Trans-retinal cellular imaging with multimodal adaptive optics

Adaptive optics (AO), when coupled to different imaging modalities, has enabled resolution of various cell types across the entire retinal depth in the living human eye. Extraction of information from retinal cells is optimal when their optical properties, structure, and physiology are matched to the unique capabilities of each imaging modality. Despite the earlier success of multimodal AO (mAO) approaches, the full capabilities of the individual imaging modalities were often diminished rather than enhanced when integrated into multimodal platforms. Furthermore, many mAO designs added unnecessary complexity, making clinical translation difficult. In this study, we present a novel mAO system that combines two complementary approaches, scanning laser ophthalmoscopy (SLO) and optical coherence tomography (OCT), in one instrument using a simplified optical design, flexible alternation of scanning modes, and independent focus control. The mAO system imaging performance was demonstrated by visualization of cells in their mosaic arrangement across the full depth of the retina in three human subjects, including microglia, nerve fiber bundles, retinal ganglion cells and axons, and capillaries in the inner retina and foveal cones, peripheral rods, and retinal pigment epithelial cells in the outer retina. Multimodal AO is a powerful tool to capture the most complete picture of retinal health.


Introduction
High-resolution retinal imaging has gained increased use for disease diagnosis and treatment outcome assessment owing to tremendous technological advances over the last two decades. Pivotal in these advances is adaptive optics (AO), which when combined with various ophthalmic imaging modalities, can probe the living retina at the cellular level. Since its first demonstration for ophthalmic imaging [1], AO has been successfully integrated into many imaging modalities, including non-confocal (i.e., flood illumination and detection) fundus photography, confocal scanning laser ophthalmoscopy (SLO), optical coherence tomography (OCT), and one-and two-photon fluorescence imaging, enabling microscopic views of single retinal cells in the living eye [2][3][4]. To date, many cells and cellular structures across the entire retinal depth have been resolved using different imaging modalities, including retinal pigment epithelium (RPE) [5][6][7][8][9], cones and rods [10][11][12][13], cone inner segments [14], Henle's fiber bundles [15,16], retinal capillaries [17,18] and vessel walls [19,20], retinal ganglion cells (RGC) [21,22], nerve fiber bundles (NFB) [23][24][25], and microglia [22]. Each of these AO imaging modalities has unique advantages for resolving particular cell types or retinal features [18], and therefore there has been increased interest in the use of multimodal AO (mAO) systems for a more complete picture of retinal health.
Some of the first attempts to develop mAO instruments resulted in easier scan navigation or stabilization by adding a non-AO wide-field imaging or tracking device to existing AO systems [26][27][28]. While those approaches improved clinical utility, they did not extend the AO performance achieved using a single AO imaging modality. Many studies used multiple stand-alone imagers, for example, adaptive optics scanning laser ophthalmoscopy (AOSLO) and conventional OCT [29], the former to provide high transverse resolution (2-3 µm) and the latter to provide high axial resolution (3-4 µm). However, because the images were acquired from two separate systems, full interpretation required additional post-processing tasks to register the SLO and OCT images. Moreover, the lateral resolution of the OCT images, acquired without AO, was sub-optimal, and the imaging session duration was necessarily long. To address these problems several investigators integrated different imaging modalities into a single AO beam path [12, 30-32], including prototype devices intended for clinical use [33,34]. Early attempts to combine AOSLO with adaptive optics optical coherence tomography (AOOCT) typically slowed the OCT B-scan rate to the SLO frame rate, resulting in under-utilization of the OCT mode. Others optimized transverse imaging speed for both modalities but necessarily sacrificed OCT volume rate and axial resolvability due to axial motion artifact [35,36], the latter issue overcome by use of depth tracking schemes [13]. Those solutions diminished the full capability of the two imaging modalities, and often added unnecessary system complexity with implementation.
In this study, we present a novel design for a mAO retinal imaging system (henceforth FDA mAO), which combines two complementary approaches (AOSLO and AOOCT) in one imager through a simplified optical design with minimized system aberration for the investigation of cells and cellular structures in the living human retina. The system provides optimal imaging performance for each modality and flexible alternation between the two imaging modes with independent focus control. The FDA mAO system performance is demonstrated by imaging cells across the entire retinal depth on three healthy subjects. The multimodal approach provides a platform to study retinal physiological and structural properties in both healthy and pathological eyes, paving the way to assess new therapeutic treatment outcomes.

Description of the FDA multimodal AO system
The FDA mAO system, which combines AOSLO and AOOCT channels in a single instrument, was designed with Zemax optical design software (Zemax LLC, Washington, USA). The system provides optimized optical performance and flexibility for joint or independent operation of the two imaging modes. Figure 1 shows a schematic of the FDA mAO system. The primary beam path (and OCT sample arm) consists of four pairs of afocal telescopes. The telescopes are configured for out-of-plane operation to compensate system astigmatism [27,37], which arises from off-axis use of spherical mirrors (SM) and is known to degrade closed-loop AO performance. The afocal telescopes conjugate the pupil of the eye with the system active components, including the Shack-Hartmann Wavefront Sensor (SHWS) lenslet array, the deformable mirror (DM, ALPAO, France), the resonant scanner (EOPC, New York NY USA), and the galvanometer scanners (Thorlabs, Newton NJ USA). SM focal length is an attribute that represents a trade-off in the system design: long focal length SMs with small rotational angles are favorable for minimizing off-axis astigmatism [37]

System control and electronics
FDA mAO system control is accomplished with a single host personal computer (PC), running two programs: AO control software and image acquisition software. The AO control software collect and displays the SHWS camera images, calculates wavefront spot centroids and slopes, performs AO closed-loop control, and controls the DM and AL. It also calculates the Zernike coefficients and wavefront aberration for real-time display and provides autofocus and preset focus settings for the DM and AL. The image acquisition software collects and displays in real-time the OCT and SLO images, sets the field size via scanner (galvanometer and resonant scanner) control, and operates the fixation target (FT). The system PC uses three framegrabbers (PCIe-1430 and PCIe-1433, National Instruments Inc., Austin TX USA and Solios eA/XA, Matrox Electronic Systems Ltd, Dorval, Quebec, Canada) to collect the SLO, OCT, and WS images and two data acquisition cards (PCIe-6363 and USB-6259, National Instruments Inc.) to process galvanometer and resonant scanner position and drive waveform signals. The DM and FT stage communicate with the host PC via USB.
Custom control, image and signal processing, user interface, and analysis software for the FDA mAO system was written in LabVIEW (National Instruments Inc., Austin TX USA), MATLAB (Mathworks Inc., Natick MA USA), and C/C ++ . Three programs were developed to use the video card graphical processing unit (GPU, GeForce GTX-760, NVIDIA, Santa Clara CA USA) via the Compute Unified Device Architecture (CUDA) parallel programming platform for OCT image processing, WS spot centroiding, and SLO image de-warping, all performed in real-time. The system is designed to operate in 'slow scan' or 'fast scan' modes by a single selection in the user interface. The WS camera operates at 10 Hz, and real-time GPU-based spot centroiding allows a closed-loop dynamic ocular aberration correction bandwidth of several Hz.

Simultaneous AOSLO/OCT imaging: slow scan mode
The 'slow scan' mode operates similarly to previously published multimodal AO systems [30,31,38], in which the SLO is the primary imaging modality. Simultaneous SLO/OCT imaging is achieved with each OCT B-Scan synchronized to every SLO raster scan, as depicted in Fig. 2. Because the SLO RS rate is fixed, all timing and image frame rates are derived from the resonant scanner frequency. The SLO RS used in the FDA mAO system has a resonant frequency of 13.5 kHz, the SLO frame rate operates at 27 Hz (500 × 500 pixels), and the OCT A-line acquisition speed is matched to the RS frequency. Because the SLO image yaxis and OCT B-scan (1024x500 pixels) are produced using the same scanner (G v ), the beams are always precisely registered, within the limit of the optical alignment (i.e., collinearity of SLO and OCT beams) and transverse chromatic aberration. Furthermore, OCT volumetric scans can be created by sweeping the second galvanometer (G h ) across a retinal patch, the size of which can either be matched to the SLO image size or different. Due to the slow OCT scan speed (set and limited by the SLO resonant scanner frequency), eye motion results in gaps in the OCT en face images [38]. To fill the gaps requires acquisition of multiple OCT volumes.
Although the SLO and OCT channels are locked in terms of acquisition timing and vertical position, independent SLO focus control is accomplished with a tunable adaptive lens (AL), placed immediately in front of the SLO collimator (see Fig. 1). The AL focus range is −1.5 to 3.5D, which is sufficient to traverse the entire retinal thickness. The independent focus control feature permits navigation and frame rate display of structure of interests with SLO. It also allows correction of the longitudinal chromatic aberration (LCA) between 760 nm (SLO) and 830 nm (OCT) imaging wavelengths and the LCA variation across population [39].

Volumetric AOOCT imaging: fast scan mode
The 'fast scan' mode takes full advantage of high-speed OCT imaging capabilities by maximizing the OCT A-line acquisition speed up to 210 kHz for the FDA mAO system. The fast acquisition speed is achieved by reading out the central 1024 pixels of the spectrometer in 8-tap, 10-bit camera mode (limited by the NI framegrabber). The measured spectrometer sensitivity and roll-off were 83 dB and −9.9 dB/mm (−7.7 dB/mm up to 1 mm depth), following the method reported by Agrawal [40]. The SLO channel is not active in 'fast scan' mode. For the applications shown below in Section 3, OCT volumes of 300 × 300 lateral pixels were collected covering a 1.5° × 1.5° FOV at a volume rate of 2.3 Hz. System focus through retinal layers was achieved with the DM.

Experimental design
FDA mAO retinal imaging performance was demonstrated in healthy human subjects. All human subject procedures were approved by the FDA Institutional Review Board and adhered to the tenets of the Declaration of Helsinki. Written informed consent was obtained after potential risk explained to each subject.
Retinal locations between the fovea and 12° retinal eccentricity were recorded in the right eye of three subjects with ages from 23 to 33 years old (S1: 31, S2: 33, S3: 23 years old). All subjects had best corrected visual acuity of 20/20 or better and were free of ocular disease. The AOSLO and AOOCT beam power at the cornea were measured to be 200 and 420 μW, respectively, and were within safe limits established by American National Standards Institute (ANSI) [41] for the retinal illumination pattern and length of the experiment used. The right eye was cyclopleged and dilated with Tropicamide 0.5%. The eye and head were aligned and stabilized using a chin and head rest attached to manually-driven XYZ patient translation stage.
Two tests were conducted to assess the FDA mAO system imaging performance with dynamic AO correction. First, SLO and OCT images were collected while the system operated in 'slow scan' mode, primarily to characterize SLO image quality. SLO/OCT videos (100-200 SLO frames and OCT B-scans) were acquired with a FOV between 0.75° × 0.75° and 2° × 2° of the photoreceptor or retinal capillary network at multiple retinal locations with AO best focus set to the corresponding layer. Pinholes with 1 and 8 Airy disc diameter were used respectively for photoreceptor and capillary imaging. SLO videos were also collected using the two independent focus control methods (AL vs. DM). The second test was designed to assess OCT imaging quality in 'fast scan' mode. For this test, 40 OCT videos with 3 volumes/video (900 total B-scans) were collected at an eccentricity 7° temporal to the fovea in two subjects with the focus set to the photoreceptor-RPE complex. This test examines the ability of the FDA mAO system to resolve retinal structures with tight physical axial separation. It has been demonstrated that individual RPE cells can be resolved by reducing speckle noise from the OCT en face image [7]. To capture the temporal dynamics of RPE cell organelle motility with speckle decorrelation, the OCT volume sets were collected with a separation of at least 3 s (typically ~5 s) [42]. One best volume was selected from each of the 40 OCT videos for post processing to produce the averaged RPE image. To assess the FDA mAO system's ability to detect weakly light scattered cells such as RGC and other fine granular structures in the inner retina, another 40 videos with 5 volumes/video (1500 total Bscans) were also acquired at 12° temporal to the fovea on two subjects with system focus set to the inner retina just below the nerve fiber layer (NFL). Volumes with blinks or significant eye motion artifact were excluded for post processing.
Depending upon the retinal target, patient imaging sessions lasted about an hour, including initial alignment and frequent rests for patient comfort. SLO image collection (e.g., photoreceptor and capillary network mosaics) took ~30 minutes to complete, while RPE and RGC imaging (collection of 40 volumes) each took ~15 minutes to complete.

Post processing and analysis
Improved image signal-to-noise ratio (SNR) can be achieved by averaging multiple images collected from the same retinal location. However, eye motion is generally several times larger than the retinal structures under investigation. Therefore, to achieve optimal results, eye motion must be corrected. We achieved registration of image sequences by applying a 2-D strip-wise registration approach [43] for SLO images, and a 3-D registration algorithm for OCT volumes [22]. Reference images/volumes were manually selected according to criteria including retinal structure sharpness, minimal eye movement artifacts, and common overlap with other reg capillaries) w as cell density Cell density w Figure 3 sum system's 3.6° graph illustra denoting the diffraction cri displacement The beam dis (0.28 mm) at the upstream two planes.  The rich tapestry of neurons, glia, and blood vessels in the inner retina were all observed in our data sets. The cross-sectional view illustrates a single neuron identified with the GCL (Fig. 9(B), red arrow). At this eccentricity (12° T) the GC somas are organized in a monolayer. En face projections provide a detailed view of other inner retina features of interest, including microglia cells and their processes at the ILM ( Fig. 9(C)), NFBs and individual GC axons (Fig. 9(D)), and the GCL soma mosaic, where different characteristic sizes and reflectance differentiate RGC subtypes (Fig. 9(E)). The averaged AOOCT volume shows clear delineation of three layers of retinal vessels (Fig. 9(E-G)), which are often visualized by OCT angiography techniques (collection of multiple B-scans at each lateral location) for higher contrast [17]. The en face fly-through for this subject is shown in Visualization 4. The soma density and diameter were quantified in the GCL for two subjects and found to be 4252 and 3592 cells/mm 2 (with subtraction of predicted Amacrine cell population [48]) and 14.57 ± 2.95 and 14.71 ± 3.28 µm, in agreement with previously reported in vivo human results [22] and histological measurements [48].

Discussion
The promise of adaptive optics has been more than adequately attained since its first demonstration imaging cone photoreceptors more than twenty years ago [1] [6,14,21]. Other times, cells were detected by the compound benefit to contrast realized by careful correction of eye motion to micron and submicron precision along with image averaging on a massive scale [5,8,10]. Parallel imaging feats have also been accomplished without adaptive optics [51]. But it should be clear by now that a state-of-the-art AO imager makes these accomplishments easier to achieve as well as to expand the patient population in which they can occur.
SLO and OCT, in all of their various incarnations, remain fundamentally better suited at different imaging tasks. And this is why they are complementary when married in a multimodal instrument such as the one demonstrated herein. SLO is a confocal technique but retains immense flexibility in the scope and manner in which scattered light is blocked or detected. It is therefore better suited to detection of light from waveguiding photoreceptors or structural boundaries and interfaces where complex refractive index differentials are highest allowing multiple photon scatter. OCT is hyper-confocal and detects predominantly singly scattered photons. However, its axial resolution is uncoupled from its numerical aperture, allowing micron-scale depth sectioning by use of high bandwidth sources. In recent years, the benefit in the en face plane has become more profound as technological developments of new sources have enabled volumetric acquisition speeds at close to video rates (~15-30 volumes/s). These benefits have made OCT better suited to detection of light from cells tightly packed in depth and those with high transparency, particularly in the inner retina. The truth of this statement is demonstrated by the fact that both RPEs and RGCs reside directly beneath highly reflective layers, and yet they are more easily resolved with OCT using AO, robust 3-D registration, and image averaging. With intrinsic contrast, the monolayer of RGL somas are resolvable over a retinal area with little or no overlying NFL with an AOSLO [21] and RPE mosaics only with use of (auto-)fluorescence [5,9,52], or under restricted conditions where the overlying photoreceptors are absent due to disease [53], or in the central fovea where the signal leakage from other cells is eliminated [6]. However, in most cases, the RPE and RGC mosaics are more readily, directly, and easily visualized with AOOCT. Moreover, because OCT collects the full complex interferometric spectrum, phase-based techniques can be further exploited to improve contrast (e.g., computational AO) [54][55][56].
The time when OCT line rates match SLO scan rates is already upon us [57]. Despite these advancements in imaging speed, OCT remains less flexible in detection of multiply scattered light, the contrast benefits of which we are only beginning to realize. Thus for a more complete in vivo cellular survey of the retina, the two modalities function on complementary terms.
Our objective in the development of the FDA mAO system was to demonstrate that all of the retinal cells and structures imaged in live human eyes reported previously, including foveal cones, peripheral rods, RPE, RGC, microglia, capillaries, etc., could be resolved with one imager. As the results demonstrate, we have gone a long way toward completing this goal. However, there are some limitations that have yet to be overcome. While the A-line acquisition rate of 210 kHz was sufficient to collect OCT volumes at 2.3 Hz for RPE and RGC imaging, the OCT speed was still too slow that it placed constraints on the optical design and system functionality. A MHz source (e.g., FDML laser) would allow a more elegant optical design with fewer scanners (i.e., a single resonant scanner for both OCT and SLO) and also simultaneous collection of SLO frames and OCT volumes at high speeds (i.e., video rates). The current design is an improvement over previous multimodal designs [12, 30, 31] because it allows collection of high speed OCT volumes ('fast scan mode'), but it still falls short of its full potential because it does this without acquisition of SLO images. Implementation of a multimodal AOSLO-AOOCT system whose OCT source has a A-scan rate as fast as the SLO pixel clock (~10 MHz) will allow more optimal operation.
Despite the rapid technological advancements and great clinical potential for early diagnosis and treatment outcome assessment, AO has yet to achieve full clinical translation. The establishment of a collaborative research project in AO between FDA and NIH is intended to help foster the clinical translation of this important ophthalmic imaging technology. Both the FDA and NIH are responsible in various ways (FDA as part of our regulatory science program and NIH in early stage and translational funding as well as through its intramural research program, among other efforts) to help provide greater access for patients to novel technologies that are proven safe and effective. The FDA mAO will serve as a platform for future studies to develop new methodology for improved clinical utility and to increase our understanding of visual physiology and eye disease.

Funding
This work was partially funded by a grant from the FDA Critical Path Initiative and the intramural research program of the National Institutes of Health, National Eye Institute. Dr. Saeedi is supported by an NIH Career Development Award (K23EY025014).