Comparison of confocal and non-confocal split-detection cone photoreceptor imaging.

Quadrant reflectance confocal and non-confocal scanning light ophthalmoscope images of the photoreceptor mosaic were recorded in a subject with congenital achromatopsia (ACHM) and a normal control. These images, captured with various circular and annular apertures, were used to calculate split-detection images, revealing two cone photoreceptor contrast mechanisms. The first contrast mechanism, maximal in the non-confocal 5.5-10 Airy disk diameter annular region, is unrelated to the cone reflectivity in confocal or flood illumination imaging. The second mechanism, maximal for confocal split-detection, is related to the cone reflectivity in confocal or flood illumination imaging that originates from the ellipsoid zone and/or inner-outer segment junction. Seeking to maximize image contrast, split-detection images were generated using various quadrant detector combinations, with opposite (diagonal) quadrant detectors producing the highest contrast. Split-detection generated with the addition of adjacent quadrant detector pairs, shows lower contrast, while azimuthal split-detection images, calculated from adjacent quadrant detectors, showed the lowest contrast. Finally, the integration of image pairs with orthogonal split directions was used to produce images in which the photoreceptor contrast does not change with direction.

Visualization of the anterior end of cone inner segments, however, is possible even in the presence of pathology, using non-confocal split-detection, a technique we first demonstrated in normal controls and congenital achromatopsia (ACHM) [10,19]. Since then, this imaging modality has been applied to the study of photoreceptors in retinitis pigmentosa [20,21], Usher syndrome [21], choroideremia [22], RPGR-associated retinopathy [23], Best vitelliform macular dystrophy [12], blue cone monochromacy [13], cone-rod dystrophy [13], Stargardt disease [24], macular telangiectasia [13,25], small hard retinal drusen [26], central serous retinopathy [27], closed-globe blunt ocular trauma [13], traumatic brain injury [28], and macular holes [29]. In point-scanning ophthalmoscopes, backscattered light can be thought of as comprised of a confocal component and a non-confocal component. The confocal component, loosely defined as light within a circle 1-3 Airy disks in diameter (ADD) centered on the geometrical image of the point illumination, is thought to be dominated by single scattering. Light outside this confocal area, is comprised by a combination of single-and multiple-scattering from various retinal layers, and is referred to here as non-confocal.
Split-detection images are generated from pairs of images captured using confocal or nonconfocal detectors, by dividing the difference of the corresponding pixel values by their sum. Because the resulting images have negative values, a positive offset is then added after multiplication by a (contrast stretching) scalar for visualization in computer displays that only accept positive intensity values. Following the demonstration of non-confocal split-detection photoreceptor imaging, confocal split-detection (differential phase contrast, [30,31]), and confocal plus non-confocal split-detection (knife-edge, [32]) retinal imaging were reported. In all three modalities individual photoreceptors appear as bright-dark opposed crescents that delineate a circle on a grey background, resembling a side-illuminated spherical dome. Confocal splitdetection, however, provides photoreceptor images with the highest transverse resolution, even allowing the occasional visualization of rod photoreceptors in normal retinas. This observation led us to formulate the questions that this work aims to answer. First, does confocal splitdetection show the same portion of the photoreceptor as non-confocal split-detection? And second, does confocal split-detection reveal cone photoreceptors in the absence of confocal and flood-illumination reflectance signal from the (EZ or IS-OS junction)? These questions are important for the interpretation of split-detection images in clinical applications. In addition, we seek to optimize the light detection geometry to maximize cone photoreceptor contrast in split-detection images.
Split-detection images of the photoreceptor mosaic resemble Nomarski differential interference contrast [33], which in transparent (phase) objects are proportional to the directional derivatives of the 2-dimensional phase profile. The automated identification of cones in these images can be computationally complex, because it requires the matching of bright and dark image region pairs within a size range and distance [34][35][36][37]. To facilitate this task, we explore the use of the mathematical integration of split-detection image pairs in which the splitting of the detection area occurs along orthogonal directions. In the resulting images, to which we refer hereon as integral images, cones appear as bright regions that could be automatically identified as local intensity maxima [38][39][40].

Quadrant detection adaptive optics scanning light ophthalmoscope
A custom AOSLO [41,42] with an internal fixation target [43] was modified through the addition of photo-multiplier tubes (PMTs) as depicted in Fig. 1. In this instrument, the output of a single-mode polarization-maintaining optical fiber coupled to a 790 nm super-luminescent diode (S-790-G-I-15-PM; Superlum, Carrigtwohill, Ireland) was used as a point source for the imaging channel. Similarly, the output of a single-mode polarization-maintaining fiber coupled to an 850 nm super-luminescent diode (SLD-MS-381-MP-PM, Superlum), and optically conjugate to the retina was used as a wavefront sensing beacon. Both light sources were modulated to be on only during the imaging portion of a 17 Hz raster scan (1.5°field-of-view), delivering an average power of 60 and 24µW at the cornea, respectively over a 7.75 mm diameter pupil. The 790 nm light scattered by the retina was captured using either four (see confocal quadrant detection in Fig. 1) or five (non-confocal quadrant detection) H7422-50 PMTs by Hamamatsu Photonics (Hamamatsu, Japan). The output of each PMT was fed to a transimpedance amplifier (HCA-10M-100K; Femto Messtechnik GmbH, Berlin, Germany) before 14-bit digitization using a 16-channel ATS9416 card by Alazartech (Pointe-Claire, Qc, Canada). The output of the transimpedance amplifiers were digitized in synchrony, resulting in four/five spatio-temporally registered images. The fifth PMT, when used, captured back-scattered light through a pinhole optically conjugated to the retina illumination. This confocal pinhole was created by a tilted custom reflective binary mask that transmits light over a circular area 1.0 Airy disk in diameter (ADD). The non-confocal light reflected by this mask was relayed using an achromatic doublet pair onto a selectable circular or annular aperture (see Table 1). This aperture was relayed onto knife-edge mirrors using achromatic doublet pairs, in order to split the light into four equal quadrants, one per PMT. When the circular aperture was used for quadrant light collection, the first pinhole mask was adjusted so that all the light returning from the retina was reflected towards the circular aperture.

Human subjects
One subject with genetically confirmed CNGA3-associated achromatopsia (ACHM) and one with no known ocular pathology were recruited for this study. Research procedures followed the tenets of the Declaration of Helsinki and were approved by the institutional review board of Stanford University. Informed written consent was obtained after explanation of the possible consequences of the study. One drop of 2.5% phenylephrine and one drop of 1% tropicamide were used to dilate the pupil and induce cycloplegia before AOSLO imaging. Artificial tear drops were administered to prevent corneal dehydration, as necessary. A bite-bar attached to a manually-operated three-axis translation stage was used to align and stabilize the subject.

Image acquisition and scaling
The gain of each PMT was adjusted to achieve similar mean pixel intensities across all quadrant detectors while avoiding digitizer saturation. AOSLO focus was first determined by finding the photoreceptor layer with maximal reflectivity in the confocal channel. Optical coherence tomography using a similar wavelength indicates that this maximally reflective layer corresponds to the photoreceptor EZ/IS-OS junction [7,45]. Then, focus was shifted anteriorly by 0.025 diopters (∼7 µm in an average eye with 16.7 mm back focal length) to maximize contrast in a split-detection channel [19]. Sequences of 300 AOSLO frames were captured for each aperture in Table 1, in random order, to avoid potential bias due to tear film degradation during the imaging session. The ACHM subject was imaged approximately 6.5°superior to the fovea, which shows both hypo-reflective cones and bright rods [10], while the control subject was imaged at an 8°temporal from the fovea. These regions were chosen so that the comparison of contrast between ACHM and normal retina was made amongst similar-sized cones. AOSLO images were scaled using the axial length of the eye [46], measured using a biometer (IOLMaster; Zeiss Meditec, Dublin, CA). The axial lengths of the control and ACHM eyes were 24.33 and 23.78 mm, respectively, which yield retinal scaling factors of 295 and 288 µm/°.

Image processing
Image distortion due to non-constant angular speed of the resonant (horizontal) optical scanner and involuntary motion of the eye were compensated in post-processing [47]. After registration, the 100 distortion-corrected frames with highest normalized cross-correlation with respect to a manually selected template were averaged to improve signal-to-noise ratio [47]. Averaged images from individual quadrant detectors were normalized across detectors to compensate for variation in PMT sensitivity, gain and optical alignment. The registered and averaged images from all experimental conditions were manually aligned in Photoshop (Adobe Systems, San Jose, CA) first, and then automatically, with the Fiji plugin bUnwarp [48,49]. The bUnwarp parameters were: zero subsample factor, very coarse initial deformation, fine final deformation, zero divergence, curl and landmark weights, unit image weight, consistency weight set at 10 and 0.01 stop threshold.
Split-detection images were generated by combining registered and averaged quadrant images Q i (with pixel values in the 0-255 range) using the following pixel operations, Here, SD x and SD y correspond to the vertical and horizontal "splitting" of the detected light, while SD P45 and SD M45 correspond to diagonal splitting. These four modalities can be thought of as radial splitting, as they involve opposite halves or quadrants. The remaining four modalities can be thought of azimuthal differences, as they involve the subtraction of adjacent quadrants. The intensity of the split-detection images was offset by 128 gray levels and linearly stretched equally for display and metric calculations.
Additional split-detection images were generated by subtracting the confocal images captured with a single PMT from a copy of itself shifted by two pixels, and then dividing by the image itself. These "calculated" confocal split detection images were compared with "experimental" confocal split-detection images.
Irrespective of the retinal structure being imaged [19,42,50,51], split-detection images resemble those obtained in Nomarski differential interference contrast imaging [33], which provides a map of finite phase differences [52]. This led us to propose the numerical integration of the differential image pairs "split" along orthogonal directions, using a least-squares global integrator, originally developed for Shack-Hartmann wavefront sensors [53] and then generalized for lateral shearing interferogram pairs with orthogonal shearing directions [54]. In this method each input image is assumed to be the difference of two laterally shifted phase maps, and thus the calculation of the phase map itself becomes solving a very large set of linear equations. Due to the large number of equations (one per image pixel), traditional linear equation solving methods are not practical. Using discrete Fourier analysis, this problem can be posed as a convolution of the split-detection images with a filter completely determined by the amplitude and direction of the shifts in units of pixels. Here, we applied this method assuming that the shift is a single pixel along directions perpendicular to the split directions. The use of a convolution with two images with orthogonal split-directions eliminates the uncertainty that remains when integrating a two-dimensional function along a single axis (which is an additive function of a variable along the orthogonal direction).
Cones in both subjects were segmented in the split-detection image using an automated method [55], followed by manual correction. Michelson contrast [56] and the standard deviation of the pixels inside each encircled cone were used as cone contrast metrics. Rod photoreceptor contrast metric were not calculated.

Statistical methods used for contrast metric comparisons
The contrast metrics of the confocal images captured across all detection apertures were compared using the intra-class correlation coefficient (ICC) to assess imaging and AO correction reproducibility. ICC(3,k) accounted for two-way mixed effects for average measures [57,58] where k is equal to the number of annular detection apertures (i.e., 6). A pairwise Wilcoxon signed right-tailed rank test was used to compare the photoreceptor contrast metrics between images corresponding to any given pair of apertures listed in Table 1 [59].

Repeatability
Tear film degradation, imperfect pupil centering and other non-controlled factors could affect cone contrast metrics throughout an imaging session. In order to test this hypothesis, we compared the cone photoreceptor metrics in the average confocal images, captured with a single PMT, across all image sequences. The ICC of the Michelson contrast and standard deviation were 0.97 and 0.98, respectively for the control subject, and 0.95 and 0.96 for the ACHM subject. These high correlation values suggest that non-controlled factors did not substantially affect image contrast, and thus, we assume that variation between cone metrics in split-detection images are due to the different experimental conditions.

Quadrant images
Image contrast and resolution appears comparable across all four quadrants (see Fig. 2), as expected, due to the rotationally symmetry of photoreceptor inner and outer segments. The quadrant images captured through the 3 ADD circular aperture resemble images captured with a single confocal circular detector, both in terms of contrast and resolution. That is, rods appear as bright spots, while cones are bright in the control subject and dark in the ACHM subject. The normalized cross-correlation between quadrant images shows relative shifts between quadrants of approximately 2 pixels along the horizontal and vertical directions (0.5 ADD).
In non-confocal quadrant images, cones in the control eye resemble those of the confocal images but broader, with irregular shape, fuzzy boundaries and contrast that decreases with increasing annulus size. In non-confocal ACHM quadrant images, cones have circular smooth sharp boundaries, with highest contrast in the 5.5-7.5 annulus. In both retinas, cones appear as domes illuminated from the side (i.e., bright on one side and dark on the opposite side).
Rods were not resolvable in non-confocal images from either retina. Intensity variations of a much larger spatial scale than individual photoreceptors were seen with orientations not obviously related to the quadrant geometry, with small to no contrast variation with annulus diameter in quadrant images from both retinas. These intensity variations resemble those in non-confocal scanning laser ophthalmoscopes [60,61], AO scanning light ophthalmoscopes [50] and flood-illumination AO ophthalmoscopes [62], and are not evident in point-scanning reflectance ophthalmoscopes when using detection apertures equal or smaller than one Airy disk in diameter [63].

Confocal vs non-confocal split-detection
In confocal images of the normal photoreceptor mosaic (see Fig. 3(A)), rods are barely resolvable white dots, while cones at the same retinal locations are irregular in shape, larger and often brighter than rods [5,64]. At large eccentricities, cones [65,66] and/or with annular imaging or illumination pupils [67] can appear multimodal. In ACHM retinas, rod spacing is typically larger than in healthy controls. Cones in ACHM (yellow arrows in Fig. 3(E)), however, are substantially dimmer than rods, often appearing completely dark [17,19].  Table 1 (in order). All images were contrast stretched equally for display purposes.
In split-detection images, resolvable cones and rods appear as a bright-dark spot pairs on a grey background (see Fig. 3(B, D, F & H)), suggestive of side-illuminated spherical domes. These domes are delimited by circles [68] that when superimposed to the confocal images contain the smaller often irregular bright spots. Split-detection cone contrast is highest 7 ± 3.5 µm anteriorly to the best confocal image captured with a single PMT.
In confocal split-detection images of the normal retina (Fig. 3, panel B), cones have high contrast and sharp boundaries and only a small fraction of rods can be resolved. In confocal split-detection images of the ACHM retina (Fig. 3, panel (F)), however, it is rods that exhibit high contrast, while cones have low contrast and smooth boundaries. In non-confocal split-detection images from both subjects, only cones are resolvable (Fig. 3(D & H)), and in the control retina with lower contrast than in confocal split-detection (compare panels B and D in Fig. 3).
Cones appear dark in confocal images in the ACHM subject, have low to no contrast in the confocal split-detection images, and high contrast in non-confocal split-detection images. This indicates a contrast mechanism that it is independent of the presence of outer segment confocal intensity and that only reveals cone inner segments in non-confocal split-detection. The high contrast cones in the confocal split-detection images in normal but not the ACHM retina suggests a second contrast mechanism, linked to cone outer segment (confocal) intensity. Importantly, the difference in cone contrast between control and ACHM confocal split-detection is not due to a substantial difference in inner segment diameter, 7.2 ± 0.05 µm (n=726) and 7.8 ± 0.2 µm (n=54), respectively, both larger than the ∼2.1 µm transverse resolution. Also, it should be noted that despite contrast differences, the boundaries of cone inner segments coincide for both sources of contrast.
The fact that the confocal split-detection images show small features with high contrast that are visible (i.e., resolvable) in the confocal images but not in the non-confocal images might suggest that the confocal split-detection is just a derivative of the confocal image (divided by the local intensity). To test this hypothesis, we generated the calculated split-detection images shown in panels C and G of Fig. 3. The calculated and experimental images have an overall similar appearance with some differences, with the most notable being that each multimodal cone appears as multiple features with poor contrast in the calculated image, and as a single high contrast feature in the experimental image. Also, although not always resolvable in the calculated image of the control subject, they seem to cover all gaps between the larger cones, while in the corresponding experimental image most of the space between cones has an almost completely uniform gray level. Finally, the rods in the ACHM subject under the shadow of a blood vessel show higher contrast in the calculated image.

Split-detection contrast variation with annulus and modality
The quadrant images in Fig. 2 were combined using the formulae in Eq. (1). The resulting split-detection images and cone contrast metrics for the control subject are shown in Fig. 4 and Fig. 5, respectively. Similarly, Fig. 6 and Fig. 7 are the images and plots from the ACHM subject. In both Fig. 4 and Fig. 6, the line separating the quadrants used to create the split-detection images matches the orientation of cone contrast in all images.    Cones and the few resolvable rods in confocal split-detection images from the control subject (Fig. 4) appear as ovals or circles with one half uniformly bright and the other half uniformly black, with a sharp transition between them. Confocal and non-confocal split-detection images of the control subject are similar, in that the cones are brighter and darker in the corresponding halves, but stark differences can also be observed. In the latter the cone outer boundaries and the boundary between the bright and dark sides of each cone are smooth with contrast decreasing with annulus size. Rod photoreceptors in these images are not resolvable for any annulus, and the space between cones varies in grey level substantially across the image, thus reducing the contrast of cone boundaries.
ACHM confocal split-detection images (Fig. 6), on the other hand, reveal the rods in all detection geometries, with cones barely visible or not at all. As in the images from the control subject, the rods appear as ovals or circles with uniformly bright and dark halves with a sharp transition between them. The photoreceptor boundaries are sharp, with the space between photoreceptors appearing uniformly gray. The lower image contrast in the vertical splitting geometries is due to imperfect alignment of the splitting mirrors.
In agreement with the images shown in Fig. 3, the ACHM non-confocal split-detection images show cones while rods are not resolvable. Cones look substantially different from those in confocal split-detection imaging of the normal retina, with softer boundaries, and continuously smooth transition between the bright and dark sides, as opposed to the more binarized appearance seen in the left column of Fig. 4. Both visual inspection of Fig. 6 and the metrics in Fig. 7, show the diagonal split-detection configurations (M45 and P45) produce equal or higher contrast than those of other configurations (p < 0.05), and that images captured using 5.5-7.5 and 7.5-10.0 ADD annuli have equal or higher contrast that other annuli (p < 0.05) for all detection geometries. The contrast metrics for the confocal split-detection (Fig. 7) should be considered carefully, and these images do not show the cones as the characteristic domes, but rather as multi-lobe structures that resemble more the surrounding rods than cones.

Split-detection contrast of digital annuli combination
Quadrant images captured with different annuli were digitally added to simulate the use of thicker annuli. The 5.5 inner diameter ADD was chosen as the starting annulus due to it showing the highest cone contrast in ACHM. The resulting images are shown in Fig. 8 for the control and Fig. 10 for the ACHM subjects, with the corresponding cone contrast metrics shown in Fig. 9 and Fig. 11 , respectively. In both the control and ACHM retina, contrast decreases monotonically with annulus thickness, although markedly less in the ACHM retina. The cone contrast in the ACHM images captured with either 5.5-7.5 or 5.5-10.0 annuli are equal or higher than for thicker annuli (p < 0.05), with SplitP45 and SplitM45 showing equal or higher contrast than other configurations (p < 0.05).  Fig. 8

Integral images
Automated identification and segmentation of photoreceptors in split-detection images is not trivial. Algorithms must search for matched pairs of bright and dark regions of certain dimensions and separation, with challenges at the image boundaries [34,36,37,55,[68][69][70]. Therefore, and given their derivative-like appearance of the split-detection images, we propose the integration of   Fig. 10 pairs of orthogonal split-detection images, a process that yields the images in Fig. 12 and Fig. 13 below. These integral images resemble those of the confocal channel albeit with lower contrast and resolution, with cones having bright centers with smooth boundaries.

Summary and discussion
The comparison of split-detection images from ACHM and control subjects indicates that cones photoreceptors reveal themselves through two different contrast mechanisms. The transition between the confocal and non-confocal contrast in both quadrant and split-detection images of the ACHM retina can also be seen in Fig. 2 of Chui, VanNasdale, and Burns' seminal demonstration of non-confocal AOSLO imaging [66]. In that figure, the transition occurs at a larger distance from the optical axis (6 ADD), which is consistent with our finding when the size of their pinhole (10 ADD) is considered.
The visualization of cones in non-confocal (annular aperture) split-detection, despite the absence of confocal and confocal split-detection signals in the ACHM retina, points to a first contrast mechanism that is unrelated to the strength of the EZ/IS-OS reflectivity. This contrast is clinically important because numerous conditions result in the attenuation or even absence of the cone confocal signal, including those that result in the photoreceptor inner and outer segments pointing away from the pupil (e.g., drusen, sub-retina fluid) [10,26,[71][72][73]. For this mechanism, we found that the annulus with inner radius ∼5.5 ADD and outer diameter equal or smaller than 10 ADD maximize split-detection cone contrast.
In the control retina, cones can be visualized both in confocal and non-confocal split-detection, indicating that that in addition to the contrast mechanism present in the ACHM retina, there is a second mechanism that delivers maximal contrast in the confocal region. This second mechanism, which seems to originate from the confocal reflectivity signal, produces crisper images, with higher transverse resolution that can reveal rod photoreceptors. Because this second contrast mechanism does not reveal hypo-reflective cones it will likely be less clinically useful than non-confocal split-detection, as it does not appear to add to the information revealed by the confocal images.
Intensity variations of a much larger spatial scale than individual photoreceptors were seen in quadrant and split-detection images that appear unrelated to the detection geometry, with small to no contrast variation with annulus diameter. It is possible that these features correspond to variations of choroidal and retinal pigment epithelium scattering and absorption properties.
Finally, the integration of split-detection image pairs with orthogonal split directions was demonstrated, generating images that do not have the directional dependence seen in splitdetection. This integration was motivated by the resemblance between split-detection images and those from Nomarski differential interference contrast [33], which also leads to posit that these integral images can be thought of as phase contrast images. Even though cone contrast in the integral images appears low, the intensity profiles of the cones should make their automated counting and segmentation less computationally complex [35][36][37].