Derivation of human retinal cell densities using high‐density, spatially localized optical coherence tomography data from the human retina

Abstract This study sought to identify demographic variations in retinal thickness measurements from optical coherence tomography (OCT), to enable the calculation of cell density parameters across the neural layers of the healthy human macula. From macular OCTs (n = 247), ganglion cell (GCL), inner nuclear (INL), and inner segment–outer segment (ISOS) layer measurements were extracted using a customized high‐density grid. Variations with age, sex, ethnicity, and refractive error were assessed with multiple linear regression analyses, with age‐related distributions further assessed using hierarchical cluster analysis and regression models. Models were tested on a naïve healthy cohort (n = 40) with Mann–Whitney tests to determine generalizability. Quantitative cell density data were calculated from histological data from previous human studies. Eccentricity‐dependent variations in OCT retinal thickness closely resemble topographic cell density maps from human histological studies. Age was consistently identified as significantly impacting retinal thickness (p = .0006, .0007, and .003 for GCL, INL and ISOS), with gender affecting ISOS only (p < .0001). Regression models demonstrated that age‐related changes in the GCL and INL begin in the 30th decade and were linear for the ISOS. Model testing revealed significant differences in INL and ISOS thickness (p = .0008 and .0001; however, differences fell within the OCT's axial resolution. Qualitative comparisons show close alignment between OCT and histological cell densities when using unique, high‐resolution OCT data, and correction for demographics‐related variability. Overall, this study describes a process to calculate in vivo cell density from OCT for all neural layers of the human retina, providing a framework for basic science and clinical investigations.

example, reductions in photoreceptor cell numbers amidst numerous other changes in the outer retinal layers have been observed in histological samples from human eyes with age-related macular degeneration (Curcio et al., 1996), and in mouse models of retinitis pigmentosa apoptosis of photoreceptors and bipolar cells has been reported (Portera-Cailliau et al., 1994;Zhang et al., 2010). Similarly, in human and primate histological studies of glaucoma, reduced retinal ganglion cell (GC) densities and optic nerve axon numbers have been correlated with reductions in visual field sensitivity (Antwi-Boasiako et al., 2021;Garway-Heath et al., 2000;Kerrigan-Baumrind et al., 2000). However, these cellular components cannot be readily visualized in vivo with widely available imaging technologies, such that this knowledge of structural changes at the cellular level and quantitative relationships with outputs of vision cannot be directly translated to applications where cell parameters are not directly measured.
Retinal thickness measurements acquired in vivo using optical coherence tomography (OCT) are commonly applied in lieu of cellular parameters in both research investigations comparing structural and functional outputs and ophthalmic clinical care (Hood et al., 2019;Leite et al., 2012;Ly et al., 2018). In these applications, there is an underlying assumption that thinner retinal layers are analogous to a loss of corresponding retinal cells. However, quantitative comparisons between retinal layer measurements derived from OCT and histological cell parameters have been limited by available OCT measurements from commercial software, which are typically averaged over concentrically arranged subfields, for example, in the Early Treatment for Diabetic Retinopathy Study grid (Brandl et al., 2019;Cvenkel & Kontestabile, 2011). First, the sizes of individual subfields are relatively large, typically leading to few averaged retinal thickness measurements being calculated, and even fewer measurements along each meridian, limiting the number of measurements comparable to each histological slice. Moreover, such grids are based on the assumption that the distributions of retinal thickness measurements in healthy and pathological retinas occur in a symmetrical and concentric manner.
However, studies using spatial cluster analysis of OCT measurements from healthy retinas have identified variably symmetric patterns of age-related change in different retinal layers (Tong et al., 2019;Trinh et al., 2021;Yoshioka et al., 2017). The relative sparseness of obtained retinal thickness measurements and variable distributions of measurements for different retinal layers suggest that commercially available OCT measurement grids may be inadequate to enable the visualization of patterns of OCT-derived retinal thickness measurements to compare to histological distributions. Furthermore, commercial OCT software calculate retinal thickness values axially (Alonso-Caneiro et al., 2016), which do not consider alterations in retinal tilt with increasing eccentricity or tilted scan acquisition.
In the present study, a customized high-density grid was applied to develop an OCT-based model representing distributions of the layers of the macula corresponding to cellular structures, namely, the ganglion cell layer (GCL), inner nuclear layer (INL), and inner segmentouter segment layer (ISOS), in healthy human retinas. These models subsequently enabled comparisons between OCT data and density data for the corresponding cell bodies, as reported by previous studies using histological samples of the human retina (Curcio & Allen, 1990;Masri et al., 2021), with the ISOS chosen to match the photoreceptor outer segment densities described in . Verification of whether OCT-derived retinal thickness measurements can adequately describe corresponding retinal cellular changes is imperative to confirm the role of retinal thicknesses as a surrogate structural measure. Should OCT reflect histological cell density data, the subsequent ability to calculate in vivo cell densities would be a powerful tool for the detection of a variety of retinal pathologies and enabling in-depth studies of the human retina.

Participant recruitment and data collection
This study adhered to the tenets of the Declaration of Helsinki  (Curcio & Allen, 1990;Masri et al., 2021).
OCT posterior pole scans acquired with the Spectralis SD-OCT (Hei- Note: Within the modeling cohort, all eyes were used to develop normative models, while the age-similar cohorts were chosen as similar to historical histological cohorts (Curcio & Allen, 1990;Masri et al., 2021), for use as comparisons to corrected data from the test cohort and derivation of cell densities from histological data. See Table S1 Figure 1), with agreement between them constituting the "ground-truth" segmentation (Tian et al., 2016;Trinh et al., 2021).

Experimental design and statistical analysis
For extraction of OCT-derived thickness measurements from the GCL, INL, and ISOS region of photoreceptors at a higher spatial resolution than available in commercial OCT review software with adjustment for OCT B-scan tilt, a custom algorithm was written using MATLAB Version 9.6 (MathWorks, Natick, MA, USA). The fundamental components of the algorithm have been described previously .
In short, after data were extracted in a RAW format,  Given previous reports of nonlinear patterns of age-related decline in various retinal layers, with different rates of change dependent on macular location (Mwanza et al., 2011;Tong et al., 2019;Trinh et al., 2021;Yoshioka et al., 2017), hierarchical cluster analysis was applied to identify retinal locations demonstrating statistically similar aging characteristics irrespective of outcomes of multiple linear regression analysis. Cluster analysis has been previously applied to characterize distributions of amino acid signatures in retinal neurons in the normal and degenerating vertebrate retina (Jones et al., 2003;Kalloniatis et al., 1996;Marc & Jones, 2002;Marc et al., 1995); in this study, a similar approach was applied to enable the development of regression models describing change in retinal thickness measurements as a function of age (Tong et al., 2019;Trinh et al., 2021;Yoshioka et al., 2017). Participants in the modeling cohort were grouped by decade brackets (20 to <30 years, 30 to <40 years, etc.), with participants 70 years of age and more grouped together due to relatively small sample size in this subcohort. For the relevant retinal layers, for each grid square measurements were averaged across each age bracket prior to hierarchical cluster analysis using within-groups linkage and squared Euclidean distance. Cluster separability was determined by calculating d′ for each cluster pair, based on the mean (x) and standard deviation (σ) retinal thickness measurement for each cluster: Tong et al., 2019) The d' threshold criterion was set at 1, equivalent to 1 SD separation between clusters. Given the axial resolution of the Spectralis OCT is 3.87 μm, this was applied as an additional criterion to ensure at least 1 pixel's difference between cluster means. Should d′ < 1 or the difference in means be less than 3.87 μm for a particular cluster pair, these clusters were deemed not statistically separable and pooled together. Subsequently, retinal thickness measurements were pooled by cluster and age bracket, and quadratic and linear regression models were applied, with extra sums-of-squares F-test comparisons performed to identify the most appropriate regression model describing aging change for that cluster and retinal layer. Where a quadratic linear regression model was preferred, the point of inflection (x), or the point at which retinal thickness begins to decline, was calculated from TA B L E 2 Demographic characteristics of histological cohorts from previously reported data (Curcio & Allen, 1990;Masri et al., 2021). To enable comparisons of rate of change described by these models between clusters and between retinal layers, the average annual percentage change was calculated from peak retinal thicknesses predicted by the quadratic and linear regression models.

GC density (Curcio & Allen
Testing the derived age-correction factors on a naïve cohort would verify the generalizability of the regression models and in turn of parameters derived from the histological data sets. From the modeling cohort, participants falling within the age range of each histological cohort were selected as age-similar subcohorts (Tables 1 and 2); that is, all participants aged between 27 and 40 years were chosen as GCL and ISOS age-similar subcohorts, similar to the age ranges of the Curcio and Allen (1990) for GC data and

Linking OCT data to historical histological cohorts
Direct comparisons of OCT and histological data would enable the estimation of retinal cell density, and subsequently retinal cell counts, from corresponding retinal layer thickness measurements derived from OCT. Histological GC data were derived from Curcio and Allen (1990), who reported GC density in cells per en face square millimeter (cells/mm 2 ) as a function of eccentricity along each principal meridian.
For the derivation of GC density per the 60 × 60 measurement grid, similar to previous studies (Garway-Heath et al., 2000;Raza & Hood, 2015;Yoshioka et al., 2018), GC densities between the principal meridians were linearly interpolated at 5 μm intervals. GC densities were subsequently averaged across each grid square area. For the subsequent application of OCT data, GC density in cells per cubed millimeter (cells/mm 3 ) was calculated by dividing the GC density in cells/mm 2 by the mean GCL thickness for the corresponding grid square in the age-similar subcohort.
Histological INL data were obtained from Masri et al. (2021), who reported sum-of-exponential equations describing cell density for various cell populations stratifying to different layers within the INL as a function of eccentricity. These equations were used to calculate cell densities at 5 μm intervals for each cell type residing within the INL; at each eccentricity, glycinergic and GABAergic amacrine cell densities were summed to derive total amacrine cell density, and OFF midget bipolar, ON bipolar, DB3a, and DB3b cell densities were summed to derive total bipolar cell density. These were then summed with total horizontal and Müller cell densities to obtain total DAPI-stained cell density, and given that 10.6% of INL cells remain unlabeled by DAPI (Masri et al., 2021), this correction factor was applied to determine the density of all cells within the INL in cells/mm 2 . Relative proportions of total amacrine, total bipolar cell, total horizontal cell, and Müller cell densities were calculated from total INL cell densities. Lastly, histological photoreceptor data were obtained from , who reported cone and rod inner segment density in cells/mm 2 as a function of eccentricity along each principal meridian.
However, due to the co-stratification of cone and rod inner segments, the corresponding ISOS represents both cell classes, with varying proportions along the en face plane with increasing eccentricity. As such, the relative contribution of cone and rod inner segments to ISOS thickness would depend on cone and rod diameter, which themselves vary with eccentricity Jonas et al., 1992;Scoles et al., 2014). Outside of the rod-free zones, reported as between 126 and 200 μm and between different principal meridians , cone diameter as a function of eccentricity was determined using Akima spline interpolation for each principal meridian from data reported by Scoles et al. (2014). Assuming maximum packing with hexagonal inner segment cross sections, the relative en face area covered by cone inner segments was calculated using corresponding cone density data at 5 μm eccentricity intervals for each principal meridian. The remainder of the surface area was assumed to be covered by rod inner segments, and once again assuming hexagonal inner segment cross sections, rod diameter was calculated based on rod density data. However, as rods are relatively sparse at the central retina, calculations at central locations may have resulted in overestimation of rod diameter, and as such linear regression models were fit through the data. From estimated rod diameters, the ratio of rod to cone inner segment areas was used as a correction factor applied to rod density data and enabled the calculation of equivalent cone density as a function of eccentricity per principal meridian. As per GC analyses, equivalent cone density was linearly interpolated between principal meridians and averaged across 60 × 60 grid square locations, and subsequently divided by ISOS thickness measurements from the age-similar subcohort to calculate equivalent cone density in cells/mm 3 . Hierarchical cluster analyses of the GCL, INL, and ISOS were conducted to identify macular locations that demonstrated similar change with age, and that would therefore be suitable to pool together in subsequent regression analyses. Although multiple linear regression analyses did not suggest significant changes in ISOS thickness with age, analyses were still performed as global ISOS measurements averaged across the entire macula may mask location-specific age-related changes. Hierarchical cluster analyses revealed nine statistically separable clusters in the GCL, five clusters in the INL, and three clusters in the ISOS (Figure 2). In the GCL and INL, these clusters demonstrated a concentric, horseshoe appearance indicating asymmetry in retinal thickness measurements along the horizontal axis, closely resembling patterns of eccentricity-dependent variation previously described in histological studies (Curcio & Allen, 1990). While the mean GCL thickness at the center-most cluster was 7.07 μm, this is likely an artifact of segmentation of the RNFL-GCL and GCL-INL boundaries, as GCs are absent at the foveola (Curcio & Allen, 1990). Of interest is the horseshoe pattern in both GCL and INL profiles, and the peaked ISOS thickness at the foveal center with steep reduction from the parafovea to mid-peripheral locations; these patterns of change closely resemble topographic cell density distributions as described in histological studies (Curcio & Allen, 1990;. These similarities in eccentricity-dependent variations provided support for the linkage of high density OCT measurements and histological cell density data. Subsequent regression models pooling data per cluster and decade bracket revealed across the GCL, INL, and ISOS, quadratic models best described age-related change at all locations except for subfoveal locations, where a linear model was preferred (p = .10-.26 at the central macula, p ≤ .0001-.01 otherwise, Table 4). Moreover, slope parameter analyses indicated that while the central-most cluster of the ISOS did not demonstrate notable change in ISOS thickness with age (p = .14), other locations did demonstrate a significant decline with age (p ≤ .0001 for both), suggesting that the results of multiple linear regression analysis were primarily influenced by foveal ISOS measurements. Points of inflection derived for quadratic regression models TA B L E 3 Multiple linear regression analyses between demographic variables and ganglion cell layer (GCL), inner nuclear layer (INL), and inner segment-outer segment (ISOS) thickness measurements. Note: Cluster numbers are labeled per Figure 2, where lower numbers indicate relatively peripheral locations. p Values indicate outcomes of F-test between quadratic and linear regression models, with a significant result indicating quadratic models best fit the data, and points of inflection (x) were derived for quadratic regression models. p Values of N/A indicate clusters where nonsignificant changes with age were detected on linear regression, and therefore comparisons between quadratic and linear regression were not performed. Annual rates of change were then calculated as a percentage of peak retinal thickness predicted from the quadratic or linear models as appropriate. a, quadratic regression coefficient; b, linear regression coefficient; c, constant. Abbreviation: y, years.

F I G U R E 2
Pseudocolor cluster maps (left column) and regression models describing age-related change (right column) in the (a) ganglion cell layer (GCL), (b) inner nuclear layer (INL), and (c) inner segment-outer segment layer (ISOS). In both, locations that are colored the same indicate those that show statistically similar changes with age, with 9, 5, and 3 statistically separable clusters identified for the GCL, INL, and ISOS, respectively. In regression models, data were pooled per the clustered locations in pseudocolor maps and per decade bracket, with mean and standard deviation for each data point depicted. Black stars indicate points of inflection for clusters where quadratic regression models were preferred, and asterisks indicate significant age-regression models. See Table 4 for coefficients describing regression models. TA B L E 5 Differences in ganglion cell layer (GCL), inner nuclear layer (INL), and inner segment-outer segment (ISOS) thicknesses between the age-corrected cohort and age-similar participants derived from the modeling cohort, via Hodges-Lehmann differences in median thicknesses, Bland-Altman analyses, and subsequent linear regression analyses to Bland-Altman data.

Correction using the derived normative models
The normative models describing variations in GCL, INL, and ISOS thickness with demographic features were tested to determine generalizability of these models on other cohorts. With data in the test cohort corrected to a 32.4 year age-equivalent for the GCL and ISOS and to a 46.4 year age-equivalent for the INL, and adjusted for sex and refractive error in INL analyses, comparisons with uncorrected data from age-similar participants in the modeling cohort (chosen to match the respective histological cohorts as closely as possible) revealed no significant differences in GCL and ISOS thicknesses (p = .85 and .28 respectively), but significantly thinner median INL in the corrected cohort (p = .03, Table 5). However, Bland-Altman comparisons revealed no notable bias with 95% limits of agreement falling well within the axial pixel resolution of the Spectralis OCT of 3.87 μm ( Figure 3 and Table 5). While linear regression models were significant across all retinal layers (p < .000-.0024), low coefficients of determination (R 2 ) indicated relatively poor fits of the linear models to the data.
Moreover, considering the 95% prediction intervals, within which 95% of data points are expected to fall, the maximum predicted differences between age-corrected and age-similar thicknesses were 3.71 μm for the GCL, 4.36 μm for the INL, and 1.32 μm for the ISOS. That is, all maximum differences fell within the axial pixel resolution of the Spectralis OCT for the GCL and ISOS, and within the equivalent of 2 pixels difference for the INL. Moreover, plots depicting differences per macular location revealed the largest absolute changes occurred around the central fovea, which is prone to inter-individual differences owing to variations in foveal pit contour (Dubis et al., 2009) (Figure 3). As such, from a practical perspective, the differences in corrected and age-similar thickness measurements can overall be considered to be insignificant.

Cell density derivations from histological data
After verifying the suitability of correction using the normative models to sufficiently minimize interindividual variability, derivations of volumetric cell density factors were performed across the 60 × 60 F I G U R E 3 Bland-Altman plots (left column) and colored location-specific difference plots (right column) describing differences in age-similar and corrected (a) ganglion cell layer (GCL), (b) inner nuclear layer (INL), and (c) inner segment-outer segment layer (ISOS) thickness. In Bland-Altman plots, bias (black dashed lines) and 95% limits of agreement (black dotted lines) relative to the axial pixel resolution of the Spectralis OCT, 3.87 μm (red dotted lines), are shown. Linear regression models through the Bland-Altman plots (blue solid lines) and corresponding 95% prediction intervals (blue shading) are also shown, with the parameters of the linear models shown in blue text.

F I G U R E 4
Percentage differences between tilt-adjusted and unadjusted retinal thickness as a function of B-scan tilt in the (a) ganglion cell layer (GCL), (b) inner nuclear layer (INL), and (c) inner segment-outer segment (ISOS) layer. The quadratic functions fit through the data and corresponding equations and coefficients of determination (R 2 ) are also shown.
grid using OCT data from the age-similar cohorts (varying slightly in mean age and sex distributions relative to corresponding histological cohorts). This was performed to enable the prediction of cell density and numbers from individuals' OCT measurements ( Figure 5, Supporting Dataset 1). While relatively concentric distributions were observed in photoreceptor densities, consistent with , asymmetries in peak and eccentricity-dependent volumetric cell densities were noted in GC data. Along the horizontal axis, peak volumetric densities were observed at .35 mm from the foveal center nasally and .38 mm temporally, with nasal volumetric cell density exceeded temporal density until an eccentricity of 1.38 mm. Meanwhile, along the vertical axis, peak cell density was noted at .33 mm inferiorly and .38 mm superiorly, with inferior volumetric cell density, exceeded superior density until an eccentricity of .93 mm. These appear to reflect asymmetries in areametric GC densities between meridians, especially within .5 mm of the foveal center as depicted in Curcio and Allen (1990), particularly in the absence of notable asymmetry in GCL thickness along the vertical meridian ( Figure 2)

Age-related changes in retinal thickness
While histological studies have assessed age-related changes in cell density to some extent (Curcio et al., 1993;Gao & Hollyfield, 1992;Harman et al., 2000), given the resource-intensive procedures, characterization over a wide age range is not feasible. In contrast, OCT can rapidly acquire in vivo retinal data, enabling more detailed characterization of age-related changes. In cross-sectional OCT studies, age has consistently been demonstrated to influence retinal thickness measurements in various layers Mwanza et al., 2011;Won et al., 2016), and the observed nonlinear patterns of change with decline beginning around the 30th decade are consistent with F I G U R E 5 Volumetric cell densities (in cells/mm 3 ) calculated from histological data and age-similar optical coherence tomography (OCT) data for the corresponding retinal layer in (a) ganglion cells, (b) total inner nuclear layer cells, and (c) photoreceptors across the 60 × 60 grid (left column). Profiles of cell densities along the central horizontal meridian (black line, middle column) and vertical meridian (red line, right column) are also displayed, where T, N, S, and I denote temporal, nasal, superior, and inferior locations respectively. For the inner nuclear layer (INL), as cell density data were only available along the temporal meridian, volumetric cell density was calculated for this meridian then extrapolated to all other angular eccentricities, resulting in the symmetrical pattern observed. See Supporting Dataset 1 for numerical density measurements across the 60 × 60 grid.
previous studies (Tong et al., 2019;Trinh et al., 2020;Yoshioka et al., 2017). While the varying regression models between different clusters within the GCL, INL, and ISOS suggested that absolute rates of change were variable across the macula, the consistency in percentage rates of change across the macula suggests that this does not reflect "true" variations in age-related change with eccentricity, but are related to retinal thicknesses at a given location.  (Holló & Zhou, 2016;Zhang et al., 2016). While consistently slower than longitudinal changes in glaucoma eyes (Hammel et al., 2017;Holló & Zhou, 2016), the varying rates of change in healthy eyes may reflect differences in the studied inner retinal complexes or ages investigated; while age ranges of

Demographic factors influencing retinal thicknesses
Previous studies have variably reported reduced GCL thickness with increasing myopia and no effect with refractive error (Mwanza et al., 2011;Omoto et al., 2020;Sezgin Akcay et al., 2017;Tong et al., 2020), and a single study investigating the INL reported significant reductions in high myopia (Kim et al., 2020). The apparent reduction with increasing myopia may be related to transverse magnification effects rather than indicative of "true" reductions in retinal thickness with myopia (Lal et al., 2021;Lee et al., 2021;Omoto et al., 2020). As this study applied relatively constrained refractive error criteria, corrections for transverse magnification may be more pertinent for highly myopic eyes. Similarly, this study found no significant relationship between ethnicity and GCL thickness, consistent with Perez et al. (2021); however, other studies have reported significant differences in inner retinal thickness between different ethnicities (Khawaja et al., 2020;Poh et al., 2020), perhaps due to various ethnicities included across studies. In the present study, participants were only categorized as Asian or white, representing the patient demographic at Centre for Eye Health; a more ethnically diverse cohort may elucidate the relationship between ethnicity and inner retinal thicknesses in greater detail.
Our finding of a significant association between sex and INL thickness appears consistent with previous studies observing thinner INL measurements in women (Ooto et al., 2011;Won et al., 2016). Meanwhile, the present study did not observe a significant association between sex and ISOS thickness; however, previous studies have variably reported significantly thicker ISOS measurements in men and no sex-dependent variations (Chua et al., 2019;Ooto et al., 2011;Won et al., 2016). Interestingly, Hermenean et al. (2020) (Chen et al., 2022;Das et al., 2018;Tsuboi et al., 2021), and INL thickness and perfusion have been significantly correlated with functional outputs such as visual acuity and visual field sensitivity (Mokrane et al., 2021;Tsuboi et al., 2021). Meanwhile, ex vivo studies of human and rodent retinas have revealed changes to the morphology and protein expression of Müller cells and amacrine cells (Gastinger et al., 2006;Hammes et al., 1995;Mizutani et al., 1998), and bipolar cell dysfunction has been observed as a consequence of GC disease and photoreceptor degeneration (Kosta et al., 2021;Shen et al., 2019). Moreover, cone bipolar cell loss and dysfunction in neurochemical signaling has been reported in rat models of retinal ischemia and reperfusion, with reductions in OCT-derived INL thickness in human eyes with retinal ischemia appearing to reflect these changes (Kalloniatis et al., 2022;Sun et al., 2007)

Combined role of OCT and emerging imaging technologies
While the distributions of OCT-derived retinal thickness measurements appear to correspond to histological distributions of corresponding retinal cells, retinal thickness measurements do not necessarily consider eccentricity-dependent variations of cell populations residing within the same layer. A classic example includes the distribution of rods and cones, with the peak in cone density at the foveal center and peak rod density reached at 1.2-1.7 mm eccentricity Scoles et al., 2014). This difference, in conjunction with possible effects of age-and sex-distribution differences between historical histological and this study's OCT cohorts, is likely to contribute at least in part to apparent discrepancies between age-related decline in OCT-derived ISOS thickness, which was −.06% in this study, compared to an equivalent annual reduction of −0.54% a year in rod density as described by Curcio et al. (1993). Similarly, the distributions of rod and cone ON bipolar cells generally follow this eccentricitydependent trend, with more rod bipolar cells located more peripherally (Lee et al., 2019;Masri et al., 2021). As such, the relative contribution of each co-stratified cell type to a given retinal thickness measurement can only be estimated based on proportional surface area calculated from cell densities and diameters.
The development of adaptive optics-OCT has enabled the resolution of numerous retinal cell types (Liu et al., 2017;Rossi et al., 2017;Scoles et al., 2014;Wells-Gray et al., 2016), and combining the identification of different cell types en face and retinal thicknesses may provide more precise, individualized estimates of cell densities. Moreover, adaptive optics studies have observed enlargement of GC somas in glaucoma Soltanian-Zadeh et al., 2021), suggesting that modifications to cell densities calculated from GCL thickness changes may be required to obtain more precise cell calculations. While the investigations in the present study provide a useful foundation for calculating cell densities from OCT, further work on cell morphology changes in disease would be highly beneficial to determine appropriateness in disease settings.

Limitations
In addition to those identified before, a key limitation of this study's design is its cross-sectional comparison of histological and OCT data from different human cohorts, including slight differences in age and sex distributions between them due to the limited sample size of the histological cohorts affecting more direct matching of participants.
While a longitudinal study design following a reasonably large human sample, including histological and OCT data, would be ideal, the ex vivo methods required to process histological data and the large age range required to sufficiently characterize change over an adult human's expected lifetime limits the practicality of such a longitudinal study design. Moreover, the comparability of corrected and age-similar retinal thickness measurements in our study indicates that the derived regression models appear to characterize age-related change appropriately. Another limitation includes reliance on human histological studies, generally including few participants from a limited age range, so cell density data obtained from these studies may not be entirely reflective of the spectrum of normative cell densities. GC and photoreceptor cell densities were also only available along principal meridians, and INL cell densities were only available along the temporal meridian, resulting in assumptions of linear change between meridians and similar densities regardless of angular eccentricity respectively.
Furthermore, calculations of rod diameter performed in this study were based on assumptions of cone and rod density and packing and appeared to overestimate rod diameter centrally compared to the limited information on human rod diameter (Jonas et al., 1992). Lastly, as photoreceptor outer segments reportedly orient toward the nodal point of the human eye (Enoch, 1972;Laties & Enoch, 1971), there may be small mismatches between retinal thickness measurements extracted perpendicular to B-scan tilt, as performed in this study, and histological photoreceptor thicknesses. Further detailed characterizations of quantitative cell density and morphological parameters with eccentricity and between principal meridians, perhaps with adaptive optics OCT, may help overcome these concerns in future work.

CONCLUSION
This study sought to develop a high spatial density OCT-based model of the neural retinal layers, with appropriate correction for age and sex, to enable the calculation of cell density parameters from OCT.
With correction factors derived from these models, the resultant minimal inter-cohort variability indicates that OCT retinal thicknesses could be applied as surrogate structural parameters to cell densities derived from histological studies. Furthermore, the derivation of volumetric cell density, based on histological cell densities and retinal thickness measurements, allows for the calculation of retinal cell density from live human eyes with relative ease, in turn enabling greater translation of knowledge of cellular processes to clinical and research investigations.