Sensorless adaptive optics multimodal en-face small animal retinal imaging

: Vision researchers often use small animals due to the availability of many transgenic strains that model human diseases or express biomarkers. Adaptive optics (AO) enables non-invasive single-cell imaging in a living animal but often results in high system complexity. Sensorless AO (SAO) can provide depth-resolved aberration correction with low system complexity. We present a multi-modal sensorless AO en face retina imaging system that includes optical coherence tomography (OCT), OCT-angiography, confocal scanning laser ophthalmoscopy (SLO), and fluorescence detection. We present a compact lens-based imaging system design that allows for a 50-degree maximum field of view (FOV), which can be reduced to the region of interest to perform SAO with the modality of choice. The system performance was demonstrated on wild type mice (C57BL/6J), and transgenic mice with GFP labeled cells. SAO SLO was used for imaging microglia (Cx3cr1-GFP) over ~1 hour, where dynamics of the microglia branches were clearly observed. Our results also include volumetric cellular imaging of microglia throughout the inner retina.

require direct measurements of the optical wavefront but instead uses an image-based aberration correction approach, such as a multi-dimensional optimization or pupil segmentation [13,14]. SAO methods have the ability to provide depth resolved aberration correction by using images acquired at specific layers within the retina. For example, AO OCT has been demonstrated using en face projections extracted from three dimensional OCT volumes to drive the optimization algorithm on the selected retinal layers [15].
The multi-modal system in this report was designed to incorporate SAO with multiple modalities including Optical Coherence Tomography (OCT), OCT based Angiography (OCT-A), confocal Scanning Laser Ophthalmoscopy (SLO), and fluorescence detection. In this work, we present a compact lens-based design of a imaging system for multi-purpose imaging of the small animal retina, which has significantly improved performance and functionality since previous reports [15][16][17][18]. The en face and cross-sectional imaging enable visualization of the retinal structure while the fluorescence imaging has the ability to visualize the biological function of the retina through labeled reporter cells. OCT and SLO can be combined to employ a multi-modal system for simultaneous and co-localized structural and functional imaging. We present representative images and analyses to demonstrate the performance, versatility, and usability of the system for small animal imaging. Images acquired prior to SAO aberration correction demonstrate the widefield and standard resolution imaging in a mouse eye. After performing SAO optimization, our results demonstrate high resolution imaging featuring in vivo volumetric and time-lapse imaging of fluorescently labeled microglia.

Optical design
A schematic of the optical layout of the system is presented in Fig. 1(a). The system components were assembled with off-the-shelf optomechanics and custom mounts designed with SolidWorks (Dassault Systèmes, France) as shown in the Fig. 1(b). The light sources for the imaging system included a near infrared (NIR) Superluminescent Diode (SLD, BLM2-D, Superlum Diodes Ltd., Ireland) for OCT using a central wavelength of 840 nm with a spectral bandwidth of ~80 nm, and a 488 nm laser (0488L-13A, Integrated Optics, Lithuania) for confocal SLO and fluorescence excitation.
The OCT subsystem was based on a fiber optic Michelson interferometer. The OCT light was split by a 70:30 single mode optical fiber coupler (AC Photonics Inc, CA, USA). The 70% portion of the light was connected to a reference arm consisting of a fiber collimator, a dispersion compensation block and a mirror. The OCT probe beam was the 30% portion of light from the coupler, which was launched from a reflective collimator (RC04FC-F01, Thorlabs Inc., NJ, USA) and transmitted through a cold mirror (ZT670rdc-xxrxt, Chroma Technology Corp, VT, USA) for combination with the 488 nm SLO light.
In the SLO subsystem, another reflective collimator (RC08FC-F01, Thorlabs Inc., NJ, USA) was used to launch the SLO light from a fiber with a polarization controller, such that the horizontally polarized light was reflected from a Polarization Beam Splitter (PBS, PBS251, Thorlabs Inc., NJ, USA). The light was then reflected from a dichroic mirror (ZT405/488/561rpc-UF1, Chroma Technology Corp, VT, USA) to the cold mirror, and then co-aligned with the OCT light.
The first pupil plane of both subsystems was defined by the Variable Focus Lens (VFL, Arctic 39N0, Corning, NY, USA) with an aperture of 3.9 mm. The imaging beams were relayed and magnified to a continuous membrane DM (DM69, Alpao, France) with a 10.5 mm aperture, and then to a mounted pair of Galvanometer-scanning Mirrors (GM, 6210H, Cambridge Technology Inc., MA, USA) with a clear aperture of 3.0 mm. Finally, the light was reduced to a beam diameter of 1.0 mm to be focused by the mouse eye and relayed from the GM to be scanned across the retina with a maximum scanning angle of 50 degrees. The optical relays were constructed using achromatic doublets with an antireflection coating for both visible a Each relay us L6 in Fig. 1 ) .
oints. For rection to view were software emission cused into rected the pan). The back-scattered 488 nm laser light was reflected from the dichroic mirror, transmitted through the PBS, focused into a multimode fiber with a core diameter ~5 ADD or ~20 ADD, and detected by another PMT (H7827-002, Hamamatsu Photonics K. K., Japan). We used a 5 ADD fiber core when performing image-based SAO with the back-scattered images, or else we used a 20 ADD fiber core, which provided the higher SNR for general navigation on the mouse retina. The PMT signal gain could be adjusted on the power supply depending on the amount of signal from the sample. The digitization (PCIe-6361, National Instrument, Austin, TX) of the PMTs was synchronized to the acquisition of the OCT A-scans for simultaneous imaging, otherwise the SLO could be operated alone at a 2 kHz line rate with a sampling density of 400 x 200 points.

Sensorless adaptive optics
The SAO could be performed on the en face projection of the OCT volumes, the backscattered confocal SLO, or the fluorescence SLO images. We implemented a hill climbing Coordinate Search (CS) algorithm similar to our previous work [16,23], which provided an exhaustive search to find the optimal Zernike coefficients. The merit function for optimization was determined by the highest image sharpness (S img ) [24,25], defined by the sum of the intensity squared of each image pixel I(x,y) in Eq. (1). .
The CS algorithm started with a flat wavefront with an RMS ~0.05 µm, which was calibrated using a SH-WFS in the location of the GM scanners. Then, for the first mode ( k ) in a sequence, a range of coefficients ( ± α) was applied to the DM. The coefficient ( * n a ) that corresponded to highest metric value was applied to the DM and the algorithm moved onto the next mode. For the first iteration, the sequence of modes began with a defocus (k = 4) value, then the astigmatisms, and continuing in ascending order up to the 21st mode for a total of 18 modes. The Zernike polynomials were ordered and reported using the mode number according to the OSA/ANSI standard [26]. The sequence of 18 modes was usually repeated for multiple iterations, typically 2 or 3 times, until the metric value no longer increased. Successive iterations would search coefficients ranges ( ± β) around the previously selected coefficients. Between iterations, the imaging FOV or location could be adjusted, as the features of interest became visible.
For high resolution imaging, SAO could be used to correct wavefront aberrations from the mouse eye using the output from of the different imaging modalities for the image-based optimization. During optimization, the sampling density of the OCT was decreased to 1024 x 400 x 20 which resulted in 19 seconds for each iteration of the optimization. When the SLO was used for optimization, the sampling was set to 400 x 100 and each iteration took a total of 12 seconds.

Animal handling
The animal imaging sessions were performed under protocols compliant to the Canadian Council on Animal Care and the approval of the University Animal Care Committee at Simon Fraser University. The mice were anesthetized with a subcutaneous injection of ketamine (100 mg/kg of body weight) and dexmedetomidine (0.1 mg/kg of body weight). A drop of topical solution (Tropicamide, 1%) was applied to dilate the ocular pupils. A rigid 0-Diopter contact lens was placed on the animal eyes to prevent the cornea from dehydration and then the animal was aligned without any further contact to the imaging system [27]. For fluorescein angiography, the mice were subcutaneously injected with 100 µL of 100 mg/mL fluorescein. Mice were purchased from the Jackson Laboratory, including wild type strain (C57BL/6J) and transgenic strain with Enhanced Green Fluorescent Protein (EGFP) labeled retinal ganglion cells (Tg(Thy1-EGFP)MJrs/J) and microglia (B6.129P-Cx3cr1 tm1Litt /J).
For retinal imaging, the OCT imaging light was limited to 620 µW. The SLO imaging light did not exceed 230 µW in this report and was limited to 100 µW when operating simultaneously with the OCT.

Image processing
Images in this work were generated by standard post-processing techniques, including steps to register and align frames to a template for averaging, using a combination of Matlab (MathWorks Inc, MA, USA) and ImageJ (National Institutes of Health (NIH), MD, USA) toolkits. The number of volumes and frames that were saved could be easily changed in the acquisition software. For the images presented in this work, we used the following parameters: for OCT, we recorded 5 volumes per acquisition in 4 seconds; for OCT-A images, only one volume was recorded per acquisition in 1.6 seconds; and for SLO, we recorded 50 to 100 frames per acquisition in 5 to 10 seconds. The OCT B-scans were aligned with a vertical translation to remove axial motion of the animal. Most of the B-scans presented in this report were an average of 5 adjacent B-scans within one volume with an exception that is explained in the results section.
The en face OCT images were generated using a Maximum Intensity Projection (MIP) between two manually selected horizontal lines corresponding to depths in the retina. Then, the en face OCT projections and the SLO frames were processed with the following procedure: 1) The registration process was initialized by manually selecting a single frame as the template to align the other frames; 2) Each frame was globally translated horizontally and vertically to maximize the cross-correlation with the template; 3) The frames were broken up into horizontal strips and each strip was translated horizontally and vertically to maximize the cross-correlation with the template [28,29]; 4) The frames were non-rigidly aligned to the template with a sum of squared differences similarity metric along cubic B-splines using the Medical Image Registration Toolbox (MIRT) [30]; 5) After registration, the frames were averaged and the black and white thresholds were adjusted to enhance the image brightness and contrast for presentation. All the B-scans in this report are presented in a linear intensity scale; 6) The images were scaled so that the vertical and horizontal dimensions have the same scale.
SLO frames from the back-scattered and fluorescence channels were simultaneously acquired, which would allow for co-registration if the fluorescence signal was insufficient [31]. However, in this work, the fluorescence images had sufficient signal to use directly for registration.

Imaging without adaptive optics
For imaging large retinal features, a widefield image is preferred and it may not be necessary to perform SAO. Figure 3(a) demonstrates a 50-degree OCT B-scan and a 44-degree en face projection of the Outer Plexiform Layer (OPL) of a wild type mouse retina. Unlike the other B-scan images in this report, in Fig. 3(a), the vertical scanning was disabled and 200 B-scans were acquired, aligned, and averaged. In Fig. 3(b) and (c), the sampling density is increased with a 22-degree FOV and the focus was shifted with the VFL from the OPL to the Nerve Fiber Layer (NFL). The B-scans and en face images were registered and averaged as described in the previous section. The location of the B-scan is indicated by the red dashed line.  uiring OCT vo rated at a faste back-scattered the NFL and m an average o acquired from lexiform layer in Fig. 4(b), wh 4. Confocal SLO im fiber layer from images of three dif -scans were cr ocation. Figure  r in Figure 6  We perfo discriminate distinct vascu a MIP, with th the axial direc umetric imagin hich are located ns between the ed by the stru w. Figure 8( Fig. 11(b) owth and isition.
The centra with an avera #2) that retrac 49. The micro 50 minutes bu left (white arr minute 24.
We perfor in order to inv reduced to 10 proceeded for points before and color-cod

Discussio
In this report functionality imaging moda with fluoresc elements, wh the-art AO im reported syste vivo imaging lapse imaging The imag represents abo calculated re al microglia in ge velocity of cted 38 µm, wi oglia branch on ut also had per row #4) appea rmed further ti vestigate the po 00 µW for 39 m r another 50 m and after the l ded image at th 12. (a) Confocal S ed microglia in th me-lapse video fro microglia images c th and retraction. S on t, we have dem for vision sci alities include ence detection ich include the maging for th ems for each in abilities that g of microglia c ging system pr out half of the solution of ~n Fig. 11 hich only still has a ality was sufficient for clearly imaging the microglia branches and to report metrics, such as movement speeds, yet maintains good quality imaging without requiring AO for imaging large features. The system was initially designed and tested for mouse imaging; however, it is also capable of imaging the rat retina as well, which is often required by many vision researchers for longitudinal studies [33]. Since the rat eye is larger than the mouse eye, this decreases the maximum attainable resolution. However, it was still beneficial to have the SAO to correct for aberrations.
During the time-lapse imaging of microglia cells, we only illuminated the retina with 488 nm since we did not require the use of a beacon for WFS measurements. The microglia timelapse in Fig. 11 appears to have more retraction than the microglia time-lapse from Fig. 12, despite the increase in laser intensity. It is possible that this was normal microglia surveillance of a healthy retina or a response to the 488 nm imaging light. If the 488 nm imaging light itself has an effect on microglia, then it may be difficult to conclude the reason for a microglia response when investigating their role in immunity studies. There is no established maximum permissible exposure (MPE) for the mouse eyes; however, other groups have scaled the MPE for SLO in human eyes [5,29,34]. The MPE for human SLO imaging decreases with imaging FOV [35], so as we image small features in small animal experiments, it will be important to continue to consider laser irradiance as a potential factor.
The imaging system was designed to be used by a non-specialist and future improvements could improve the reliability and robustness of the SAO. For example, a Region of Interest (ROI) within the display could be selected by the user instead of reducing the entire imaging FOV, which further increases the exposure during the ~10 to 20 seconds required for the optimization iteration. Real-time image tracking on the ROI would also enable the optimization algorithm to follow an object of interest or reject frames with a large amount of motion artifact [36][37][38]. In this work, we were using a multi-iteration exhaustive search, which was robust to the occasional motion artifact over the ~30 to 60 seconds required for the entire optimization. However, accurate image tracking would encourage the use of faster optimization algorithms, such as model-based approaches [39,40] that require much fewer measurements, thereby decreasing optimization time. This would be advantageous to further reduce the exposure of the entire imaging process and the potential for damage over time.
In this work, we optimized up to the 21st Zernike mode for 18 modes in total. The improvement in the image quality after each mode is optimized is represented in Fig. 7(b), which demonstrates that there is an increase in the metric value in the 5th radial order in the first and second iteration. Using higher orders in the optimization algorithm could improve results but it would come at the cost of algorithm time. Since time is limited for in vivo imaging, the algorithmic execution time is better spent on further iterations [41]. For example, the step sizes between coefficients can be reduced to improve the wavefront correction. Furthermore, successive iterations have an improved SNR, which will also improve the performance of the AO correction.
In conclusion, we have demonstrated a lens-based system, capable of high-resolution en face small animal imaging with multiple modalities. The compactness and simplicity of the system allow for the potential translation to vision scientists that require tools for in vivo and longitudinal studies. Our results demonstrate the potential for studying individual cells, such as RGCs and microglia, in healthy and diseased animal models.