C1. Instrument Development
Fig C1 shows all four instruments, each of which is fully functional and is providing valuable new information about the retina and retinal disease.
Based on the success of the first adaptive optics scanning laser ophthalmoscope (AOSLO) (Roorda 2002), the aim in the first 5 years of BRP funding called for the development of 4, second generation AOSLOs, each constructed at a different BRP site (Fig. C1): Steve Burns' lab, formerly at Schepens Eye Research Institute and now at Indiana University (IU), Scot Olivier's group at Lawrence Livermore National Laboratories (LLNL), Austin Roorda's lab, formerly at Houston and now at UC Berkeley (UCB-AR), and David Williams' lab at the University of Rochester (UR). The instrument developed by LLNL has been successfully deployed in Vas Sadda's group at the Doheny Eye Institute (DEI). The instrument at Rochester is used in part for experiments with John Flannery's group at UC, Berkeley (UCB-JF).
The important similarities and differences between these instruments are shown in Fig. C2, which share a core design but differ so as to achieve different scientific goals. The next sections describe some of the improvements that have been made over the original instrument built by Roorda (2002).
Fig C2: The schematic shows the basic AOSLO core design. Each subsystem is labeled and distinct aspects developed by members of the BRP system are indicated.
C1.1. Improvements in Wavefront Correction
All instruments have incorporated new MEMS deformable mirror technology. Until now, the gold standard for wavefront correction in both vision science and astronomy has been a line of deformable mirrors from Xinetics, Inc. These mirrors are large: a 97 actuator mirror is 75 mm in diameter which scales up the size of the entire instrument due to the need to magnify the eye's pupil in the plane of the deformable mirror. Moreover, they are expensive (>$100K) which precludes the commercialization of AO technology in ophthalmology. Boston Micromachines Corporation (BMC) (Watertown, MA) has worked closely with the BRP to design mirrors specifically for vision science applications. That effort was funded by the NSF Science and Technology Center for Adaptive Optics in which all partners in the BRP participate and Williams directs the research program in vision science. BMC mirrors have comparable performance (Doble 2002), more actuators (140), are more than an order of magnitude smaller than Xinetics mirrors (4 x 4 mm), and are more than 4 times less expensive (~$25K). BMC provided the first commercially available MEMS mirror that met the needs for vision science applications. Nonetheless, a limitation of current MEMS mirrors is that their stroke is relatively small, so that the wave aberrations of some severely aberrated eyes, such as those of aging (McLellan 2001) and diseased (Rajagopalan 2005) populations, cannot be corrected. The BRP has ameliorated the stroke limitation with two different approaches: 1) IU demonstrated that reflecting the imaging beam twice from a single deformable mirror doubles the effective stroke of the mirror (Webb 2004). 2) LLNL demonstrated two deformable mirrors in series, a bimorph deformable mirror (Aoptix Technologies, Inc.) to correct lower order aberration (the woofer) and a BMC MEMS mirror dedicated to higher order aberrations (the tweeter), both placed at planes conjugate to the entrance pupil of the eye and to the wavefront sensor. The largest achieved displacement of the bimorph deformable mirror was16 microns over a 10 mm aperture, which can compensate up to +- 3 diopters. Fig C3 shows images acquired from a healthy subject using the dual deformable mirrors. The additional degree of correction obtained with the MEMS mirror (a factor of 2 in the wavefront aberration amplitude) is important for achieving excellent image quality, even in the presence of focus error or high levels of aberrations.
Fig C3: Retinal images from a healthy subject taken at 3° Superio-Nasal over a 1.1° field of view at wavelength of 840 nm after AO correction using the bimorph DM (left panel) and after additional AO correction using the MEMS DM (center panel), along with the corresponding wave aberration over a 6 mm pupil (right panel). Aberrations decreased when the bimorph mirror was activated and decreased further when the MEMS mirror was activated, resulting in improved image quality. This approach has been shared with another BRP invested in AO applied to OCT (Jack Werner, PI).
The deformable mirror industry, encouraged by the BRP and the CfAO, is making steady progress to increase mirror stroke, BMC is close to production of 6 micron stroke mirrors and Imagine Eyes (Orsay, France) has recently introduced a 50 micron stroke mirror that is very promising, since it may be capable of serving the role of both woofer and tweeter.
C1.2. Incorporation of a Low Resolution Wide-Field View of the Fundus
AOSLOs provide a very high resolution view of the fundus, allowing the imaging of single cells, but this comes at the expense of field of view, which is typically 4 degrees or less. Two of the new instruments include a secondary wide-field fundus view that has a resolution comparable to conventional fundus cameras. This field has proven to be very useful in providing retinal landmarks (eg. major vessels) that orient the operator, help to identify regions of interest for imaging, and make it easy to return to the same retinal cells day after day.
C1.3. Use of Low Coherent Light Source to Improve Fidelity of Cone Images
The use of low-coherent light in the new instruments reduces interference artifacts in images of cone photoreceptors (Zhang 2006), allowing much higher fidelity images of the cone photoreceptor mosaic. The introduction of low coherent light sources in the AOSLO is a signature advance of the new BRP instruments.
C1.4. Eye Movement Measurement/Correction and Real-time Image Stabilization
Eye motion has long been troublesome for retinal imaging systems. The amplitudes and velocities of the spectrum of eye movements (saccades, drift, oculomotor tremor, multiple eccentric fixation loci) produce image displacements and distortions with few real-time remedies. Eye motion is especially problematic for high-magnification systems such as the AOSLO because the amplitude of such movements, especially for a clinical population, often exceeds the field size. Since the beginning of the BRP, we have added two new partners, Physical Sciences, Inc. and Montana State University who are contributing to our efforts to stabilize retinal images against eye motion artifacts. The BRP has made significant progress for correction of eye motion in the following three areas:
C1.4.1. Dual Wavelength Imaging and Registration
Fig C4: Fluorescence labeled retinal ganglion cells over an RPE intrinsic fluorescence background. These images show the benefit of dual-registration (right) vs. no registration (left) of a 600-image sequence.
Low light levels from in vivo fluorescence imaging require that hundreds of frames be averaged to increase the signal-to-noise ratio (SNR). Inter-frame eye movements preclude averaging many fluorescence frames directly. In most cases the fluorescence signal is so weak that each fluorescence frame does not contain enough spatial structure for registration. To solve this problem, UR has implemented dual-wavelength imaging for registration of the low SNR fluorescence images. In this approach, images of the same retinal location are recorded through two different imaging channels. The reflectance channel produces high SNR images using a near-infrared light source. The high resolution of small retinal structures in these images enables a very precise estimation of retinal motion using cross-correlation. The estimated motion is then applied to the very low SNR fluorescence images, producing sharp in vivo images which otherwise could not be produced with safe light exposure levels. This technique has been applied to yield in vivo visualization of the RPE mosaic and ganglion cells in primates (Gray 2006) and humans (Wolfing 2006). Roorda's group at UCB has also implemented a dual wavelength imaging approach (Grieve 2006), and demonstrated that the method is invaluable not only for eye movement correction of dim images but also for stimulating the retina with one channel while imaging with the other (see C4.2).
C1.4.2. Software-Based Real-time Stabilization
MSU and UCB-AR have developed software for real-time, high-accuracy video stabilization. The software corrects for low amplitude image distortion at frequencies up to 1 kHz. Furthermore, the stabilization is accurate to a couple of pixels which, for AOSLO magnifications, is a fraction of a cone. The scanning nature of the AOSLO is the key that allows high frequency stabilization. As the image is being recorded by the frame grabber, data is read into the stabilization software, strip by strip. Each strip is individually processed to estimate the eye's current location with respect to the scanning beam. Strips are then adjusted and repositioned in the stabilized video, using a smooth spline-based interpolation scheme. The second key to successful real-time stabilization is the use of a novel computational technique called the Map Seeking Circuit algorithm (Arathorn 2002) for patch-based correlations. The algorithm has significantly lower complexity and higher efficiency than conventional fast Fourier transform-based approaches (Stevenson 2005) and runs on current PC technology.
Fig C5: AOSLO image motion with and without removing distortions from eye movements. Image motion after correction is reduced to less than 1 minute of arc. The right image is a registered frame from the stabilized video shown in the same scale as the plot.
Currently, not all video is correctable in real time, and the software is frame-grabber specific. For off-line purposes a version of the software is available which allows a user to remove distortions from any saved AOSLO video. The offline version allows the user to adjust parameters to improve the eye movement correction. A by-product of the software approach to image stabilization is a that a high-frequency, high-accuracy trace of eye movements is automatically generated representing, to our knowledge, the best traces of eye movements ever made (Stevenson 2005).
The MSU and UCB-AR have used the software-based eye tracking to control presentation of dynamic, complex stimuli (Poonja 2005), providing stabilized stimulus delivery. The procedure is as follows: While processing a frame in real-time, the MSU software estimates the location of the user-selected target position at a time shortly before the raster scans over it. The stimulus delivery arm of the software generates the appropriate signals for driving the acousto-optic modulator to modulate the laser while it scans over the target position. Any complex, grayscale stimulus as small as a single pixel can be projected. We are aiming to reduce the critical time to ~ 4 msec, thereby achieving tracking frequencies on the order of 250 Hz.
C1.4.3. Hardware Stabilization
Fig C6: Hardware-based stabilization. A 10sec video with saccades and drift was recorded in a healthy subject with poor fixation. Shifts in location were eight points in each frame. The center graph is a histogram of frame-to-frame displacements. The right graph shows the probability for a given location to move. During the untracked epoch, this procedure failed, since the frame had excursions larger than the image region, meaning the eye movements were often greater than 100 microns.
While the previous software techniques work well in normal humans with good fixation and in animals that have been administered ocular anesthesia, they may be more problematic in patients with poor fixation. A hardware approach based upon the retinal tracking systems previously developed by PSI (Ferguson 1998, Hammer 2002), was developed and integrated into the AOSLO instruments at Indiana and PSI (Burns 2006, Hammer 2006). The hardware retinal tracking approach uses a low-power beam projected into the eye, aligned to a retinal target (usually the optic nerve head), and dithered at high frequency (16 kHz).
Fig C7: Montage of single frames in a patient with recurrent CSR. The top left region of the retina appears normal, the central region is the focus of the recurrent serious elevations, with fluid elevation (revealed by scattered light imaging) along the right.
The reflectance of the return signal is processed with a phase sensitive detection scheme and fed back in a closed-loop configuration to two galvanometer-driven mirrors through which the AOSLO beam is steered and stabilized through eye movements. The system operates at very high bandwidth (>1 kHz) but in its current configuration is less accurate than the real-time software registration approach (RMS error: ~10 µm for hardware vs. <1-2 µm, or a single pixel diameter for software). The hardware-based tracker also includes an algorithm for automatic blink detection and re-lock. Since the hardware-based eyetracker is not coupled to the AOSLO image, it can track a significantly larger range of eye movements. It can also be used to steer the scanning raster to different locations on the field allowing the system to automatically generate a large field image by combining a series of AO-corrected views from a grid of locations. Fig. C7 shows an example of such a montage.
C1.5. Third Generation AOSLO for Noninvasive Imaging of Human Ganglion Cells
The original specific aims called for the development of two, third generation instruments in the last year of the grant. The first instrument was to be a AO-equipped surgical microscope equipped to facilitate delicate retinal surgeries, such as arterio-venous sheathotomy for branch retinal vein occlusions, and submacular surgery for choroidal neovascularization. We abandoned this plan because changes in retinal disease therapy since the original application, especially the lack of demonstrable benefit of submacular surgery (Hawkins 2004) and optimism about pharmacological therapy for branch vein occlusion, made the immediate need for this instrument less clear. The second proposed third generation instrument was an AOSLO that would be robust enough for deployment in a typical clinical research environment. As described below, our advances for imaging single cells in the inner retina have generally relied on the use of invasive fluorescent dyes, which restricts the usefulness of the technology to animal models. A major goal of the BRP is to develop entirely noninvasive methods for imaging all major human retinal cell types, overcoming the challenges posed by the small refractive index differences and high transmissivity of many retinal cells. A 3rd generation device, currently under construction at UR, is designed to explore the feasibility of this goal, with the demonstration of noninvasive imaging of ganglion cell bodies being the primary goal. This instrument design combines all the best practices and accumulated wisdom of all the BRP partners to date (e.g. eye tracking, dual deformable mirror wavefront correction), and will provide a testbed for exploring various imaging modalities that, in combination with adaptive optics, may allow imaging of retinal neurons besides photoreceptors.