Measuring the spatial distribution of multiply scattered light using a de-scanned image sensor for examining retinal structure contrast

: An optical platform is presented for examining intrinsic contrast detection strategies when imaging retinal structure using ex vivo tissue. A custom microscope was developed that scans intact tissue and collects scattered light distribution at every image pixel, allowing digital masks to be applied after image collection. With this novel approach at measuring the spatial distribution of multiply scattered light, known and novel methods of detecting intrinsic cellular contrast can be explored, compared, and optimized for retinal structures of interest.


Introduction
Adaptive optics (AO) ophthalmoscopy enables noninvasive imaging of retinal structure at optical resolutions of 2 µm or less [1,2]. However, little is known about the origin of contrast mechanisms for highly light-scattering retinal structures (e.g. nerve fiber, blood vessels, photoreceptors, pigment). In addition to structural assessment of these microscopic structures, there is increasing interest in functional assessments being developed for AO retinal imaging [3]. Functional assessment requires both resolution and contrast of the microscopic processes being visualized. If we gain a better understanding of how light scattering is contributing to the contrast that makes retinal structures visible, we can more rapidly develop sub-cellular structural and/or quantitative imaging methods to assess retinal function.
Detection schemes that use "multiply scattered", "off-axis", or "non-confocal" light have allowed visualization of vessel walls using "knife-edge" detection [4], inner segments of photoreceptors using "split detection" [5], and retinal ganglion cells using "multi-offset" detection [6]. However, these non-confocal modalities detect a relatively small number of off-axis points and may not take full advantage of the information contained in the spatial distribution of the collected multiply scattered light. In addition, testing different detection schemes can be laborious, requiring detection path realignment prior to each image capture and introducing potentially undesirable temporal confounds to the dataset. Recent studies have taken advantage of fiber coupling with pixel reassignment [7] and spatial light modulation [8] to improve upon some of these limitations and are able to examine additional modes of structural contrast, but the number of detectors for image capture can be limited by space on the device or optical table. Developing a method to allow multiple simulations of image plane detection approaches from a single dataset could complement these efforts to optimize the intrinsic contrast of noninvasive imaging techniques.
Here, we present a de-scanned scattered light detection (DSLD) microscope to assess detection mask geometries that capture multiply scattered light from ex vivo samples, and can subsequently inform strategies for improving the contrast of cellular imaging. This system collects an entire 2D distribution of scattered light surrounding the illumination spot for each pixel in the confocal scanned image. This collection of backscattered and reflected light (epi-illumination) mimics clinical imaging methods used in vivo. This microscope can also be used to collect light passing through the retina (trans-illumination) without scanning to produce wide-field image at fast frame rates to assist focus and navigation. The 2D scattered light distribution from epi-illumination is collected for each pixel in the 2D scanned image of the tissue. Digital detector geometries can be applied after data collection, leading to streamlined contrast exploration and optimizing imaging methods tailored towards microscopic structures of interest. In this work we describe the DSLD system design, data collection, subsequent detector simulation analysis, and provide demonstrations of the DSLD microscope's capabilities to image different retinal structures. We then discuss potential applications for using the DSLD microscope as a platform to improve intrinsic retinal cell contrast.

Instrument design
A custom microscope referred to as the de-scanned scattered light detector (DSLD) system ( Fig. 1(A)) was constructed by placing a CMOS camera (DCC3240N, ThorLabs, Newton, NJ, USA) at the de-scanned confocal plane of custom microscope with a 2D galvanometer scanning system (GVS202, ThorLabs, Newton, NJ, USA) supporting a 5 mm beam diameter at a minimum of 100 Hz or 1 kHz with 0.3 ms settling time for small angles less than ±0.2 • . Since our pixel step size is far smaller (<0.002 • ), the mirrors are fast enough to support step and hold during camera integration rather than using a continuous scan. This system uses a 780 nm light emitting diode (M780L3, ThorLabs, Newton, NJ, USA) for trans-illumination, and an 800 nm super-luminescent diode (MOPA-SLD-800, Superlum, Cork, Ireland) for epi-illumination. The latter source is what has been used in the AOSLO designs described by Dubra et al. [9,10]. For epi-illumination, the scanning mirrors are imaged approximately onto the objective pupil plane (Olympus Plan Achromat 10× 0.25 NA) with 2× magnification by a scan lens (75 mm, AC254-075-B) and tube lens (150 mm, AC254-150-B). The illumination beam (less than 5 mm at the galvos) is limited by the objective pupil ensuring that the microscope is operating at the 0.25 NA providing a theoretical lateral resolution of 1.95 µm and axial depth of field of 12.8 µm. Back-scattered light from the sample is de-scanned and the confocal illumination spot is imaged on the camera with an imaging lens (75 mm, AC254-075-B) for a combined 7.5× magnification from the sample to the camera plane. Camera coordinates (x c ,y c ) represent position relative to the confocal illumination spot. The image is formed by scanning the illumination spot and is represented by galvo coordinates (x g ,y g ). A camera area of interest (AOI c ) is recorded for each galvo coordinate resulting in 4-dimensional data that includes the 2D scattered light distribution for each illumination point in the sample (Fig. 2). In this configuration, individual camera pixels correspond to approximately 1 µm at the sample, and one Airy disk diameter (ADD) is approximately 3.9 camera pixels. These data can then be processed using digital masks that represent a virtual detector geometry to produce sample images with different contrast mechanisms. Alternatively, the data can be used to examine the scattered light distribution as an average or as a function of scan position.

Sample preparation
Retina from 13-lined ground squirrel (13-LGS) was used for this study for practical and experimental reasons: The 13-LGS is a diurnal, cone-rich (>85% of photoreceptors are cones) rodent that is increasingly important for vision research, due to the need for developing therapies for cone photoreceptor diseases [11,12]. In addition, this animal model is highly amenable to in vivo retinal imaging for comparison to previous AOSLO imaging results [10,13]. Retinal  Single-frame excerpt from a video simulation of scattered light collection (Visualization 1). De-scanned camera frames are collected as the illumination point is scanned across the sample resulting in 4D data: 2D scattered light distribution for each pixel in the 2D scanned tissue image. This visualization shows an example of the scattered light profile from the corresponding sample region (green circle) as the data is collected across the sample (right panel). A wide-field image of a blood vessel is shown in this example where a largely fixed pattern of scattered light translates across the detector plane as the illumination point is scanned. This provides interesting insights into detection modalities that were not previously possible. This intact retina sample was imaged fresh in PBS without chemical fixation. tissue was collected from 13-LGS that had been taken out of hibernation and arose from torpor for at least 6 hours for a separate experiment. Recent data suggests photoreceptor function returns to 13-LGS out of hibernation by this time [14]. Whole-globes were either fixed in 4% paraformaldehyde overnight or imaged fresh (within 2 hours of euthanasia) in PBS. Retina samples were dissected out of the whole-globe from the superior region of the 13-LGS retina (an area with larger cones and thinner retina) with choroid and sclera attached and will be referred to as "intact" retina hereafter. Choroid and sclera were removed from one retina sample to allow light to pass through photoreceptors from both directions (Figs. 1(B) and 1(D)). Samples were then mounted in a PBS-filled 0.5 mm deep well between two 0.17 mm thick cover-slips for imaging. The 0.5 mm PBS well was critical to avoid any tissue alterations that might otherwise arise from compressing the approximately 0.3 mm thick 13-LGS retina samples with more traditional mounting techniques.

Data collection and image analysis
A custom MATLAB (MathWorks Inc, Natick, MA, USA) program was developed to collect the scattered light from the sample with control of galvo range and step size as well as CMOS camera parameters including pixel clock, exposure, AOI c and location of the confocal illumination spot on the camera sensor. The retina structure is first brought into focus with a fast non-scanning mode using 780 nm LED trans-illumination ( Fig. 3(A)). Once the desired sample depth is brought into focus, the scanning mode is used to acquire scattered light (x c , y c ) at each galvo position (x g , y g ). As depicted in Fig. 2, camera frames are collected as the galvo mirrors scan and light is collected for each pixel across the AOI c on the sample (256 2 or 512 2 locations are sampled for the images shown in this study). The scan speed is limited by the framerate of the CMOS camera. Using a smaller AOI c will decrease the readout time, which depends on the number of detector rows in the AOI c . For example, a 32×32 AOI c allows a camera framerate just over 1 kHz, thus allowing a 256×256 sample image to be collected in 65 seconds. This pixel rate is about 5000× slower than PMT based systems, and therefore is not viable for in vivo imaging without multiplexing. However, for ex vivo imaging, this trade-off is acceptable as the goal is to provide the detailed map of the distribution of scattered light.
Once the custom software collects the 4D data, digital detection masks can be applied to the scattered light distribution of each galvo pixel (Fig. 4). This detection plane mask determines the spatial distribution of scattered light used to form the image. While these detection masks can take any shape within the CMOS camera pixel array, we simulated a few common detection geometries similar to those used in AOSLO imaging. In this study, all data was collected using 32×32 pixels (approximately 8.2 ADD, Fig. 3(B)) to maintain reasonable image acquisition time and memory usage, but could theoretically examine a range of 16×16 pixels to 1024×1024 pixels (approximately 4.1 to 262.5 ADD) for the CMOS camera used here. As shown in Figs. 3(C)-3(E) and Fig. 4, we examined digital masks similar to traditional AOSLO detection methods, including confocal [15], offset-pinhole [16], split detection [5], as well as a modified version of dark field imaging using an annulus-shaped aperture instead of blocking out the central-most light [17]. The digital apertures are defined as manually input pixel radius values from the center of the AOI c . Similar to using two detectors as in Scoles et al. 2014 [5], the digital split detection images are generated when one resulting half aperture image's pixel intensities (S1) are combined with the other half aperture image's pixel intensities (S2) by the following (S 1 − S 2 )/(S 1 + S 2 ), which can be done using horizontally or vertically opposed half widths (Fig. 3(D) and Fig. 3(E), respectively).
Contrast modulation ((L max − L min )/(L max + L min ), where L is gray-scale luminance) of the vessel wall relative to the adjacent retina 5 µm away, was measured using the image's gray values collected in FIJI [18]. Images were aligned in FIJI prior to analysis to ensure the same regions were being measured. Whole image contrast was measured from images of photoreceptors using Digital mask simulation and image processing. First, the structure of interest is brought into focus with trans-illumination, a retinal blood vessel for example (A). Then scattered light is collected in epi-illumination from each scanned pixel. The scattered light profiles (insets) are the average of camera frames collected over the entire scan. Customized digital masks shown in the insets are applied to each camera frame or scan pixel to produce the subsequent image simulation (B). For example only the central-most 4 pixels were used to simulate confocal imaging (C), and the normalized difference of half masks was used to simulate split detection (D and E). Structures vary in appearance depending on the pixels used or the orientation in which they are combined. The vessel wall, for example, is more visible in this data set with vertical orientation apertures for split detection (red arrow, compare D and E). This intact retina sample was imaged fresh in PBS without chemical fixation. Scale bar = 50 µm. Fig. 4. Single-frame excerpt from a video demonstrating sixty-five unique simulations of light detection (Visualization 2). We examined digital masks similar to traditional AOSLO detection methods, including confocal, split detection, and dark field. The left panel shows the resulting image that corresponds to the digital mask applied that includes the outlined pixels shown in the right panel. This highlights the effect of detector plane geometry since the entire scattered light distribution is acquired at once for each scan pixel and there are no registration artifacts. All apparent lateral shifts, apparent changes in focus, and changes in contrast are due solely to the detection geometry within this 32×32 pixel (8.2×8.2 ADD) region. This intact retina sample was imaged fresh in PBS without chemical fixation. the Brenner's focus measure [19,20], which detects contrast edges by measuring the difference between a pixel and its neighboring pixel. This was applied to images using the "fmeasure" MATLAB function suite of contrast metrics [21].

Scattered light measurements
The DSLD system collects the dynamic off-axis light scattering information as it scans the retina within the area of interest. Figure 2 shows an example of the scattered light profile from the corresponding galvo coordinate location (green circle) as the data is collected across the sample. The scattering properties can change depending how the tissue is prepared prior to imaging and what retinal structures are being analyzed. Figure 5 shows examples of the scattered light measurements from different retinal structures of interest under different tissue preparation conditions. We were able to more consistently image identifiable retinal structures using fresh tissue compared to fixed tissue. Photoreceptors became more challenging to bring into focus after retina opacification due to fixation or fresh tissue denaturation, the latter causing a progressive decline in image quality which inhibited imaging after approximately two hours (in PBS) from the time of tissue collection. As shown in Fig. 5, mean scattered light from inner retina structures appeared as a central peak whereas imaging the photoreceptors through the entire retinal tissue produces less uniform scattering patterns. The less uniform light scattering patterns were similar when imaging fresh or fixed tissue, but could be eliminated with the removal of the sclera and choroid (Fig. 5).

Digital mask outputs
While many aperture patterns and subsequent arithmetic combinations can be examined with this system, we demonstrate a few multiply scattered light collection modalities to highlight the capability and flexibility of the DSLD system. Differently sized halved aperture combinations were examined after imaging 13-LGS inner retina blood vessels and photoreceptors to simulate different split detection aperture masks (Fig. 6). By changing the orientation of the halved aperture masks, structures are prone to directionality artefacts. As shown in Figs. 3(D) and 3(E), the vessel wall has greater contrast with the horizontal orientation (25.0% contrast) compared to the vertical orientation (8.7% contrast) of halved apertures used for the split detection simulation of these data. Further, as shown in the top row of Fig. 6, blood vessels and cells were easier to distinguish from the surrounding retina with increasing digital aperture sizes. Also, vessel walls have higher contrast compared to the adjacent retina with increasing digital aperture sizes (1.9% to 38.3%).
However, an inverse contrast relationship was seen while resolving photoreceptors and wholeimage contrast with split detection simulations. As shown in the bottom row of Fig. 6, individual photoreceptors were easier to distinguish with smaller digital masks and whole-image Brenner's focus measure decreased with increasing pixel aperture sizes (399.5 to 41.6).
Digital apertures that simulate annular dark field microscopy were also examined. Figure 7 shows examples of the striking variability of contrast when data is collected focused on the inner retina (nerve fiber layer) and outer retina (photoreceptors) with different diameters of digital annular apertures. In the top row of Fig. 7, nerve fiber bundles have similar punctate reflectivity with a small (0.8 ADD) annular aperture compared to confocal AOSLO images of 13-LGS nerve fiber [22]. The contrast these hyper-reflective features of nerve fiber varies with increasing annular aperture sizes. The specific cellular or extracellular origin of these puncta in the nerve fiber are not known. In the bottom row of Fig. 7 , structures that resembled retinal pigment epithelial (RPE) cells were visible with a 3.6 ADD annular aperture. The visible and presumed  Figs. 6, 7 and 8). However photoreceptors could be consistently visualized with the sclera and choroid removed (bottom row, also see Fig. 1), which may be useful for investigating wave-guiding properties of photoreceptors. #Fresh tissue imaged in PBS within 2 hours after euthanasia. * Fixed in 4% paraformaldehyde for 24 hours, then imaged in PBS. ∧Sclera and choroid removed prior to imaging.  We also examined the effect of off-set apertures on intrinsic retinal structure contrast. The top row of Fig. 8 shows a few examples of the orientation dependent nature of detecting the variable hyper-reflective contrast of wave-guiding photoreceptors using a sub-Airy aperture (0.5 ADD) at increasing distances away from the center. The bottom row of Fig. 8 shows off-set aperture contrast of a blood vessel and surrounding nerve fiber layer, revealing variable hyper-reflective contrast of blood cells, as well as the elongated hypo-reflective structures in the nerve fiber layer, thought to be Müller cell septa [24]. The specular reflective nature of these images is likely due to the relatively small ADD offset (2.5 Airy disc offset) of these apertures, as seen in previous work using similar aperture offsets [16]. Fig. 8. Simulating off-set aperture microscopy. The top rows shows images of photoreceptors with variable contrast depending on the position of the pinhole, with a pair of hyper-reflective photoreceptors appearing with a −10x c ,−10y c camera pixel ( 2.5 ADD) offset, 0.5 ADD aperture, compared to other 0.5 ADD aperture arrangements at different offsets. The intact retina in the top row was imaged in PBS after 24 hour fixation in 4% paraformaldehyde. The bottom row shows images of a blood vessel with 2.5 ADD central and clockwise apertures, revealing variable contrast for blood cells (orange circles) and the hypo-reflective features of the neighboring nerve fiber layer (red arrows); thought to be Müller cell septa [24]. The intact retina sample in the bottom row was imaged fresh in PBS without chemical fixation. Scale bars = 50 µm.

Discussion
The growing number of detection strategies for imaging retinal structures reported in the literature makes it clear that contrast optimization needs to be tailored towards the cell type or structure of interest. Yet this optimization is challenging due to an incomplete understanding of light scattering in the retina. As a small step toward improving that understanding, we have demonstrated a proof-of-principle DSLD microscope as a platform for examining contrast detection strategies when imaging retinal structure. This is a novel approach of collecting multiply scattered light for subsequent simulation of contrast mechanisms and has potential for improving known or uncovering novel methods of imaging contrast. We have shown examples from retinal photoreceptors, RPE, blood vessels and nerve fiber layer, and demonstrated that this platform can be used to explore contrast from these structures that compliments other methods to provide a more complete understanding of retinal cell contrast.

Applications
There is a need to find cellular biomarkers that indicate the early onset of retinal disease before severe visual consequences manifest. Non-invasive imaging of unlabeled cell types with AO ophthalmoscopy is revolutionizing the way we interpret disease states and track treatment strategies [1,2,25], but cellular biomarkers that warn of disease onset remain elusive. The DSLD platform presented here allows for rapid testing of detection modalities using multiply scattered light from microscopic retinal structures and/or subsequent mathematical combinations of different image results in order to improve cellular contrast (as in split detection and multi-offset AOSLO [5,6]). Identifying optimal geometry for imaging distinct retinal layers and cell types by manually adjusting the mask for each image limits the number of detection schemes that can be evaluated at a single time-point [6,26]. Using digital mask results ex vivo from the DSLD system to design and evaluate confocal-plane apertures and detection schemes in vivo using AOSLO represents the next logical step for this work. Figure 2 shows that a largely fixed pattern of scattered light translates across the detector plane as the illumination point is scanned. This provides interesting insight into detection modalities that was not previously possible. For example, as a bright spot in this pattern shifts from one side of the confocal spot to the other, split detection signal would change sharply. Further investigation of this may lead to a better understanding of the contrast mechanism. The opposing trends in contrast for inner and outer retina structures (Fig. 5) supports the idea that index gradient at the focal plane deflects forward scattered light that is then backscattered, and that offset detection increases contrast of structures with less specular reflection [16]. Further, Guevara-Torres et al. revealed that axial detector displacement is also a critical consideration when creating cell-specific detection masks [27]. Future investigation of the translating pattern we observed as a function of relative focus between illumination and detection could complement this interesting study.
Due to the large number of potential detector geometries, especially when arithmetic combinations are considered, machine learning approaches may be useful to identify which aspects of the scattered light distribution contain the information to yield improved contrast of distinct retinal cells and structures. Machine learning may also eliminate subjective selection of digital mask arrangements to test effects on contrast. We used Brenner's focus measure [19] for quantifying image contrast in this work, due to this metric effectively distinguishing between depths of focus in AO-flood images of photoreceptors [28]. However, further validation work should be performed with regard to image quality and contrast metrics with retina images collected by the DSLD microscope. Such metrics can be used to drive the aforementioned machine learning approaches for investigation of contrast origins and mechanisms.

Limitations
In its current design, the DSLD microscope is limited to data collection from ex vivo samples. Image distortions of retinal imaging due to eye motion of living subjects would be exacerbated by the low scan rate used here, though faster cameras exist. Future work will prioritize re-designing the DSLD system for in vivo application using a faster camera, increasing ADD range of the AOI c , increasing scan speed with multiplexed spots and evaluating line illumination geometry.
While there are many benefits afforded with the experimental control of using ex vivo tissue preparation and microscopy, the current work is limited in its ability to confidently simulate applications that can be directly applied to in vivo retinal imaging. Tissue degradation is expected when using ex vivo retina, even as quickly as the 2 hour time-frame post-euthanasia used in this study. Also, while the DSLD system here has a purposely similar numerical aperture (0.25) to a dilated human eye (0.23), the optics of this system do not accurately replicate native eyes, nor the aberration correction of AO. While direct cellular activity can be measured for several hours after dissection if maintained in the proper conditions [29], the native translucent properties fade and scattering properties almost certainly change when using excised or fixed retina. It is unclear to what extent these changes will alter our ability to translate simulated contrast to applications of in vivo retinal imaging. Data collection from the innermost surface retinal layers was consistent and resulted in the highest quality images. When we imaged through the entire tissue to measure scattering properties from photoreceptors, image quality was negatively affected, likely due in part to rapid tissue opacification and denaturation of the naturally transparent retina. While we were not able to view 13-LGS photoreceptors at the reliable consistency and high resolution of imaging them in vivo with AOSLO [10], we were able to image through inner layers to visualize some photoreceptors in the superior retina of fresh (Fig. 6) and fixed ( Fig. 7 and Fig. 8) intact samples. This peripheral superior region of retina (above the optic nerve) is thinner and has larger photoreceptors compared to inferior retina (below the optic nerve) [10,30,31], such that visualizing cells in this region may be less susceptible to aforementioned consequences of ex vivo tissue denaturation. However, the increased non-uniformity of scattered light detected when imaging photoreceptors with this system (Fig. 5) suggests that the reduced image quality is related to an increase in light scatter caused by the denatured state of the ex vivo retinal samples being used. Image quality improved when sclera and choroid were removed from the sample (Figs. 1(B) and 1(D)), likely due to decreased variability in light scatter (Fig. 5). The ability to image using both trans-and epi-illumination with this system may be a useful preparation for learning more about photoreceptor wave-guiding principles. However, such unnatural modifications to the retina only further distances translatability of any contrast observations for the use with in vivo imaging.

Conclusion
We have presented a digital microscope and detection simulation platform that can be used for exploring methods of cellular contrast, and has the potential to uncover novel methods of cellular contrast; either to view structures in new ways or view previously invisible structures without exogenous contrast agents. This multi-detection simulation approach increases the number of off-axis points compared to current in vivo methods to "image" the distribution of scattered light, and can provide useful insights for improving intrinsic contrast detection of non-invasive retinal imaging approaches such as AOSLO. Most notably in this work, our split detection simulations suggest that the use of smaller apertures may increase the resolution for imaging the putative photoreceptor inner segments when using this type of contrast detection. The scattered light measurements highlight the variability of light capture when imaging a structure as optically complex as the retina.