We have developed a compact retinal imager that integrates adaptive optics (AO) into a line scanning ophthalmoscope (LSO). The bench-top AO-LSO instrument significantly reduces the size, complexity, and cost of research AO scanning laser ophthalmoscopes (AOSLOs), for the purpose of moving adaptive optics imaging more rapidly into routine clinical use. The AO-LSO produces high resolution retinal images with only one moving part and a significantly reduced instrument footprint and number of optical components. The AO-LSO has a moderate field of view (5.5 deg), which allows montages of the macula or other targets to be obtained more quickly and efficiently. In a preliminary human subjects investigation, photoreceptors could be resolved and counted within ~0.5 mm of the fovea. Photoreceptor counts matched closely to previously reported histology. The capillaries surrounding the foveal avascular zone could be resolved, as well as cells flowing within them. Individual nerve fiber bundles could be resolved, especially near the optic nerve head, as well as other structures such as the lamina cribrosa. In addition to instrument design, fabrication, and testing, software algorithms were developed for automated image registration and cone counting.
© 2009 Optical Society of America
Adaptive optics (AO) has recently achieved success in a range of applications in ophthalmology. It has been integrated into flood illumination full-field retinal cameras, confocal scanning laser ophthalmoscopes (SLO), and optical coherence tomography (OCT) instruments for high resolution reflectance imaging [1–10]. AO is used as a tool to study the structural and functional aspects of vision, image elegant and complex retinal circuitry, and the deterioration of such structures and processes during the progression of disease. It has been used in direct measurements of the foveal avascular zone, retinal capillary erythrocyte and leukocyte velocity, pulsatility, and other functional dynamics [11, 12]. AO stimulation is also being used to measure and begin to understand the characteristics and causes of intrinsic retinal signals [13, 14]. It has been used to study the organization of the human trichromatic cone mosaic and defects in color vision compared to standard psychophysical tests [15–17]. The waveguiding properties and aspects of cone directionality (i.e., Stiles-Crawford effect) have also been confirmed with measurements using adaptive optics instruments . AO-corrected microperimetry to probe fixation loci and retinal microscotomas have also been recently reported [19, 20]. The RPE cell mosaic is being mapped in monkeys and humans and correlated to the cone mosaic [21, 22]. Ganglion cell function has been explored with exogenous dyes [21, 23]. In many of the aforementioned studies, genetic defects are confirmed and explored directly in the live eye . Diseases such as rod-cone dystrophy are also being examined [25–27]. Systems are slowly migrating from the research lab into the clinic for use on patients with a variety of diseases and conditions [28–30]. AO is also being applied to advanced molecular and gene therapies, both in their development and as the primary method to determine treatment efficacy at the cellular level.
AO systems use a wave-front sensor to sense ocular aberrations that arise primarily from the cornea, crystalline lens, and tear film, and correct them using a wavefront compensator in a closed-loop manner. A Hartmann-Shack wavefront sensor (HS-WS) consisting of a lenslet array and CCD camera are typically used for rapid detection of ocular aberrations , but other techniques such as interferometry can also be used. Likewise, there are several methods to achieve wavefront correction including MEMS-based deformable mirrors (DM) , magnetic actuator DMs , and liquid crystal phase modulators [34, 35].
There are currently three barriers to widespread use of AO by clinicians and research scientists. The first obstacle is the high cost of deformable mirrors (DM) and liquid crystal phase modulators. The second barrier is system complexity. Currently, AO systems can be built and operated only by researchers with extensive expertise in optics, engineering, and instrumentation; AO systems have been designed and constructed for the best imaging performance possible, to the exclusion of all other factors (size, cost, complexity, ease-of-use, etc.). This has slowed the transition of AO from the research lab to the clinic. Lastly, compelling applications of this powerful imaging technology essential to future standard of eye care have yet to be integrated into clinical practice.
These barriers must be addressed together. The high DM cost will presumably decrease with on-going technological innovation as more companies become involved in system manufacturing and foster AO commercialization. With reports of significant performance improvement of AO methods and systems for a growing number of ophthalmic applications, demand for a commercial AO instrument will increase.
In order to reduce the complexity of AO, we have designed, developed, and tested a compact adaptive optics line-scanning ophthalmoscope (AO-LSO). The goal is to build an instrument that nearly anyone can operate in order to extend the perspective and optical access that AO provides to as many research and clinical applications as possible. The instrument is designed to fill the niche between SLO instruments that provide wide field but limited resolution for routine clinical use and the complex, high resolution, high-performance AO instruments that are currently confined to research laboratories. A compact, simplified AO instrument like the AO-LSO that can be used by ophthalmologists, optometrists, and vision scientists will facilitate the introduction of this technology and the development of new techniques to detect and treat retinal diseases.
2.1 Optical setup
The fundamental principle behind any confocal imaging system (e.g., scanning laser ophthalmoscope, confocal microscope, etc.) is the rejection of light scattered from adjacent or out-of-plane voxels by use of a pinhole in front of the detector conjugate to the focal plane. This significantly improves image contrast and resolution over flood illumination and detection schemes. The line scanning approach multiplexes the illumination and detection in one dimension, while rejecting scattered light outside of the confocal range gate defined by the pixel size in a similar fashion as a flying-spot SLO. We have previous shown that the LSO retains a substantial portion of the confocality of a flying spot SLO with several key advantages . The line scanning approach reduces the hardware complexity by elimination of a high-speed scanner and associated mirror or lens relays.
Figures 1 shows the optical schematic of the AO-LSO. All optics fit easily on an 18×24 inch (45×60 cm) optical breadboard. The system uses two optical relays each consisting of two spherical mirrors to transfer the pupil to a large-stroke deformable mirror (DM, Mirao, Imagine Eyes Inc.) and to a single imaging galvanometer. The galvanometer pivots about the pupil to scan the line beam, created by a cylindrical lens, across a small patch of retina. A fifth spherical mirror is the final ocular imaging element. Spherical mirrors are preferred because they produce no back-reflections and to reduce overall system aberration. The angles between the spherical mirrors are kept small to minimize astigmatism. The magnification of the relays is constrained in the design by the DM diameter (~15 mm), the linear array detector (LAD, e2v Inc.) length (10.24 mm), and the desired input beam diameter at the pupil. The input beam diameter was set to 7.5 mm to achieve high lateral resolution for a dilated pupil with minimization of ocular aberrations. The magnification from the LAD to the retina is 6.4 and so the field size on the retina is 1.6 mm (5.5 deg for an emmetropic eye) and the pixel size of the retinal images is 1.56 µm. Zemax modeling of the optical system indicates that diffraction-limited performance within an Airy disc radius of ~2.2 µm (at 820 nm) can be achieved in a paraxial eye with aberration correction using only a small fraction of the DM dynamic range. Residual off-axis astigmatism is corrected by the DM.
The overall pathlength is fairly long (~2.6 m from illumination fiber to eye); however, the number of optical elements is small, and with the high stroke DM the system is amenable to reduced focal lengths and folding for increased compactness. Instrument focusing (de-focus correction) is performed with the DM in addition to wavefront correction.
A dual superluminescent diode (SLD) source with a bandwidth of ~65 nm and a center wavelength of ~820 nm (Broadlighter, Superlum Inc.) is used for illumination to eliminate coherent speckle. The illumination beam is spread in one axis with a cylindrical lens and transferred via the relays to the DM and the galvanometer, where it is scanned in the other axis to produce a raster on the retina. The raster light returned from the retina is de-scanned with the galvanometer and the fixed line is passed back through the system and focused on a LAD to produce retinal images.
A separate beacon (735 nm laser diode, LD) is used to sense aberrations. The beacon is introduced into the setup with a pellicle beamsplitter (PBS) and the return light is separated with a dichroic beamsplitter (D2). The pellicle is used to pick off only a small amount (~8%) of the LD light available at the input and to maximize the return signal. The dichroic beamsplitter is designed for efficient separation of imaging and beacon light (>95% transmission at 735 nm and >90% reflection at 820 nm). The main disadvantage of a separate beacon is the need for another light source and a dichroic to separate the two paths. However, the flexibility in positioning the beacon within the AOLSO raster allows AO correction to be optimized for particular image regions. The beacon is parfocal with the imaging beam and positioned near the center of the line. An additional lens relay is used to de-magnify the returning beacon light to 7.5 mm on the HS-WS camera (Uniq Inc.). The iris blocks light from corneal reflections from corrupting the HS-WS images. The HS-WS and LAD are synchronized to the same frame-sync signal to ensure aberration correction of the same retinal patch. Two flip-mounted mirrors are used for HS-WS calibration using the imaging beam. With low system aberrations, this calibration could be performed at other retinal conjugates and thus is not absolutely required for operation.
A wax impression bite bar and forehead rest were used for head stabilization and a system of translation stages for pupil centration. The subject’s fixation was coarsely controlled with a computer-controlled 8×8 LED array fixation target. A cold mirror is used to transmit imaging and beacon beams while reflecting the fixation target into the eye.
The system was initially tested in a limited number of human subjects without retinal disease. The measurements were performed under a New England IRB approved protocol. The incident power at the cornea for the beacon and for the extended imaging beam was 50 µW and 650 µW, respectively. This exposure level is safe according to ANSI standards  for 8 hours continuous intrabeam viewing (without adaptive optics or retinal tracking, not explicitly discussed by ANSI standards). In our case, line illumination would allow for even higher safe exposure levels. For the 5.5 deg line illumination (without scanning), extended source calculation increases the maximum permissible exposure (MPE) by a factor of 8 (square root of the factor for square illumination ) as compared to intrabeam viewing. AO correction increases the power density by reducing only the line width by a factor of 5 (see below the reduction of the Strehl ratio). Therefore, extended line illumination more than compensates for the power density increase associated with AO correction. We note here that if the scanner should fail, the patient could be exposed to the stationary laser line formed by a fixed cylindrical lens. The system can never focus to a point; therefore the LSO is inherently safer than an SLO.
Ten subjects (7 male, 3 female) with an average age of 47 years were imaged. 3 subjects were myopic, 4 subjects were emmetropic (visual acuity 20/40 or better), 3 subjects were hyperopic, and 6 subjects were presbyopic. The average refractive error was approximately +1.5 D for the hyperopes and -5.0 D for the myopes. No subject’s eyes were dilated during the investigation. However, most subjects had at least 4–5 mm diameter pupils, which was sufficient for lateral and depth resolution improvement with AO compensation. The room was darkened and a baffle was used to provide dark-adapted pupil dilation. Some subjects had 7–8 mm pupils when not looking directly at the NIR beacon and illumination beams. The pupils always constricted when imaging the fovea. Irregular pupil dilation accounted for some of the variability seen in resultant imaging performance and also accounted for the relatively smaller differential in AO correction (as measured by Δ[wavefront error]RMS) for images collected from the fovea.
The central macula was the image target in most individuals, although in some subjects, the optic nerve head (ONH) was also imaged. Photoreceptors could be resolved in all subjects except two, one of which had a mild cataract. The deformable mirror was able to statically correct de-focus (prior to AO compensation) in all individuals. LSO and HS-WS videos with durations of 5–30 seconds were acquired.
3.1 AO performance
Adaptive optics dynamically corrected the input wavefront for ocular aberrations and improved the lateral resolution in LSO images collected from human subjects. Figure 2 shows AO-LSO images (and video) of the fovea with and without AO correction with the best focus. Photoreceptors can be resolved within ~1 deg. of the fovea. Figure 3 shows the measured RMS wavefront error for one subject (8-mm pupil). The average RMS error (Strehl ratio) improved from 0.34 µm (2.2×10-4) to 0.07 µm (0.70) with AO correction. All Zernike orders up to 7 showed improvement, although the most improvement was seen in the first four orders. The 10–90 % risetime for the correction was 0.52 s, slower than that achieved previously for other deformable mirrors . The overall correction time is determined by several factors including the DM response, the wavefront sensor acquisition frame rate – 15 fps in our case — and the relaxation rate (fraction of the total correction made per wavefront) which requires about 5 frames with our algorithm . Other AO performance metrics (e.g., typical wavefront error maps) are shown elsewhere .
3.2 Imaging performance
Photoreceptors could be resolved in 80 % of the individuals imaged in this small pilot proof-of-principle human subject investigation. Photoreceptors could be resolved within ~0.3 mm (~1.0 deg.) of the fovea for some individuals. According to previously reported histology , the cone size at this eccentricity is roughly equivalent to the size of 3 pixels for the AO-LSO instrument. Photoreceptors could be reliably counted (manually) within ~0.5–1 mm (~1.5–3 deg.) for all subjects where they could be resolved. Thus, photoreceptors could be counted across roughly 80% of the surface area of the macula (central 5.5 mm). Figure 4 shows an AO-LSO image and video of the photoreceptor mosaic for one subject.
The photoreceptor mosaic was examined more thoroughly in 4 of 10 subjects. Three of these subjects were emmetropes and one had relatively large astigmatism (OS: -1.75D/065, OD: -2.5D/098). For these subjects, videos were acquired for generation of montages across the entire macula (central 15 deg.). Additional analysis included manual and automated cone counting and frequency analysis as a function of eccentricity. Figure 5 shows examples of the photoreceptor mosaic imaged at various eccentricities for these four subjects (cropped magnified regions). Note that subject (a) appear to have a higher density than subject (b) at roughly equivalent eccentricities. For good fixators (Fig. 5a), very little distortion in individual photoreceptors is seen. For other subjects, intraframe distortions from microsaccades caused the cones to appear elongated (Fig. 5c) along the direction of motion.
In several individuals, short video clips (~5 sec.) were collected while different fixation target LEDs were illuminated to move fixation to nine offset positions with respect to the imaging raster. These were used to map the photoreceptor mosaic across the entire macula. The best individual images from the videos (i.e., those with the best photoreceptor contrast and absent shear from saccades) were used to build up a montage of a region sized ~3.5–4.5 mm (~12–15 deg.) depending upon overlap and the subject’s ability to fixate. The images were aligned manually (Fig. 6), where individual photoreceptors could be matched in the overlap region. Software is under development to automatically generate the montage. Torsional eye motion was not corrected for and sometimes prevented perfect overlap of the individual frames.
Generally, AO correction was left active (i.e., loop closed) while the fixation target LED was changed, and with slight adjustment of the beam to maintain pupil centration, the videos could be collected in rapid sequence. In the four subjects where this sequence was performed, the imaging time to collect all the videos ranged from 8 to 18 minutes. Figure 6 shows a central macula montage from one individual.
3.3 Photoreceptor counting
Photoreceptor counting at eccentricities from 0.5–2.5 mm was performed on four subjects. A region sized 50×50 pixels (78×78 µm) was chosen where photoreceptor contrast was good and there was no (or very little) shadowing from overlying retinal capillaries. For the AO-LSO, regions larger than this usually had vessel shadows that provided inaccurate cone density measurements. A program was developed (as part of our AO analysis software libraries) to perform manual cone counting. The manual cone counting software allowed the user to click on the location of an individual cone, which immediately updated the image overlay with a mark indicating that the cone has been counted. When all the cones in a region were fully marked, the image with overlay and cone coordinates was saved. We have also developed automated cone counting software described below, similar to previously published algorithms [42–44]. Besides selecting the region on which to perform cone counting, the software is fully automated. Depending on the region size and cone density, the software takes a few seconds to several minutes to complete in the current Matlab implementation.
To robustly automate the cone counting procedure, several factors have to be controlled or accounted for, including inherent differences in photoreceptor reflectivity and size, variable brightness across the image field, degree of AO correction, and eye motion induced distortions. Photoreceptor waveguiding, caused by light interference between individual discs within the photoreceptors, naturally varies from cone to cone. In addition, blood vessel shadows and other overlaying structures and layers (e.g., NFL) may induce changes in the reflectivity and contrast of the photoreceptor layer across the field. Also, the AO-LSO has slightly non-uniform illumination (from the cylindrical lens), and the field size is such, relative to the isoplanatic patch, that AO correction may be optimal in some regions but not others.
Our automatic algorithm first corrects for a non-uniform image background. A 2×2 median filter is applied. The background is then obtained by morphologically opening the image using a disk-like structuring element with a radius of 5 pixels. After subtracting the background, contrast stretching is performed for image enhancement. All local maxima in the image are then identified, and statistical analysis provides the centroid for the identified maxima as an initial guess for the cones’ locations. Finally, cone locations are filtered to prevent duplicates from closely-spaced centroids. Filtering is the most time consuming operation and is accomplished as follows: a rough estimate for the cone diameter is obtained by thresholding, morphologic transformations (to remove spur, isolated, and H-connected pixels), and statistical analysis to determine the average cone area across the entire region; a region 1.4 times the average cone diameter around each centroid (cone) is examined to determine if multiple centroids exist within the region, indicating potential overlapping cones; if multiple centroids exist, the program calculates the mean intensity within the expanded region for each centroid and keeps only the cone with the maximum mean intensity. The 1.4 factor was found empirically and future development will determine and use frequency analysis to adaptively correct this factor for different cone densities (found at different eccentricities, for example).
Figure 7 shows a comparison between manual and automated cone counting at an eccentricity of 0.8 mm. In this case, a region sized 50×50 pixel was chosen. (When regions sized 100×100 pixels were chosen so as not to overlap with retinal vessels, the automated software produced similar cone densities as the 50×50 pixel regions.) The cone counts for the 50×50 pixel region was 140 for the manual counts and 121 for the automated counts, respectively (13% difference). The locations of the counted cones overlap for manual and automated analysis for a large fraction of cones. Even when the locations don’t overlap because of pixel-to-pixel intensity variation and the manner in which the automated software chooses the centroid, the automatic software algorithm still counted a cone with nearly 90% accuracy. For the AO-LSO at eccentricities <0.8 mm, the 2×2 median filter begins to influence the cone counts because cone peaks and contrast can be totally smoothed out by the median filter. Conversely, removing the median filter will cause the counts for large eccentricities to become artificially inflated.
The results from both manual and automatic counting are shown in Fig. 8 in comparison to previously reported histology . The manual counts achieve fairly good agreement, albeit with slightly larger densities, in comparison to histology at all eccentricities. Subject 1 generally had a higher cone density at all eccentricities. This subject is the same as that shown in Fig. 5(a) where a qualitatively high density was noted. The automated cone counts tended to be closer to histology for larger eccentricities and were inaccurate for the smallest eccentricities (<0.8 mm). At this eccentricity, where the cone densities are greater than ~20,000/mm2, the diameter of the average cone for the AO-LSO is ~4.5 pixels. It is clear that below this the automated software is unable to resolve individual photoreceptors. In general, the manual and automated cone counts were reasonable well matched (<20% difference between them 82% of the time), with the manual counts typically higher than the automated counts (77% of the time) and the largest differences occurring at the smallest eccentricities. For eccentricities greater than 0.8 mm, therefore, the automated software appears to be quite robust and accurate at counting cones.
There are several possible explanations for cone count differences from histology and the fact that the manual cone counts were generally higher than automated counts. First, there is naturally a spread in cone densities across the entire population, but as importantly, there is a variation in the length/size of the eye which was not measured in this study. The proper conversion factor from image angle to distance on the retina depends on this value. Second, it is evident that the eager analyzer may incorrectly label some areas as cones (false positives), and likewise, the software can fail to identify cones in ambiguous regions that fail to trigger detection criteria (false negatives), each with systematic tendencies to one or the other error type that are image- and eccentricity-dependent. Third, it is possible that when choosing a sampling location, the analyzer tends to choose a non-representative region (a process akin to tendency to choose one’s best images for scientific publications). We did not distinguish in our analysis among different retinal sampling location (i.e., nasal, temporal, inferior, superior), though it is known that differences exist between cone densities along the meridional and equatorial directions . This could account for some of the apparent fluctuations in measured cone densities as a function of eccentricity.
3.4 Registration and frame averaging
Involuntary eye motion is always problematic for ophthalmic diagnostic techniques. Motion effects relative to field size obviously increase as the field size decreases, and there is frequently little overlap between adjacent AOSLO images with 1–2 deg fields with poor fixators. Retinal tracking hardware has been previously integrated into AOSLOs, successfully stabilizing images to within ~15 µm . Real-time software registration algorithms have also been demonstrated, stabilizing images to less than a cone diameter when AO correction and patient fixation are adequate [46–48]. However, for simple, low-cost AO systems, some type of post-processing image registration algorithm is required for optimal information extraction. Image registration and frame averaging are important to increase the signal-to-noise ratio, to better resolve moving particles (e.g., blood cells flowing in capillaries), when long duration scans are required (e.g., autofluorescence of lipofuscin or other exogenous fluorophores), and when a specific feature (e.g., photoreceptor, small capillary, etc.) is to be tracked over time. The software described here is suited for both AOSLO and AO-LSO images (with some limitations), does not require perfect AO correction, and accepts a lower than optimal frame yield, discarding non-overlapping and wholly distorted frames.
In general, eye motion can cause intra-frame distortions: compressing, expanding, or shearing pixels depending upon whether the movement is in the same, opposite, or orthogonal direction as the image scanner. Eye motion within a frame will also cause inter-frame translation between adjacent frames. Torsional eye motion can also cause one frame to rotate with respect to another. Eye motion is not fast enough to occur during the integration time of the linear detector (for the AO-LSO with line rates in the tens of kHz the integration time is usually <100 µs), and this fact can be used as an advantage in the registration software. Thus all processing is done on strips 20 pixels in width, which together take ~1.3 ms to acquire (for a frame rate of 15 fps). The registration software only performs translation of these small image strips. No image de-warping (expansion, compression, or shear to compensate eye motion) is currently applied. If there are not matched features because of excessive warping, the region is discarded. Image de-warping to increase yield is a task planned in future analysis software development.
There are three basic steps to the registration software algorithm. First, a key (or anchor) frame is chosen. This is currently performed manually, although refinement of the software will allow for this to be performed automatically using discrimination based upon image brightness, contrast, and/or frequency content. Second, a feature-matching routine is executed to find similar structural features in the key frame and all others. Third, all frames are aligned to the key frame.
In the first step, the operator makes an initial pass on the images to select a key frame. Images with blinks or excessive eye motion are also currently removed in the initial step. Although these frames would be automatically removed in subsequent steps because of a lack of matching features, until the software is fully automated, this step increases the processing speed by reducing the number of frames to process. All images to be processed are smoothed with a median filter to remove some image noise and contrast stretched for image enhancement.
The second step identifies structural features that are identical in the two frames (anchor frame and frame to be aligned). For this purpose we use the scale invariant feature transform (SIFT) developed by Lowe . SIFT provides the coordinates and a descriptor (key) for each identified feature. The descriptor contains a series of parameters related to the geometrical characteristics of the feature and the intensity gradient around the feature, properties that are scale invariant in addition to rotation and translation invariant. Depending on the image content, there can be as many as 1000 SIFT features in a 500×500 pixel image. Figure 8 shows two frames with SIFT-matched features.
Once the features are identified in a pair of images, the next step is to match them. The program calculates the angle (the arccosine of the dot product between the two descriptors), rather than the Euclidian distance, between two descriptors. For small angles, the ratio of angles is a close approximation – but more efficient to calculate – than the ratio of Euclidean distances. We only keep matches in which the ratio of vector angles from the nearest to second nearest neighbor is less than a pre-selected value (0.6). A larger value would generate more matching points but add more false matches. There are usually on the order of 100 matches for a pair of images. If there are less than ten matches we consider the result unreliable, discard the frame from the set, and move on to the next frame.
Once the coordinates of all matched points in the anchor frame and frame to be aligned are found, a transformation between frames is calculated. For simple whole frame translation, the average translation angle for all pairs of matched points is found and converted into a single set of x-y displacement coordinates. More typically, however, intra-frame motion will cause distortions where different parts of the image will translate by different distances and in different directions. In this case, the matched points need to be filtered. A histogram (bin size=20 deg.) of the translation angles for all pairs of matched points is calculated and analyzed. If the histogram has only one maximum, this indicates simple translation and we keep only the matched points from the bin corresponding to this maximum. Translation angles outside of this bin most likely indicate false feature matching. If the histogram has two maxima, this indicates that two parts of the frame moved in different directions. We only keep the matched points from the bins corresponding to the two maxima, and consider angles outside of these bins unreliable.
To perform a sophisticated transformation in the presence of complex eye motion, the frame to be aligned is split into 20-pixel thick horizontal strips. When a strip contains one or more matched points, it is displaced by the average translation angle of the matched points within the strip. The final registered image generally contains some discontinuities (seen as horizontal black stripes) since individual strips are translated differently, some strips have no matching points and are disregarded, and full image de-warping is not performed. To create a composite image of all frames without discontinuities, only the non-zero image regions of the registered images are averaged. Both the average image and each registered image are filtered with a contrast enhancement filter.
Figure 10 shows a single foveal image (a), a composite image of 89 frames with full frame registration and averaging (b), a composite image with strip registration and averaging (c), and a composite image without registration (d). Note that improvement in vessel contrast (especially in the lower left portion of the image) in Fig. 10(c) compared to Fig. 10(b).
3.5 Imaging blood flow and retinal structures
To examine flow in the finest retinal capillaries, the AO-LSO system focus was pulled forward (with the DM) to the inner retinal layers where the vessels reside (inner nuclear and inner plexiform layers). Flow could be visualized in the smallest capillaries that form a ring around the foveal avascular zone (FAZ). Figure 11 shows the ring of capillaries around the FAZ for three individuals and a video of flow in the subject shown in Fig. 11(a). The images are averaged from 50–70 individual registered video frames using the post-processing registration software described above.
On several subjects, the ONH was also examined. Figure 12 shows three images of the ONH at different depths collected from three subjects. Individual nerve fiber bundles entering the ONH can be visualized when the system focus is set to more anterior (i.e., inner) layers (Fig. 12(a)). As the focus is adjusted deeper, photoreceptors along the rim and in the scleral crescent can be seen (Fig. 12(b)). When the focus is pushed deep into the ONH, the pores and sieve-like structure of the lamina cribrosa can be seen (Fig. 12(c)).
The AO-LSO system is not intended to replace current AOSLO’s , but rather to supplement it for those users that choose to sacrifice a small amount of performance for a reduction in complexity. As such, it is worthwhile to discuss several limitations to the AO-LSO approach, but also some advantages that it may have in comparison to a conventional flying-spot AOSLO.
First, because the illumination optics involves a focused line rather than a focused spot, there is a broad Gaussian variation in illumination intensity across the field for the AO-LSO. In our implementation, we have chosen the cylindrical lens focal length such that the 1/e power points are beyond the edges of the field. This effect is observed in the images presented throughout the paper as a brightness changes from top to bottom (e.g. Figs. 4, 10, and 11) or from left to right in the images that were not rotated 90° to correspond correctly to retinal orientation (e.g. Figs. 7, 9). However, the illumination is fixed and scalable across the field and does not degrade resolution or impair clinically relevant diagnosis in any way.
Second, unlike an AOSLO, the minimum resolution of the AO-LSO is constrained by the detection hardware. The image resolution of the instrument is essentially determined by the LAD pixel size (projected onto the retina) and the Airy disc of the input beam. Because an AOSLO uses a single point detector in which the image information is encoded into temporal variations, any arbitrary detection aperture size relative to the Airy disc can be projected and detected, depending upon the relay optics and confocal pinhole size. Therefore, fairly large pinholes can be used while efficiently collecting light and over-sampling an arbitrarily small field to achieve the highest attainable resolution. In contrast, it is not so simple to oversample the field for the AO-LSO, since the separation of adjacent (square) pixels and the confocal aperture are the same, fixed by the LAD. The Nyquist sampling limit sets the maximum resolvable spatial frequency which corresponds to twice the pixel separation. This should be a fraction of the Airy disc diameter to produce ideal confocality and diffraction-limited resolution with AO-LSO. However, such a small pixel relative to the Airy disc results in a hyperconfocal condition where light collection efficiency is significantly diminished. Sampling below the Airy radius results in rapid loss of signal strength without appreciable improvement in resolution. Furthermore, cross-talk from adjacent illuminated line elements degrades contrast. Thus, the pixel/field size cannot be made arbitrarily small for an AO-LSO However, since a focused line is projected onto the retina rather than a focused spot, more power can be safely delivered to the eye (ANSI extended source factor) to partially compensate for the collection losses in the hyperconfocal case. In the AO-LSO presented here, we are slightly hyperconfocal: the Airy disc radius is ~2.2 µm (at 820 nm in a 6 mm pupil paraxial eye) and the pixel width reference to the retina is ~1.5 µm assuming an eye length of 17 mm.
For most research AOSLO’s with less than ~3 deg field the main advantage is the ability to resolve cones very near the fovea. Outside of that single ~2–3 deg patch such high resolution is not required for cone counting; small fields are less efficient for rapid survey of large regions of the posterior pole. Of course, no technical obstacle precludes the development of a flying spot AOSLO with a wider field. However, to map a large region of the retina with our simpler AOLSO system with a field of view two or more times larger requires less than a quarter of the number of mosaic patches, and therefore less than a quarter of the clinical imaging time.
One of the consequences of the larger field size for the AO-LSO is that a region larger than the so-called “isoplanatic patch” is imaged. The isoplanatic region was effectively set as part of the system design (with AO mirror corrections through the third order) so that the total residual aberrations beyond these low orders across the 5.5 deg field would not exceed the Airy disc diameter at the perimeter, and so that the required low-order AO corrections (other than sphere) would not exceed the stroke of the mirror in any part of the field of regard of the large final mirror. Furthermore, there is no single astigmatism correction (the dominant aberration of off-axis spherical mirrors) that will compensate fully across large field angles. All these aberrations have the effect of limiting how short the focal lengths can be in the optical train relative to the field size, independent of the real eye, that represent an irreducible “pseudo-isoplanatic” patch size of the optical system over which high quality diffraction-limited focus is maintained. Of course, the eye itself with large pupil (though already reasonably isoplanatic by evolutionary design) will contribute irreducible aberrations and further reduce the effective size of the isoplanatic region.
In practice, a more severe limitation on the effective isoplanaticity of the focal region may arise from the gross tilt of the focal plane with respect to the retina across the field. This is especially severe in myopes where the lengthened eye may tilt the retina away from the ideal emmetropic posterior pole surface or with eccentric pupil alignment. Fortunately, the AO-LSO can address this in some circumstances (i.e., nasal-temporal tilt) by tilting the detector surface (with a rotary mount) in the plane of the line focus to compensate – this is equivalent to moving a point detector axially  at the line rate to adjust focus on the fly – a practical impossibility for large refractive errors with an SLO. The vertical extent of the scan can be reduced to limit the (superior-inferior tilt) error of the mosaic patch, and compensated with more patches. In addition, the ability to position the separate beacon at any point in the image field allows corrections to be optimized for specific locations in the image.
Besides the advantages discussed above, i.e. reduced footprint, complexity, number of optical components, and cost and larger scan area, the AO-LSO is also more amenable to very high frame rates (>70 fps), using high speed linear detectors now available (e.g., Sprint, Basler Vision Inc.). This advantage can be exploited for applications that require image averaging and registration in the presence of eye motion, but also to study the dynamics of retinal blood flow.
We have designed a second generation AO-LSO that further reduces the size and complexity of the system. The overall footprint of the optical setup is reduced to 12×18 inches (30×45 cm), a size that can be mounted to a slit lamp stand. Also, we have begun to explore wavefront sensorless algorithms [51–55] exploiting unique LSO features for further hardware reduction. The wavefront sensorless approach to AO uses information extracted from the image sensor to calculate the sample aberrations that are applied to the correcting element, obviating the need for a separate wavefront sensor. These techniques are often limited to small aberrations for AOSLOs because large aberrations will direct light completely out of the pinhole. So the line field approach may offer some advantages in correction of larger aberrations. Also, line image-based algorithms can be implemented in extremely fast loops near the line acquisition rates (tens of kHz), so that the DM sets the limiting rate for AO correction.
We have successfully demonstrated the integration of adaptive optics into a line scanning ophthalmoscope. The instrument was able to achieve AO correction with comparable fidelity to a research AOSLO and resolution of photoreceptors, capillaries, and other clinically significant retinal targets. While there are some optical and physical constraints fundamental to the line scanning technique that define system capabilities (e.g., the field cannot be made arbitrarily small), the AO-LSO has significant advantages over research AOSLOs that make it well-suited to further development into a clinical prototype package and eventual commercialization. Future system development, including use of wavefront sensorless algorithms, may lead to a further reduction in hardware, footprint, cost, and overall instrument complexity. This development may eventually make high resolution AO instrumentation more accessible to vision researchers and ophthalmologists for a wide variety of applications from disease diagnosis to the evaluation of the effectiveness of new drug therapies.
This work was supported by NIH grant EY018509. We thank Steve Burns for assistance in the development of cone counting and frame registration software.
References and links
1. C. E. Bigelow, N.V. Iftimia, R. D. Ferguson, T. E. Ustun, B. Bloom, and D. X. Hammer, “Compact multimodal adaptive-optics spectral-domain optical coherence tomography instrument for retinal imaging,” J. Opt. Soc. Am. A 24, 1327–1336 (2007).
2. S. A. Burns, R. Tumbar, A. E. Elsner, D. Ferguson, and D. X. Hammer, “Large-field-of-view, modular, stabilized, adaptive-optics-based scanning laser ophthalmoscope”, J. Opt. Soc. Am. A 24, 1313–1326 (2007).
4. E. J. Fernandez, B. Hermann, B. Povazay, A. Unterhuber, H. Sattmann, B. Hofer, P.K. Ahnelt, and W. Drexler, “Ultrahigh resolution optical coherence tomography and pancorrection for cellular imaging of the living human retina,” Opt. Express 16, 11083–11094 (2008), http://www.opticsinfobase.org/oe/abstract.cfm?URI=oe-16-15-11083.
5. B. Hermann, E. J. Fernandez, 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).
6. J. Z. 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).
7. 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), http://www.opticsinfobase.org/oe/abstract.cfm?URI=oe-10-9-405. [PubMed]
8. R. J. Zawadzki, B. Cense, Y. Zhang, S. S. Choi, D.T. Miller, and J. S. Werner, “Ultrahigh-resolution optical coherence tomography with monochromatic and chromatic aberration correction”, Opt. Express 16, 8126–8143 (2008). http://www.opticsinfobase.org/oe/abstract.cfm?URI=oe-16-11-8126 [CrossRef]
9. R. J. Zawadzki, S. M. Jones, S. S. Olivier, M. T. Zhao, B. A. Bower, J. A. Izatt, S. Choi, S. Laut, and J. S. Werner, “Adaptive-optics optical coherence tomography for high-resolution and high-speed 3D retinal in vivo imaging”, Opt. Express 13, 8532–8546 (2005), http://www.opticsinfobase.org/oe/abstract.cfm?URI=oe-13-21-8532. [CrossRef]
10. Y. Zhang, B. Cense, J. Rha, R. S. Jonnal, W. Gao, R. J. Zawadzki, J. S. Werner, S. Jones, S. Olivier, and D. T. Miller, “High-speed volumetric imaging of cone photoreceptors with adaptive optics spectral-domain optical coherence tomography”, Opt. Express 14, 4380–4394 (2006), http://www.opticsinfobase.org/oe/abstract.cfm?URI=oe-14-10-4380.
12. Z. Y. Zhong, B. L. Petrig, X. F. Qi, and S. A. Burns, “In vivo measurement of erythrocyte velocity and retinal blood flow using adaptive optics scanning laser ophthalmoscopy”, Opt. Express 16, 12746–12756 (2008), http://www.opticsinfobase.org/oe/abstract.cfm?URI=oe-16-17-12746.
15. J. Carroll, M. Neitz, H. Hofer, J. Neitz, and D. R. Williams, “Functional photoreceptor loss revealed with adaptive optics: An alternate cause of color blindness”, Proc. Nat. Acad. Sci. USA 101, 8461–8466 (2004).
18. A. Roorda and D. R. Williams, “Optical fiber properties of individual human cones”, J. Vis. 2, 404–12 (2002).
19. W. Makous, J. Carroll, J. I. Wolfing, J. Lin, N. Christie, and D. R. Williams, “Retinal microscotomas revealed with adaptive-optics microflashes”, Invest. Ophthalmol. Vis. Sci. 47, 4160–4167 (2006).
20. N. M. Putnam, H. J. Hofer, N. Doble, L. Chen, J. Carroll, and D. R. Williams, “The locus of fixation and the foveal cone mosaic”, J. Vision 5, 632–639 (2005).
21. D. C. Gray, W. Merigan, J. I. Wolfing, B. P. Gee, J. Porter, A. Dubra, T. H. Twietmeyer, K. Ahmad, R. Tumbar, F. Reinholz, and D. R. Williams, “In vivo fluorescence imaging of primate retinal ganglion cells and retinal pigment epithelial cells”, Opt. Express 14, 7144–7158 (2006), http://www.opticsinfobase.org/oe/abstract.cfm?URI=oe-14-16-7144.
22. J. I. W. Morgan, A. Dubra, R. Wolfe, W. H. Merigan, and D. R. Williams, “In Vivo Autofluorescence Imaging of the Human and Macaque Retinal Pigment Epithelial Cell Mosaic”, 50, 1350–1359 (2009).
23. D. C. Gray, R. Wolfe, B. P. Gee, D. Scoles, Y. Geng, B. D. Masella, A. Dubra, S. Luque, D. R. Williams, and W. H. Merigan, “In vivo imaging of the fine structure of rhodamine-labeled macaque retinal ganglion cells”, Invest. Ophthalmol. Vis. Sci. 49, 467–473 (2008).
24. M. K. Yoon, A. Roorda, Y. Zhang, C. Nakanishi, L.-J. C. Wong, Q. Zhang, L. Gillum, A. Green, and J. L. Duncan, “Adaptive Optics Scanning Laser Ophthalmoscopy Images Demonstrate Abnormal Cone Structure in a Family with the Mitochondrial DNA T8993C Mutation”, Invest. Ophthalmol. Vis. Sci. iovs.08–2029 (2008).
25. S. S. Choi, N. Doble, J. L. Hardy, S. M. Jones, J. L. Keltner, S. S. Olivier, and J. S. Werner, “In vivo imaging of the photoreceptor mosaic in retinal dystrophies and correlations with visual function”, Invest. Ophthalmol. Vis. Sci. 47, 2080–2092 (2006).
27. J. I. Wolfing, M. Chung, J. Carroll, A. Roorda, and D. R. Williams, “High-resolution retinal imaging of cone-rod dystrophy”, Ophthalmology 113, 1014–1019 (2006).
28. S. S. Choi, R. J. Zawadzki, J. L. Keltner, and J. S. Werner, “Changes in cellular structures revealed by ultrahigh resolution retinal imaging in optic neuropathies”, Invest. Ophthalmol. Vis. Sci. 49, 2103–2119 (2008).
29. J. L. Duncan, Y. H. Zhang, J. Gandhi, C. Nakanishi, M. Othman, K. E. H. Branham, A. Swaroop, and A. Roorda, “High-resolution imaging with adaptive optics in patients with inherited retinal degeneration,” Invest. Ophthalmol. Vis. Sci. 48, 3283–3291 (2007).
30. D. X. Hammer, N. V. Iftimia, R. D. Ferguson, C. E. Bigelow, T. E. Ustun, A. M. Barnaby, and A. B. Fulton, “Foveal fine structure in retinopathy of prematurity: An adaptive optics Fourier domain optical coherence tomography study”, Invest. Ophthalmol. Vis. Sci. 49, 2061–2070 (2008).
31. P. M. Prieto, F. Vargas-Martin, S. Goelz, and P. Artal, “Analysis of the performance of the Hartmann-Shack sensor in the human eye”, J. Opt. Soc. Am. A 17, 1388–1398 (2000).
32. N. Doble, G. Yoon, L. Chen, P. Bierden, B. Singer, S. Olivier, and D. R. Williams, “Use of a microelectromechanical mirror for adaptive optics in the human eye”, Opt. Lett. 27, 1537–1539 (2002).
33. E.J. Fernandez, L. Vabre, B. Hermann, A. Unterhuber, B. Povazay, and W. Drexler, “Adaptive optics with a magnetic deformable mirror: applications in the human eye”, Opt. Express 14, 8900–8917 (2006), http://www.opticsinfobase.org/oe/abstract.cfm?URI=oe-14-20-8900.
35. F. Vargas-Martin, P. M. Prieto, and P. Artal, “Correction of the aberrations in the human eye with a liquid-crystal spatial light modulator: limits to performance”, J. Opt. Soc. Am. A 15, 2552–2562 (1998).
37. A. N. S. I. (ANSI), American National Standard for the Safe Use of Lasers. 2000, American National Standard Institute, Inc.
38. F. C. Delori, R. H. Webb, and D. H. Sliney, “Maximum Permissible Exposures for Ocular Safety (ANSI 2000), with Emphasis on Ophthalmic Devices”, J. Opt. Soc. Am. A 24, 1250–1265 (2007).
39. D. X. Hammer, R. D. Ferguson, C. E. Bigelow, N. V. Iftimia, T. E. Ustun, and S. A. Burns, “Adaptive optics scanning laser ophthalmoscope for stabilized retinal imaging”, Opt. Express 14, 3354–3367 (2006), http://www.opticsinfobase.org/oe/abstract.cfm?URI=oe-14-8-3354.
40. D. X. Hammer, M. Mujat, N. V. Iftimia, and D. FergusonCompact adaptive optics line scanning laser ophthalmoscope in Photonics West -BIOS; Ophthalmic Technologies XIXF. Manns, P.G. Söderberg, and A. Ho, eds., Proc. SPIE Vol.7163, 71630J, 2009.
41. C. A. Curcio and K. R. Sloan, “Packing Geometry of Human Cone Photoreceptors - Variation with Eccentricity and Evidence for Local Anisotropy”, Vis. Neurosci. 9, 169–180 (1992).
42. T. Y. P. Chui, H. X. Song, and S. A. Burns, “Adaptive-optics imaging of human cone photoreceptor distribution”, J. Opt. Soc. Am. A 25, 3021–3029 (2008).
43. K.Y. Li and A. Roorda, “Automated identification of cone photoreceptors in adaptive optics retinal images,” J. Opt. Soc. Am. A 24, 1358–1363 (2007).
44. B. Xue, S. S. Choi, N. Doble, and J. S. Werner, “Photoreceptor counting and montaging of en-face retinal images from an adaptive optics fundus camera”, J. Opt. Soc. Am. A 24, 1364–1372 (2007).
45. T. Y. P. Chui, H. Song, and S. A. Burns, “Individual variations in human cone photoreceptor packing density: Variations with refractive error”, Invest. Ophthalmol. Vis. Sci. 49, 4679–4687 (2008).
46. D. W. Arathorn, Q. Yang, C. R. Vogel, Y. Zhang, P. Tiruveedhula, and A. Roorda, “Retinally stabilized cone-targeted stimulus delivery”, Opt. Express 15, 13731–13744 (2007), http://www.opticsinfobase.org/oe/abstract.cfm?URI=oe-15-21-13731.
47. C. R. Vogel, D. W. Arathorn, A. Roorda, and A. Parker, “Retinal motion estimation in adaptive optics scanning laser ophthalmoscopy”, Opt. Express 14, 487–497 (2006), http://www.opticsinfobase.org/oe/abstract.cfm?URI=oe-14-2-487.
48. S. B. Stevenson and A. RoordaCorrecting for miniature eye movements in high resolution scanning Laser Ophthalmoscopy in Ophthalmic Technologies XV, F. Manns, P. Soderberg, and A. Ho, eds., Proc. SPIE Vol.5688A, 145–151, 2005.
49. D. Lowe, “Method and apparatus for identifying scale invariant features in an image and use of same for locating an object in an image”, US Patent number: 6,711,293 (2004).
50. E. J. Botcherby, M. J. Booth, R. Juskaitis, and T. Wilson, “Real-time extended depth of field microscopy,” Opt. Express 16, 21843–21848 (2008), http://www.opticsinfobase.org/oe/abstract.cfm?URI=oe-16-26-21843.
51. D. Debarre, E. J. Botcherby, M. J. Booth, and T. Wilson, “Adaptive optics for structured illumination microscopy”, Opt. Express 16, 9290–9305 (2008), http://www.opticsinfobase.org/oe/abstract.cfm?URI=oe-16-13-9290.
52. D. Debarre, M. J. Booth, and T. Wilson, “Image based adaptive optics through optimisation of low spatial frequencies”, Opt. Express 15, 8176–8190 (2007), http://www.opticsinfobase.org/oe/abstract.cfm?URI=oe-15-13-8176.
54. A. J. Wright, D. Burns, B. A. Patterson, S. P. Poland, G. J. Valentine, and J. M. Girkin, “Exploration of the optimisation algorithms used in the implementation of adaptive optics in confocal and multiphoton microscopy”, Microsc. Res. Tech. 67, 36–44 (2005).
55. L. Sherman, J. Y. Ye, O. Albert, and T. B. Norris, “Adaptive correction of depth-induced aberrations in multiphoton scanning microscopy using a deformable mirror”, J. Microsc. 206, 65–71 (2002).