered an upper threshold as the necessary exposure depends strongly on the subject’s ability to fixate, which can be noticeably worse in the aged or diseased eye.
Current adaptive optics flood-illumination retina cameras operate at low frame rates and employ bulky and expensive arc and flash lamp sources. As an alternative, we investigate a novel AO flood-illumination camera that achieves significantly higher frame rates using a current-modulated SLD and multimode laser diode coupled to a stretch of multimode optical fiber. Retinal images were successfully collected with exposure durations as short as 1/3 msec. Continuous frame rates of 10, 30, and 60 Hz were achieved for imaging 1.8, 0.8, and 0.4 deg retinal patches, respectively. In all cases, readout speed of the CCD limited frame rate. Through focus of the retina with the camera clearly revealed cellular details at both the photoreceptor and retinal vasculature layers. Little could be observed in the highly transparent neural tissue lying in between, illustrating a significant weakness of flood illumination compared to optical sectioning instruments such as OCT. Short-burst imaging up to 500 Hz was also achieved by temporarily storing sequences of images on the CCD. The 60 Hz (continuous) and 500 Hz (burst) frame rates are 9 and 75 times higher than the 6.7 Hz upper limit  of current AO flood-illumination retina cameras.
The combination of high resolution (realized with AO) and high frame rates provides the ability to quickly navigate through the retina, recognize individual cells of relative high contrast without image warp and motion blur, and monitor retinal dynamics occurring at the cellular level (e.g., capillary blood flow). High speed AO flood illumination represents a potentially straightforward and non-invasive approach to quickly map the retinal capillary network in three dimensions. As a demonstration of the value of the camera for imaging over a broad range of exposure durations (1/3 to 100 msec), we evaluated the tradeoff between exposure duration and image blur associated with retina motion. Finally, the high speed of the camera permitted observation of surprisingly rapid fluctuation in the reflectance of single cones when the eye was dark adapted. The oscillatory behavior of these cones may depend on cone class, which will require further investigation.
The authors thank William Monette and Daniel Jackson’s group for electronics and machining support. A special thanks goes to Edgar Alvarez for early work on the power spectra analysis. Financial support was provided by the National Eye Institute grant 5R01 EY014743. This work was also supported in part by the National Science Foundation Science and Technology Center for Adaptive Optics, managed by the University of California at Santa Cruz under cooperative agreement No. AST-9876783.
References and links
1. R. K. Tyson, Principles of Adaptive Optics (Academic Press, New York, 1998).
2. J. Liang, D. R. Williams, and D. T. Miller, “Supernormal vision and high resolution retinal imaging through adaptive optics,” J. Opt. Soc. Am. A 14, 2884–2892 (1997). [CrossRef]
3. H. Hofer, L. Chen, G. Y. Yoon, B. Singer, Y. Yamauchi, and D. R. Williams, “Improvement in retinal image quality with dynamic correction of the eye’s aberrations,” Opt. Express 8, 631–643 (2001). [CrossRef] [PubMed]
4. N. Ling, Y. Zhang, X. Rao, X. Li, C. Wang, Y. Hu, and W. Jiang, “Small table-top adaptive optical systems for human retinal imaging”, in High-Resolution Wavefront Control: Methods, Devices, and Applications IV, J. D. Gonglewski, M. A. Vorontsov, M. T. Gruneisen, S. R. Restaino, and R. K. Tyson, eds., Proc. SPIE4825, 99–108 (2002). [CrossRef]
5. A. V. Larichev, P. V. Ivanov, N. G. Iroshnikov, V. I. Shmalhauzen, and L. J. Otten, “Adaptive system for eye-fundus imaging,” Quantum Electron. 32, 902–908, 2002. [CrossRef]
7. M. Glanc, E. Gendron, F. Lacombe, D. Lafaille, J.-F. Le Gargasson, and P. Léna, “Towards wide-field retinal imaging with adaptive optics,” Opt. Commun. 230, 225–238 (2004). [CrossRef]
8. P. Fournier, G. R. G. Erry, L. J. Otten, A. Larichev, and N. Irochnikov, “Next generation high resolution adaptive optics fundus imager,” in 5th International Workshop on Adaptive Optics for Industry and Medicine, edited by Wenhan Jiang, Proceedings of SPIE Vol. 6018 (SPIE, Bellingham, WA, 2005).
10. A. Roorda, F. Romero-Borja, W. J. Donnelly, H. Queener, T. J. Hebert, and M. C. W. Campbell, “Adaptive optics scanning laser ophthalmoscopy,” Opt. Express 10, 405–412 (2002). [PubMed]
11. D. T. Miller, J. Qu, R. S. Jonnal, and K. Thorn, “Coherence gating and adaptive optics in the eye,” in Coherence Domain Optical Methods and Optical Coherence Tomography in Biomedicine VII, V. V. Tuchin, J. A. Izatt, and J. G. Fujimoto, eds., Proc. SPIE4956, 65–72 (2003). [CrossRef]
12. B. Hermann, E. J. Fernández, A. Unterhuber, H. Sattmann, A. F. Fercher, W. Drexler, P. M. Prieto, and P. Artal, “Adaptive-optics ultrahigh-resolution optical coherence tomography,” Opt. Lett. 29, 2142–2144 (2004). [CrossRef] [PubMed]
13. Y. Zhang, J. Rha, R. Jonnal, and D. Miller, “Adaptive optics parallel spectral domain optical coherence tomography for imaging the living retina,” Opt. Express 13, 4792–4811 (2005). [CrossRef] [PubMed]
14. R. Zawadzki, S. Jones, S. Olivier, M. Zhao, B. Bower, J. Izatt, S. Choi, S. Laut, and J. Werner, “Adaptive optics optical coherence tomography for high-resolution and high-speed 3D retinal in vivo imaging,” Opt. Express 13, 8532–8546 (2005). [CrossRef] [PubMed]
15. E. J. Fernández, B. Považay, B. Hermann, A. Unterhuber, H. Sattmann, P. M. Prieto, R. Leitgeb, P. Ahnelt, P. Artal, and W. Drexler W, “Three-dimensional adaptive optics ultrahigh-resolution optical coherence tomography using a liquid crystal spatial light modulator,” Vision Res. 45, 3432–3444 (2005). [CrossRef] [PubMed]
17. I. Iglesias and P. Artal, “High-resolution retinal images obtained by deconvolution from wave-front sensing,” Opt. Lett. 25, 1804–1806 (2000). [CrossRef]
18. K. E. Thorn, J. Qu, R. J. Jonnal, and D. T. Miller, “Adaptive optics flood-illuminated camera for high speed retinal imaging,” Invest. Ophthalmol. Visual Sci. 44, E-Abstract 1002 (2003).
19. J. Rha, R. S. Jonnal, Y. Zhang, and D. T. Miller, “Rapid fluctuation in the reflectance of single cones and its dependence on photopigment bleaching,” Invest. Ophthalmol. Visual Sci. 46, E-Abstract 3546 (2005).
20. K. E. Thorn, R. S. Jonnal, J. Qu, and D. T. Miller, “High-speed imaging of the retinal microvasculature with adaptive optics,” Society of Photo-Optical Instrumentation Engineers’ 2004 International Symposium on Ophthalmic Technologies XIV, San Jose, CA, January 24–25, 2004.
21. J. Rha, R. S. Jonnal, Y. Zhang, and D. T. Miller, “Video rate imaging with a conventional flood illuminated adaptive optics retin,a camera,” 88th Optical Society of America Annual Meeting, Rochester, New York, October 10–14, 2004.
22. ANSI, American National Standard for the Safe Use of Lasers, ANSI Z136.1 (Laser Institute of America, Orlando, FL, 2000).
23. B. Crosignani, B. Diano, and P. Di Porto, “Speckle-pattern visibility of light transmitted through a multimode optical fiber,” J. Opt. Soc. Am. 66, 1312–1313 (1976). [CrossRef]
24. B. Dingel and S. Kawata, “Laser-diode microscope with fiber illumination,” Opt. Commun. 93, 27–32 (1992). [CrossRef]
25. E. G. Rawson, J. W. Goodman, and R. E. Norton, “Frequency dependence of modal noise in multimode optical fibers,” J. Opt. Soc. Am. 70, 968–976 (1980). [CrossRef]
26. F. M. Mims III, A Practical Introduction to Lightwave Communications (Howard W. Sams & Co., Indiana, 1982).
27. L. N. Thibos, X. Hong, A. Bradley, and R. A. Applegate, “Accuracy and precision of methods to predict the results of subjective refraction from monochromatic wavefront aberration maps,” J. Vis. 4, 329–351 (2004). [PubMed]
28. Y. Zhang, J. Rha, R. S. Jonnal, and D. T. Miller, “Indiana University AO-OCT system,” in Adaptive Optics for Vision Science: Principles, Practices, Design, and Applications, J. Porter, et al., eds. (John Wiley & Sons, New Jersey, In Press).
31. A. G. Bennett, A. R. Rudnicka, and D. F. Edgar, “Improvements on Littmann’s method of determining the size of retinal features by fundus photography,” Graefes Arch. Clin. Exp. Ophthalmol. , 232, 361–367 (1994). [CrossRef] [PubMed]
32. M. Iwasaki and H. Inomata, “Relation between superficial capillaries and foveal structures in the human retina,” Invest. Ophthalmol. Visual Sci. 27, 1698–1705 (1986).
33. B. Cense, N. A. Nassif, T. C. Chen, M. C. Pierce, S. Yun, B. H. Park, B. E. Bouma, G. J. Tearney, and J. F. de Boer, “Ultrahigh-resolution high-speed retinal imaging using spectral-domain optical coherence tomography,” Opt. Express 12, 2435–2447 (2004). [CrossRef] [PubMed]
34. A. R. Wade and F. W. Fitzke, “In-vivo imaging of the human cone photoreceptor mosaic using a confocal LSO”, Lasers Light Ophthalmol. 8, 129–136 (1998).
35. A. Pallikaris, D. R. Williams, and H. Hofer, “The Reflectance of Single Cones in the Living Human Eye”, Invest. Ophthalmol. Visual Sci. 44, 4580–4592 (2003). [CrossRef]
36. H. Nishiwaki, Y. Ogura, H. Kimura, J. Kiryu, and Y. Honda, “Quantitative evaluation of leukocyte dynamics in retinal microcirculation,” Invest. Ophthalmol. Vis. Sci. 36, 123–130 (1995). [PubMed]