Temporal speckle-averaging of optical coherence tomography volumes for in-vivo cellular resolution neuronal and vascular retinal imaging

Abstract. It has been recently demonstrated that structures corresponding to the cell bodies of highly transparent cells in the retinal ganglion cell layer could be visualized noninvasively in the living human eye by optical coherence tomography (OCT) via temporal averaging. Inspired by this development, we explored the application of volumetric temporal averaging in mice, which are important models for studying human retinal diseases and therapeutic interventions. A general framework of temporal speckle-averaging (TSA) of OCT and optical coherence tomography angiography (OCTA) is presented and applied to mouse retinal volumetric data. Based on the image analysis, the eyes of mice under anesthesia exhibit only minor motions, corresponding to lateral displacements of a few micrometers and rotations of a fraction of 1 deg. Moreover, due to reduced eye movements under anesthesia, there is a negligible amount of motion artifacts within the volumes that need to be corrected to achieve volume coregistration. In addition, the relatively good optical quality of the mouse ocular media allows for cellular-resolution imaging without adaptive optics (AO), greatly simplifying the experimental system, making the proposed framework feasible for large studies. The TSA OCT and TSA OCTA results provide rich information about new structures previously not visualized in living mice with non-AO-OCT. The mechanism of TSA relies on improving signal-to-noise ratio as well as efficient suppression of speckle contrast due to temporal decorrelation of the speckle patterns, enabling full utilization of the high volumetric resolution offered by OCT and OCTA.


Introduction
The retina is a visually accessible portion of the central nervous system situated at the focal plane of a natural optical system and therefore affords a unique opportunity for noninvasively visualizing the changes associated with neurodegenerative disorders and vascular alterations of neural tissue. 1,2 Recent progress in high-resolution retinal imaging has achieved sufficient resolution and sensitivity to take full advantage of this opportunity in humans. 3 The opportunity also applies to the mouse, a primary animal model for studying human biology and disease. 4 Mouse models are important not only in basic vision science and ophthalmic research, 5 including studies of the mechanism of phototransduction, 6,7 but also in general neuroscience. 8 Interestingly, the special optical features of the mouse eye, which has a numerical aperture (NA) twice than that of the human eye, 9 offer a potentially higher resolution for retinal imaging. 10 However, the retinal neurons anterior to photoreceptors-i.e., bipolar, horizontal, amacrine, or ganglion cells (GCs) -are nearly transparent, making them very difficult to image in the living eye using imaging modalities employing only labelfree backscattered light.
Scanning laser ophthalmoscopy (SLO) 11 and optical coherence tomography (OCT) 12 are two complementary in-vivo retinal imaging modalities. [13][14][15][16] Both detect backscattered light, and SLO can also detect fluorescence. SLO and OCT have been successful in imaging the photoreceptor mosaic and retinal microvasculature with adaptive optics (AO), 17,18 but only very recently there have been reports of imaging the highly transparent cells in the inner retina. Rossi et al. 19 reported imaging GC somas visualized by AO-SLO, implementing aperture split detection; Liu et al. 20 and Wells-Gray et al. 21 reported visualization of GC somas using AO-OCT with volume averaging. Interestingly, AO-OCT with volume averaging produced images of GCs with unprecedented clarity, allowing reliable mapping of morphometry of retinal neuron somas across the living human retina. 20 There have been many reports of imaging inner retinal neurons in the mouse retina labeled by extrinsic fluorescent proteins or dyes and detected by fluorescence emission SLO. 13 There have also been reports of imaging retinal neuron somas using AO-SLO with split detection like in human studies. 22 However, there have been no reports on the visualization of inner retinal neurons based on volume averaging of OCT or AO-OCT data from the mouse eye. Considering the important role of the mouse model in biomedical research, the ability to visualize these cells is clearly desirable.
In this paper, we present a general framework developed in our laboratory which uses mouse retinal OCT/optical coherence tomography angiography (OCTA) to achieve high-contrast volumetric images of inner retinal neurons (with OCT) and corresponding microvasculature (with OCTA). After registration and application of temporal speckle-averaging (TSA) method, cellular-level resolution volumetric neurovascular retinal mapping is achieved. The key to successful TSA lies in using an OCT/OCTA volume registration method that corrects for mouse eye movement occuring during imaging under general anesthesia. We develop a stepwise method to register a series of successively acquired three-dimensional (3-D) OCT volumes using three consecutive two-dimensional (2-D) registrations with reliable and robust registration tools available in ImageJ. 23 B-scan and en face images from both OCT intensity and OCTA 24,25 volumes are directly compared side by side to illustrate the effect of speckle-averaging, which is further quantified by various image metrics and analyses. Finally, two examples, one reconstructed from 150-OCT-volumes of a pigmented (C57Bl/6J) mouse and the other from a 50-OCTA-volumes of a B6 albino mouse, are presented. Two important retinal structures, retinal neuron somas and the choriocapillaris, are visualized for the first time in the living mice without extrinsic contrast agents.

Methods
Unlike in humans, in-vivo animal retinal imaging is often performed under anesthesia to reduce unnecessary stress. Common problems associated with long-term imaging of anesthetized animals include tear film breakup, drying of the corneas, and development of cold cataracts. These effects can be overcome by the use of a contact lens and transparent gel to maintain the cornea clear and moist, 13,26 and by maintaining and monitoring body temperature during imaging. Most eye movements, including involuntary drift and saccades, are not present in mice under anesthesia. 27 Despite the reduction in eye movements, some eye motion, caused by animal breathing and consequent head movements, occurs with consequent image artifacts during retinal imaging. We observed that these residual motions are mostly small (with noted exceptions) and give rise to distortions that can be characterized as rigid-body displacements between volumes, which can be corrected with a simple and robust stepwise correction method employing three serial 2-D registrations in two separate image planes.

Mouse Positioning and Anesthesia
All mice husbandry and handling were in accordance with protocols approved by the University of California Animal Care and Use Committee, which strictly adheres to all National Institutes of Health (NIH) guidelines and satisfies the Association for Research in Vision and Ophthalmology guidelines for animal use. Adult pigmented C57BL/6J and B6 albino mice were obtained from Jackson Laboratories and maintained on a 12:12, ∼100-lux light cycle. During the experiments, mice were anesthetized with the inhalational anesthetic isoflurane (2% in O 2 ), and their pupils were dilated with medical grade tropicamide and phenylephrine. A contact lens and gel (GelTeal Tears, Alcon) were used to maintain the cornea transparency for in-vivo retinal imaging. 26,28 A commercially available bite bar (Kopf Instruments, Inc., Model 1923-B Mouse Gas Anesthesia Head Holder) was used to stabilize the mouse head and deliver the oxygen and isoflurane.

University of California Davis Multimodal SLO/OCT Mouse Retinal Imaging System
A combined SLO/OCT system was used for imaging the mouse retina. 13 The beam at the mouse pupil was 0.5 mm, which corresponds to a NA of 0.12 for the imaging system. 29 The light power is 100 μW for SLO (488 nm laser, Coherent) and 600 μW for OCT (center wavelength 860 nm with a bandwidth of 132 nm, SLD, Superlum Ltd.). A custom spectrometer using Balser line scan camera with 2048 pixels was used to acquire the OCT spectra. OCT spectra were acquired at a 100-kHz A-scan rate using custom Labview software that can operate in continuous volume acquisition mode. The volume acquisition time was ∼3 s for OCT (512 A-scans/B-scan, 512 B-scans/volumedata in Fig. 6) and ∼6 s for OCTA (540A-scans/B-scan, 1080 B-scans/volume-data in Fig. 7). The OCT volumes were saved at 5 volumes= min (ΔT OCT ¼ 12 s), and OCTA volumes were acquired at 2.5 volumes= min (ΔT OCTA ¼ 24 s). Initial postprocessing was implemented by custom Matlab™ code implementing standard OCT processing, including direct current subtraction, wavelength-to-k (wave number) resampling and interpolation, dispersion compensation, Hann windowing, and FFT. 30,31 The OCTA data were processed by a custom phase-variance OCT (pv-OCT) C++ code to visualize blood flow and generate OCTA B-scans. 32,33 The serially recorded volumetric data sets, including intensity and OCTA data, were further coregistered to correct for motion artifacts and allow averaging.
The serial volume registration was implemented by image processing in MATLAB and ImageJ. As described in detail below, serially collected volumetric data set were initially processed with the custom MATLAB code for registration, and then the ImageJ "MultiStackReg" and "StackReg" plugins were applied to coregister prearranged series of stacks.

Retinal Image Movements in Anesthetized Mice during Imaging
Although the general anesthesia used during retinal imaging greatly helps to reduce eye movements that distort images, the remaining eye movements are not negligible, especially for long acquisition times if cellular resolution averaging is desired.
Here we present a simple observation of the retinal movements that occur in anesthetized animals and their effects on volumetric image data. Figure 1(a) shows the Cartesian coordinate system on the retina used to represent the 3-D data. Here the x and y axes correspond to OCT galvo-scanners' fast and slow scanning directions, respectively, while the z axis is the OCT depth or A-scan direction (along which data are acquired within the pixel dwell time of 10 μs). The fast OCT x-scanner, which generates B-scans, scans at ∼200 Hz, while the slower y-scanner, which shifts the B-scans to generate the volumes, sweeps across the image field at ∼0.3 Hz. Figure 1(b) depicts the most common image displacements that we observe. Slower motion artifacts that arise from small eye drift result in translation ðx; y; zÞ and rotation ðφ; θÞ between successive volumes, and can be easily corrected by rigid-body transformation using ImageJ stack registration. We previously used a similar framework to correct for motion artifacts within a single volume in human retinal imaging with AO-OCT, 34 but here we apply a Neurophotonics 041105-2 Oct-Dec 2019 • Vol. 6(4) multistep process to achieve volume-to-volume coregistration.
In addition to the relatively small volume displacements just described, there are also rarely occurring distortions that create scaling changes between volumes [ Fig. 1(c)]; these rare distortions are produced by rotations of the eye along the x or y axis or axial displacement of the eye as a whole. Unlike in most human retinal imaging, a contact lens and gels were used in our mouse retinal imaging system, which resulted in the creation of more degrees of freedom for image distortions. For example, moving the contact lens 1 mm away from the cornea (resulting in gel thickness changes) can easily increase the scaling factor to 30%. 26 If the gel thickness changes unevenly across the pupil (e.g., during eye rotation along x or y axis one side becomes thicker than the other side), nonisotropic scaling between x and y axes will occur. Although such displacements are rare, they can nonetheless be detected in the image analysis and corrected by the ImageJ affine transformation. However, based on our experience the most efficient approach is to examine registered volumes using en face intensity projection after first correcting them for the common motion artifacts, and simply discard these volumes whose en face projections show scaling deformations, rather than to use more complex registration. The interval between mouse breaths can be ∼1 to 3 s, and the eye movements caused by the mouse breathing will be recorded during the OCT volume acquisition [ Fig. 1(d)]. We noticed that the duration of the eye movement during breathing is much shorter than the interval between breaths; thus, only small portion of the data is distorted [ Fig. 1(d) middle and left panels]. These distortions do not affect the registration between en face images and are not noticeable after averaging. Thus, the image distortions caused by mouse breathing were ignored in our method.

Correction of the Mouse Retinal Movements Present during Anesthesia Using Stepwise OCT/OCTA Volume Registration
To correct the retinal motion artifacts described above, a method of image registration between consecutive volumes is needed. However, no broadly accepted serial volume registration is commonly available. Thus, custom volume registration software had to be developed. In developing, we tried to use well vetted and simple freely available registration tools. We found that one of the simplest and most reliable 2-D image alignment tools in ImageJ-StackReg plugin 35 -and its extensions such as MultiStackReg 36 work very well for registering successive B-scans, as well as en face projections of OCT Volumes.
Notably, the MultiStackReg plugin aligns stacks of images to one single reference image, thereby eliminating the error accumulation that occurs with the StackReg plugin. Thus, we implemented a stepwise 2-D registration for serial OCT/OCTA volume registration. Importantly, when several imaging channels are acquired simultaneously (e.g., OCT intensity and OCTA), one needs to be able to align the data to one of the channels only (e.g., OCT intensity) and then apply the correcting transformations directly to the other channels. This important feature is also available in MultiStackReg plugin. We now describe the stepwise volume registration method, which uses a series of 2-D registrations to correct the motion artifacts between volumes (Fig. 2). After acquisition of a series of OCT volumes, for each OCT volume we first create an en face intensity projection image from the whole volume or, in case when the vascular features are not visible enough, projection from the inner retina [from outer plexiform layer (OPL) to nerve fiber layer (NFL)] was used to improve the overall registration performance. Then we use the MultiStackReg ImageJ plugin to register the full set of en face images to a reference image from the middle of the series. Because the intravolume distortion only occupies a small portion of the data, the single middle frame (in the middle of the time series stack) is usually sufficient to be selected as a proper reference. The transformation matrix for the registration of each en face image was saved, and then applied to the entire volumes in the OCT series (including the intensity and other channels, e.g., OCTA). After registration and motion correction with the transformations, the volumes were next cropped to maintain only the overlapping regions, as shown by the smallest rectangle in Fig. 2(b); this removes empty regions to allow efficient second step of registration between the B-scans. In the next step, a temporal series of B-scans from the same y location in the series of OCT volumes was extracted and enclosed in a stack for registration (BM stack). Then, the motion errors between successive B-scans in the BM stack were corrected with MultiStackReg and the transformations were recorded; the same transformation matrices were then applied to all other channels of the data sets (e.g., OCTA). Specifically, this latter step must be performed for each B-scan within the serial OCT volumes [steps in Figs. 2(d)-2(e)] until the entire overlapping volume region has been corrected. After BM-stack registration, we then averaged the series of B-scans from each y location to achieve the TSA volume. However, it was noted that in the TSA volumes there were sometimes slight offsets (possibly arise from intravolume motions caused by mouse heartbeat or scanners off-axis vibration) between adjacent B-scans [as shown in Fig. 2(g) top image, arrows]. Usually, a simple registration using StackReg plugin with intensity OCT data assigned to one color (e.g., red) channel and OCTA information assigned to another color (e.g., green) channel was performed to correct for the fine position errors between adjacent TSA B-scans. This allowed the production of a smoothed TSA volume of OCT intensity and/or OCTA data. The core of this method relies on three serially performed registrations, marked by three dash-dotted rectangles in Fig. 2.

Effects of TSA on OCT Data Sets
The average volume created by TSA can be expected to provide improved visualization of weakly scattering structures in the 3-D sample volume on the basis of general principles. For example, if backscattering from a structure spanning several adjacent voxels results in an average of n photons collected by the detector, Poisson fluctuations of magnitude ffiffiffi n p from voxel to voxel could make the structure undetectable in a single volume. TSA with N volumes would reduce the fluctuations to ffiffiffiffiffiffiffiffiffi n=N p , while the signal from all voxels remains n, so that TSA results in increased sensitivity and dynamic range of OCT intensity images. Averaging without the precise alignment provided by TSA, however, would not confer these benefits because the nominally identical voxel in a series of poorly aligned volumes would have variations arising from sample-to-sample variation in the underlying structure due to motion. What would result in a reduction of both resolution and speckle contrast? Interestingly, averaging without the precise 3-D alignment provided by TSA, but with the simple 2-D BM-scan registration as it is done in most commercial clinical OCT systems, still confers most of the TSA benefits. This is due to the fact that even imperfectly coregistered B-scans in y axis (slow scanning direction) have nominally similar voxel values in the same retinal layer. This allows achieving speckle contrast reduction and signal-to-noise ratio (SNR) benefit of averaging but at the cost of a reduction in lateral resolution. 37 There are additional benefits of TSA that goes beyond this classical improvement in image sensitivity and dynamic range. One arises from improved phase stability associated with each voxel: this increases the fidelity of the OCTA data set. 24 A second is a visible reduction of speckle contrast, which greatly improves the visibility of cellular and subcellular features on OCT intensity data. 38 To quantify these benefits, we calculated cross correlations between intensity BM stacks to estimate the time needed for full decorrelation of the speckle fields. We also measured changes in speckle contrast and intensity histograms for TSA performed on a large number (N ¼ 150) of volumes. Here, we introduce three parameters that are used to quantify changes in speckle contrast.

Contrast-to-noise ratio
To estimate both effects of TSA, the improvement in SNR, and reduction in speckle contrast, we use the contrast-to-noise ratio (CNR). Here the intensity from a region of interest (ROI) from the retina was assigned as a signal, and a background area from the vitreous was defined as the background (noise). 39 Then, the CNR was calculated as E Q -T A R G E T ; t e m p : i n t r a l i n k -; e 0 0 1 ; 3 2 6 ; 5 6 8 CNR ¼ where I ROI and I bg are the mean values of the intensity within an ROI and background regions, respectively, and σ ROI and σ bg are their corresponding standard deviations (SDs).

Normalized speckle contrast
Normalized speckle contrast (NSC) was used to quantify the speckle contrast suppression effect due to temporal averaging. NSC is defined as E Q -T A R G E T ; t e m p : i n t r a l i n k -; e 0 0 2 ; 3 2 6 ; 4 1 6 where σ I is the SD of the image intensity (I) in a given region and I is its average. 40 The averaging of independent speckle patterns is widely understood to be the primary mechanism by which speckle is reduced. If in a depth-restricted portion of a B-scan, the intensity of all pixels is independent and identically distributed, and M uncorrelated speckle patterns are averaged, the speckle contrast of the average will obey the following general relation: 40 E Q -T A R G E T ; t e m p : i n t r a l i n k -; e 0 0 3 ; 3 2 6 ; 2 8 8 While this effect is related to the general effect of averaging described above, it is distinctive, because speckle contrast is voxel-to-voxel variation arising from scattering entities smaller than the 3-D point-spread function (PSF). 41 The scan-to-scan variation in speckle underlying Eq. (2) would arise if the number of scatterers in each PSF volume or their mutual interference fluctuated from volume to volume about the same mean value. Here we assume that the variance in intensity arising from speckle is much higher than the one arising from the noise.

Effective number of uncorrelated speckle patterns (m)
Actually, the mean intensity of speckle from voxels at different locations will vary in a sample. In this case, the speckle intensity distribution can be described as a case of the full where I sp represents the possible values of intensities in the fully developed speckle pattern. Here α and μ are the shape parameter and expected value, respectively, of the local mean intensity variation (assumed to follow a gamma distribution), K α ðxÞ is the modified Bessel function of the second kind with order α, and m is the number of uncorrelated speckle patterns. 42 We fitted OCT intensity histograms of TSA scans in the BM stack with this function to estimate the number of uncorrelated speckle patterns in the original temporal series.

Results
In this section, we provide the results of our analysis of mouse eye movements during anesthesia, followed by a comparison between single and TSA OCT intensity and OCTA data sets. Then we quantify the effect of TSA on CNR improvement and speckle contrast reduction. Finally, examples of applying TSA to OCT volumes of pigmented (C57Bl/6J) and B6 albino mice are provided.

Experimental Analysis of Mouse Retinal Movements during Serial Volume Acquisition
To characterize the motion artifacts (described above, Fig. 1) that underlie motions in the images of anesthetized mice, we acquired and subsequently analyzed a multimodal volumetric data set centered on a single fluorescent retinal ganglion cell (RGC) (Fig. 3). This thy1-YFP-16 transgenic mouse (The Jackson Laboratory) expresses a yellow fluorescent protein (YFP) at high levels in motor and sensory neurons, as well as in subsets of central neurons. The complex, unique pattern of the RGC's submicrometer fluorescent dendrites provides an ideal target for examining motions over an extended imaging session. During the ∼14min imaging session, 400 consecutive SLO images and 73 consecutive OCT volumes were simultaneously collected at the same retina location, which scanned ∼300 × 300 μm 2 with 400 × 400 pixels sampling. To illustrate the effect of motion between volumes, Fig. 3(a) displays the SLO fluorescence images of a single RGC color-coded by time: motions are manifest as blurring of the image and color fringes. Figure 3(b) shows the result of combining the two images after assigning them with red and green for rigid body and affine registration images, respectively. The improvement in sharpness reveals that these registrations have removed much of the motions of the unregistered average image [ Fig. 3(a)], while the uniform color of the combined image indicates that there was little difference between corrections obtained with the two registration methods. Figure 3(c) shows an image of the absolute value of the pixel-by-pixel difference between the images obtained with two different registrations, further revealing that there are no major differences. Nonetheless, a contrast-stretched 32-bit difference between them reveals a small residual difference [ Fig. 3(d)]. The pseudo-3-D relief appearance of the cell dendrites is due to slightly sharper and brighter local fluorescence after registration by rigid-body transformation as compared to that with affine. Figure 3(e) provides the averaged line profiles from nine different locations across the GC dendrites [ Fig. 3(b) arrows] extracted from two averaged images, one obtained by the rigid body and one by affine registrations. Note that rigid-body registration creates a sharper line profile. This result also suggests that the shadow near the dendrites seen in Fig. 3(d) is mainly caused by the difference in sharpness and brightness, and not by a lateral mismatch between two averaged images. This comparison also reveals that there is a negligible scaling transformation [ Fig. 1(c)], because rigid-body correction (no scaling correction) works better than affine correction (correction with scaling transformation).

Direct Comparison of Images from Single and Temporal Speckle-Averaged OCT Volumes
Because OCT involves the interference of partially coherent light, OCT images necessarily have high speckle contrast, which arises from scattering particles in the sample that are smaller than the 3-D PSF. 41 Speckle contrast degrades image quality, making it impossible to resolve fine structures near diffraction-limited resolution. Despite its spatially random appearance, the speckle pattern of a nonvarying sample is fully determined by the underlying spatial distribution of scatterers and would not vary from scan to scan if nothing in the sample moves. However, in-vivo sub-PSF scale scattering particles (e.g., mitochondria and other organelles) are not static, resulting in variation in the speckle pattern between consecutive volumes. 44 At the same time, larger morphological structures such as cell bodies and vasculature are static on the timescales of image acquisition, allowing image registration. As a consequence, averaging of serial volumes reduces speckle contrast, while maintaining structural information from objects larger than the PSF, yielding images with improved SNR and image contrast. Comparison of single and TSA volumes provides a means of visualizing the speckle-reducing effect of averaging (Fig. 4). Figure 4 shows the comparison between single OCT and OCTA volume and TSA OCT and OCTA volumes (from 150 and 50 serially acquired OCT and OCTA volumes, respectively). In the TSA OCT B-scan, one can distinguish the walls of two large blood vessels in the NFL (arrows). The TSA OCT en face image clearly shows much better blood vessel contrast (including blood vessel walls, arrows), nerve fiber bundles, and novel structures resembling cell somas (arrowhead). TSA also greatly improves OCTA data. A comparison between a single volume and TSA OCTA is shown in Figs. 4(e)-4(h), where even smallest capillaries are clearly visible in the OCTA B-scan (red arrows). Similarly, the en face projection of capillaries from OPL shows great improvement of signal (capillaries)-to-noise (random noise between capillaries) ratio. For example, even some parts of the capillaries hidden in the shadows of larger NFL blood vessels in single volumes appear continuous (arrows) in the TSA volume.

Quantification of the Effect of Temporal
Speckle-Averaging As described above, the improved cellular contrast offered by TSA relies on averaging uncorrelated speckle patterns to suppress the speckle contrast and improve the SNR of resulting images, what can be quantified by CNR. To estimate the effect of TSA, we analyzed OCT intensity data from the experiment, as shown in Figs Fig. 4(a): while generally true that the longer time interval between two B-scans, the lower the correlation, the results suggest that the maximum decorrelation is achieved in about 2 min. While this implies that waiting 2 min between scans might be optimum for reducing speckle, it needs to be kept in mind that TSA of volumes obtained more frequently, nonetheless, improves the visibility of weakly scattering structures according to 1= ffiffiffiffi N p , as discussed above. This decrease in the noise of the background σ bg as a function of a number of averaged volumes, N, is shown in Fig. 5(c). However, the actual data visibility defined by CNR in Eq. Figures 5(f) shows the estimation of uncorrelated speckle patterns in averaged B-scans implementing fitting of the full K-distribution to intensity histograms [Eq. (4)]. As expected, the number of uncorrelated speckle patterns increases steadily with N but clearly deviates from the theoretical limit. The potential benefit of increasing the time interval between acquisition of consecutive volumes, what results in number of uncorrelated speckle fields, is also presented (square and star). Figure 5(f) also shows the comparison between independent speckle patters present in N-averaged volumes calculated using Eq. (3) for M and Eq. (4) for m. It is evident that simple Eq. (3) underestimates the number of independent speckle patterns compared to results obtained with Eq. (4). This suggests the possible underestimation of the CNR.

Example Results of TSA of OCT and OCTA
To explore the effect of TSA on the visualization of neuronal and vascular structures in mouse retinal OCT data sets, we applied our technique to imaging pigmented C57BL/6 and B6 albino mice, which are two commonly used mouse genotypes. The C57BL/6 mouse line is one of the most popular mouse strains used in biomedical research. Addition of B6-albino mice allows the investigation of the effect of melanin on retinal data acquired with TSA OCT. Lack of melanin allows the imaging light to penetrate deeper into the choroid in albino mouse, which makes the imaging of choroidal vasculature less challenging than it is in pigmented C57BL/6 mice.

TSA of 150 intensity volumes from a C57BL/6 mouse retina
To further explore the value of TSA in retinal imaging, in Fig. 6, we present the results from applying TSA to 150 serially collected OCT volumes from a region of the retina of a pigmented (C75BL/6) mouse spanning a 600 × 600 μm 2 and sampled by 512 × 512 pixels. The OCT beam was focused on the superficial retinal layers to better visualize inner retina. Figure 6 shows the results from seven representative layers extracted from the TSA OCT volume, clearly illustrating the improved visibility and contrast. Specifically, the nerve fiber bundles in the NFL, the blood vessels and capillaries in NFL, inner plexiform layer (IPL), and OPL are clearly visible. In addition, one can identify bright circular objects in ONL whose size appears to match the size of the photoreceptor cell bodies by comparing the OCT image with histology data 45    OCTA is a powerful technique that allows the visualization of blood vessels without an extrinsic contrast agent. 25,32 Here we illustrate the application of TSA to serially acquired OCTA volumes. As mentioned above, one of the features of our volume registration method (Fig. 2) is that it allows multichannel registration based on the structural (intensity) OCT data alone. Here we demonstrate this multimodal functionality by coregistering 50 OCTA volumes acquired over 1 × 1 mm 2 with 540 × 1080 (360 × 3) A-scans (Fig. 7). In this data acquisition protocol, three B-scans are collected consecutively at each y position to extract the phase variance that arises from the blood cells scattering as they transit the vasculature. 32,33 For this experiment, a B6 albino mouse was imaged, and the optical focus of the system was set in the middle of the retina near the ONL to achieve good quality imaging of both the retinal and the choroidal vasculature. Similar to the previous example (Fig. 6 Fig. 7(g)], which has not been previously visualized in the mouse in vivo without contrast agent. 46,47 OCTA data sets can be used to produce very high contrast maps of retinal vasculature, as illustrated with a depthcolor-coded projection of the inner retina vessels [ Fig. 7(h)].

Discussion and Conclusions
In this paper, we presented a framework for applying TSA to cellular resolution volumetric neuron and vascular imaging of the living mouse retina. Successful visualization of novel cellular structures was achieved by the application of a stepwise registration method (Fig. 2). The method was built on reliable and robust registration tools available in ImageJ, the common open-source data-processing platform developed by NIH. 23 As a result, the proposed method is simple to adopt if compared to dedicated custom registration, which may lower some technical barriers to its adoption. Although the examples shown in this paper use data from mouse retinal imaging, we believe that this method can be easily extended to OCT volumes acquired in any anesthetized subjects, including other experimental animals and humans (e.g., infants) (the eye movements of nonanesthetized subjects present additional challenges). In addition, our method allows extraction of the retina motions present during imaging, which might help to troubleshoot sources of unwanted motion in serially acquired volumes. In addition to the application of TSA to OCT intensity data sets (Fig. 6), we also showcased TSA of OCTA volumes (Fig. 7). In this case, TSA provided high contrast for retinal vasculature and choroid. To the best of our knowledge, TSA yielded the first contrast-free visualization of the mouse choriocapillaris. The importance of noninvasive choriocapillaris visualization has been pointed out by clinical studies in humans, [48][49][50][51] and the general understanding that this distinctive vascular bed supplies nutrition and oxygen to the retinal pigment epithelium (RPE) and photoreceptors, 52 and that its deterioration may play a key role in the development of many retina diseases including agerelated macular degeneration (AMD). 52 TSA OCTA should provide improved visualization of neovascularization, an important metric for monitoring the progress of AMD toward the "wet" form, and for ocular tumors. 53 TSA greatly improves the image quality by two distinctive but overlapping statistical mechanisms. The first is the general process by which averaging reduces noise-which in general may arise from shot-noise fluctuations in photon capture rates or instrumental noise-while simultaneously increasing the dynamic range of the signal. However, in the case of Fourierdomain OCT detection that is used in acquiring our volumetric data sets, we know that the shot-noise is limiting. [54][55][56] This process enables structures that weakly scatter to become visible. Without TSA, however, sample-to-sample distortions arising from ocular motion can reduce resolution and reduce the effectiveness of averaging.
The second mechanism by which TSA improves image quality is by suppression of speckle contrast. It is increasingly clear that many of the sub-PSF scale scatterers that produce speckle in imaging with partially coherent light are mobile. TSA reduces speckle contrast because of the time-dependent decorrelation of the speckle patterns in the living retina (Fig. 5). Investigation of the decorrelation time for speckle in different cellular and subcellular locations may provide information about the identity and mobility of the underlying scatterers. [57][58][59] While decorrelation times are set by the biological properties of the underlying scatterers, when considered purely as noise, it is possible to speed decorrelation by subtle manipulations of the PSF, as we described recently in an active method that creates uncorrelated speckle patterns with aperture phase modulation (APM) via AO-OCT. 45 This latter technique can more efficiently reduce speckle contrast since it need not wait on the biological decorrelation time but does so at the cost of the increased complexity of the imaging system. The TSA method needs no additional components beyond those available with any mouse retinal OCT system. It is important to mention that decorrelation time for speckle patterns acquired in vivo in clinical AO-OCT systems (NFL, GCL, IPL < 1 s; 58 RPE < 3 s 60 ) could be shorter than the one reported in this paper (τ ¼ 5 to 40 s). There can be several reasons besides biological difference between species that could explain this discrepancy, including: (1) difference in pupil stability between awake human and anesthetize mouse. Our recent study on the influence of ocular aberrations on speckle patterns suggests that variations in pupil wavefront during clinical imaging could result in unintended aperture phase modulations, 45 which could affect speckle decorrelation times. (2) Lateral and axial resolution between OCT and AO-OCT system. Narrower PSFs should be more sensitive to local variations in scattering, resulting in shorter decorrelation time.
(3) The speckle decorrelation times reported in this study may be affected by the relatively slow volume acquisition rates. Clearly, more studies are needed to resolve these differences and weight influence of each factor.
One of the limitations of the registration method as currently implemented is its relatively long processing time, 3 to 4 h for a 50-volume data set. Even though the analysis is mostly automated, the lengthy processing time limits the ability to perform this analysis routinely. The long processing time is a consequence of using open source tools developed for 2-D image registration. Thus, there is a need for a dedicated graphics processing-unit-based image registration process that could, for example, perform parallel registrations of BM stacks. This would greatly reduce the processing time, opening the door to iterative registration that would further reduce motion artifacts.
The data in this paper were acquired using a non-AO-OCT mouse retina system. We expect that this framework can be extended to AO-OCT data acquired over smaller fields of view, usually a few degrees of the visual field, as long as the serial en face intensity projection and B-scans can be registered using the ImageJ plugins. Our retinal neuron soma images have lower contrast than those from human retina acquired with AO-OCT. 16,20 This lower contrast could be because the NA of the mouse retina imaging system used in this paper is only one-half of that used in the human system. In future, we will apply TSA to data sets acquired by a compact AO-OCT system 45 which has an NA of 0.23, roughly equal to the human full-pupil AO-OCT system. This increased NA will improve lateral resolution, and potentially, the cellular contrast in our data. However, the ability to visualize the highly transparent somas of the neurons in the mouse retina without the use of AO makes the TSA very attractive. As an alternative to AO, it is possible to use a somewhat larger imaging beam size on the mouse pupil to increase the system NA, 29 or switch to shorter wavelength (visible light OCT 61,62 ) imaging. Both strategies will predictably improve the system resolution and yield a better contrast of cell somas. Finally, other promising alternatives to hardware-based AO, such as computational AO techniques, 63 could also be combined with our method to provide high-quality retinal images.

Disclosures
All authors declare no conflicts of interest.