Peak cone density predicted from outer segment length measured on optical coherence tomography

Abstract

Purpose:

To compare peak cone density predicted from outer segment length measured on optical coherence tomography with direct measures of peak cone density from adaptive optics scanning light ophthalmoscopy.

Methods:

Data from 42 healthy participants with direct peak cone density measures and optical coherence tomography line scans available were used in this study. Longitudinal reflectivity profiles were analyzed using two methods of identifying the boundaries of the ellipsoid and interdigitation zones to estimate maximum outer segment length: peak-to-peak and the slope method. These maximum outer segment length values were then used to predict peak cone density using a previously described geometrical model. A comparison between predicted and direct peak cone density measures was then performed.

Results:

The mean bias between observers for estimating maximum outer segment length across methods was less than 2μm. Cone density predicted from the peak-to-peak method against direct cone density measures showed a mean bias of −155 cones/mm² with 43% of participants displaying a 10% difference or less between predicted and direct cone density values. Cone density derived from the slope method showed a mean bias of −24,895 cones/mm² relative to direct cone density measures, with only 24% of participants demonstrating less than a 10% difference between direct and predicted cone density values.

Conclusions:

Predicted foveal cone density derived from peak-to-peak outer segment length measurements using commercial optical coherence tomography show modest agreement with direct measures of peak cone density from adaptive optics scanning light ophthalmoscopy. The methods used here are imperfect predictors of cone density, however, further exploration of this relationship could reveal a clinically relevant marker of cone structure.

Introduction

The human fovea comprises only 0.2% of the total retinal area but is central to arguably the retina’s most important function – to facilitate high acuity vision. The fovea is characterized by several anatomical specializations:¹ It is devoid of inner retinal layers (and lacks inner retinal vasculature) and the central fovea is almost exclusively comprised of cone photoreceptors. Increased cone packing at the fovea results in an increased outer nuclear layer thickness compared to the peripheral retina, with foveal cone cell bodies stacked on top of one another. To accommodate this high density, cone outer segments (OS) in this region are only about 1.5–2 μm in diameter, and are elongated relative to parafoveal cones (by an average of ~19μm).^{2, 3} Disruptions to typical foveal structure occur in a variety of congenital and acquired pathologies. For example, an underdeveloped fovea (i.e., foveal hypoplasia) is common in many inherited retinal disorders, including albinism, aniridia, and achromatopsia.^4–6 The foveal avascular zone is abnormal in patients with diabetic retinopathy,^7–9 sickle cell disease,^{10, 11} and in individuals born prematurely.¹² Additionally, altered foveal cone structure has been observed in patients with a variety of conditions, including but not limited to achromatopsia,^{13, 14} occult macular dystrophy,¹⁵ Usher syndrome,¹⁶ some forms of red-green color blindness,^17–19 and blue cone monochromacy.^{17, 20, 21} There is growing interest in understanding how these and other disruptions to foveal anatomy relate to visual dysfunction.^{22, 23}

Central to these efforts is the ability to noninvasively and quantitatively assess the various anatomical specializations of the fovea. Optical coherence tomography (OCT) has been used extensively to examine foveal pit morphology in a myriad of patient populations.^24–29 Directional OCT allows visualization of the Henle fiber layer of the retina by altering the pupil entry position of the imaging beam and allows for more precise measurement of outer nuclear layer thickness,^{30, 31} while OCT angiography enables assessment of the foveal avascular zone.^32–34 However, foveal cone photoreceptors are of particular interest as the packing density of these cones represent a fundamental limit to visual acuity for a given retina.³⁵ Adaptive optics scanning light ophthalmoscopy (AOSLO) allows direct visualization of the foveal cone mosaic with single-cell resolution, enabling direct measures of cone density and cone spacing.^36–38 Relative to standard clinical modalities, AOSLO imaging is time consuming, requires advanced training to operate, and is not available for widespread use. Additionally, certain populations can be more challenging to image with AOSLO, including individuals with dense cataracts, dry eye, small pupils, large refractive errors, or nystagmus (seen in albinism and aniridia).^{4, 13, 39–42}

With challenges surrounding the use of AOSLO technology and emerging techniques for the diagnosis, monitoring, and treatment of retinal disease, there is interest in validating OCT-based measures of foveal cone structure. The length of foveal cone OS has a theoretical relationship to foveal cone density, based on the observation that cone photoreceptors have a constant volume and roughly triangular packing geometry.⁴³ Using a model to convert foveal cone OS length to an estimate of foveal cone density, a previous study found good agreement between the OCT-based estimates of cone density and direct AOSLO-based measurements in individuals with albinism, but not healthy controls.⁴³ In a similar study, cone spacing was significantly correlated to OS length in individuals with inherited retinal disease but not healthy controls.⁴⁴ However, another study found a significant inverse relationship between cone spacing Z-score and OS+retinal pigment epithelium thickness in healthy controls, but not in individuals with rod-cone degeneration,.⁴⁵ More recently, a strong correlation between maximum OS length and peak cone density was observed in four healthy eyes.⁴⁶ The inability to establish a consistent relationship between OS length and cone density in healthy retinae may be attributed to errors in the method for measuring OS length, inaccurate cone density measurements (especially in participants with higher cone densities), or spatial misalignment in the retinal location of the different measures. Here, we reexamine the relationship between maximum OS length and peak cone density in a larger sample of participants with normal vision. We utilize a revised method for measuring and calculating maximum OS length from OCT as well as an alternative AOSLO imaging protocol to better visualize the foveal cone mosaic.^{38, 47}

Methods

Participants

This study was performed following the tenets of the Declaration of Helsinki and was approved by the Institutional Review Board at the Medical College of Wisconsin (PRO00030741). Data from a total of 42 participants (15 male, 27 female; age range: 12–61 years) who previously agreed to have their data used for unspecified future research (i.e., banked) were utilized for this study. Axial length was measured using a Zeiss IOL Master 500 (Carl Zeiss Meditec, Dublin CA, USA). The pupil was dilated and accommodation was suspended prior to imaging with one drop of 2.5% phenylephrine hydrochloride (Akorn, Lake Forest, IL, USA) and one drop of 1% tropicamide (Akorn, Lake Forest, IL, USA) in participants age 18 or older or one drop of 1% cyclopentolate hydrochloride (Cyclomydril, Alcon Laboratories Inc., Fort Worth, TX, USA) in participants under age 18.

Measuring peak cone density with AOSLO

AOSLO peak cone density measures from the participants in this study have been analyzed to evaluate interocular symmetry³⁸ and to produce average peak cone density measures across five raters⁴⁸ that are described in detail elsewhere. Most (40 of 42) participants had OCTs acquired on the same day as AOSLO imaging. The difference in time between OCT and AOSLO imaging for the other two participants was three and six months. In 34 participants, the right eye was selected for analysis and enabled comparison to direct AOSLO peak cone density measures averaged across five graders.⁴⁸ However, there were eight participants with poor OCT quality and/or lack of a B-scan capturing the area of peak OS length in the right eye. Therefore, the left eye was used for analysis and was compared to direct AOSLO peak cone density measures from one rater for the same eye.³⁸

AOSLO imaging and processing of the foveal cone mosaic were completed as a part of a previous study.³⁸ Briefly, foveal cone imaging was attempted in both eyes of each participant using the confocal modality of a custom AOSLO.³⁶ Imaging sequences (150 – 200 frames) were collected using a 790nm super luminescent diode to obtain a 1.5° square grid encompassing the region of peak cone density (1° field-of-view, 0.5° intervals). To ensure adequate visualization of the foveal cone mosaic, either a time series or through-focus imaging protocol was used for further averaging during post-processing, as discussed previously.³⁸ In circumstances where the smallest foveal cones were not resolved, additional imaging was collected using a visible light source (680nm), smaller field of view (0.5° or 0.75°), and a sub-Airy disk pinhole (0.5–0.7 Airy disk diameter).

Each imaging sequence was processed and averaged to create high-resolution images^{49, 50} that were semi-automatically aligned ⁵¹ into a montage of the retinal mosaic using custom software. The alignment of individual images within the retinal montage were manually corrected using Adobe Photoshop CS6 (Adobe Systems, Inc., San Jose, CA, USA). Images encompassing the area of peak cone density underwent additional processing to produce a single image of the foveal cone mosaic with high clarity.³⁸ A 300×300μm² region of interest was extracted after accounting for differences in lateral image scale due to axial length.³⁶ Cones were identified semi-automatically to produce a cone coordinate matrix (Mosaic Analytics, Translational Imaging Innovations, Hickory, NC) for analysis of cone topography. Cone density was calculated for every pixel within the image, averaged across a window size sampling 150 cones, and summed across the image to create a density matrix (https://github.com/OCVL/Metricks, version number 1.5.7)³⁸ that included the coordinate location and value of peak cone density.

Measuring OS length and predicting peak cone density

OCT image acquisition and processing

Horizontal line scans of the macula were acquired (1000 A-scans/B-scan; 100–200 repeated B-scans; nominal scan width of 7 mm) on each eye using high resolution OCT (Bioptigen, Research Triangle Park, NC). Line scans were registered and averaged to produce a single high-resolution OCT image capturing the region of apparent maximum OS length as described previously.⁵² The number of frames from the raw line scan averaged to produce the final OCT image varied across participants (mean: 5.16 ± 3.69 frames; range: 1 – 17 frames). Averaged OCT images were manually compared to a macular volume scan by one human rater (M.T.A.) to verify that the line scan contained the region with the highest OS length within the retina. Our human rater flagged averaged OCT line scans that were dissimilar in appearance to the macular volume scan B-scan containing the area of maximum OS length. In eight participants, the OCT line scan acquired did not capture the region of maximum OS length, therefore, a single frame from the volumetric scan was utilized. Even though averaging fewer frames (or only utilizing one frame) resulted in increased speckle noise, this was deemed a necessary trade-off to produce an image representative of the retinal location containing the maximum OS length. Images were then scaled laterally based on axial length using previously published methods and subsequently resampled to a common width of 5.26mm (lateral scale: 5.26 μm/px; axial scale: 2.40 μm/px).⁴³ OCT images used for this analysis can be found in Supplement 1.⁵³

OCT image analysis to determine maximum OS length

Custom Java software (OCT Reflectivity Analytics [ORA]⁴³) was used to analyze the longitudinal reflectivity profiles (LRPs) from individual A-scans of the log-scale OCT image to identify retinal layers of interest.⁴³ We defined OS length as the distance between the 2^nd and 3^rd outermost hyperreflective bands in the retina, corresponding to the ellipsoid zone (EZ) and interdigitation zone (IZ), respectively (Fig. 1).⁵⁴ OS length was measured using two different approaches, that we define as peak-to-peak and slope methods. The peak-to-peak distance between the EZ and IZ is a common practice for measuring the OS layer in OCT images.^{43, 55, 56} However, peak-to-peak distance may not include the full length of the OS. To mitigate this, other groups have suggested utilizing the boundaries of the bands to more accurately evaluate axial distances measured with OCT.^{57, 58} These boundaries represent the maximum slope identified on the second derivative of the LRP (thus we refer to this as the slope method). For the EZ band, observers were instructed to identify the inner boundary of the EZ band and the outer boundary of the IZ band. Thus, the distance between the EZ and IZ bands measured using the slope method will have a greater distance than measurements utilizing the peak-to-peak method.

Figure 1.

Method for extracting OS length using the peak-to-peak and slope methods. The foveal OCT image on the left has the ellipsoid zone (EZ) band highlighted in blue and the interdigitation zone (IZ) band highlighted in orange, along with a white rectangular box representing an individual LRP used for analysis. OS length was measured as the distance between the EZ (blue) and IZ (orange) bands, using two methods that are shown on the right. A portion of the LRP containing the EZ (blue) and IZ (orange) bands has been enhanced to display the point of peak pixel intensity as a red star and slope point of interest as an open red circle for either band. The pixel distance between the EZ and IZ peaks (red stars) was used to quantify OS length for the peak-to-peak method, while the distance between EZ and IZ slope points of interest (open red circles) was used for the slope method.

Two observers (A.U. and H.M.) identified peaks and slope points in the LRPs generated from each OCT scan. First, an LRP (26.3μm, or 5pixels) was manually placed at the location where OS length appeared greatest. To fully capture the region of foveal OS elongation, five consecutive LRPs (spaced ~20μm or ~4pixels apart) were generated on either side of the initial LRP, producing eleven total LRPs over a 200μm (or 38 pixel) region. For each LRP, peak EZ and IZ points (peak-to-peak method) were always chosen, but slope points within the EZ and IZ were selected only if one appeared within the desired band. If no slope points within the EZ and/or IZ bands were selected in a given LRP, the peak points for the EZ and/or IZ were used for the slope estimates of OS length, while the peak-to-peak method used the peak EZ and IZ points to predict OS length. This analysis was conducted using MATLAB R2020b (MathWorks, Natick, MA, USA) and output the raw maximum OS length measured using either method for each OCT. Maximum OS lengths for each method determined by either observer were assessed for interobserver reproducibility using intraclass correlation⁵⁹ and Bland-Altman analysis.⁶⁰ The data were then averaged between the two observers to provide a predicted value of peak cone density for each method. Spearman correlation⁶¹ was used in this analysis based on the non-linear relationship between OS length and cone density, as observed previously.⁴³ Cone density predicted from maximum OS length was compared to direct measures of peak cone density using Bland-Altman analysis,⁶⁰ and the percent difference between predicted and direct cone density measures for either method were plotted as histograms.⁶² All statistical analyses were completed using R (v4.2.0; R Foundation for Statistical Computing, Vienna, Austria).⁶³

Predicting peak cone density from OS length measurements

To convert the above-measured OS length measurements to a predicted value for peak cone density, we used the following previously described model⁴³ (equation 1):

$$\mathrm{D}=\mathrm{K}\left({\mathrm{A}{h}_{OS}}^3+{\mathrm{B}{h}_{OS}}^2+\mathrm{C}{h}_{OS}\right)$$ (1)

where $\mathrm{D}=\text{peak}\text{cone}\text{density}$ (in cones/mm²) with constants $\mathrm{A}=152.1111$, $\mathrm{B}=3.7419\mathrm{C}=0.0230$, and $\mathrm{K}=(\sqrt{3}\times \mathrm{\pi })/6\times \left(136.6\times {10}^{-9}\right)$; and $h_{OS}=\text{maximum}\text{OS}\text{length}$ (in cm). The model is based on the following assumptions that was optimized to a dataset from individuals with albinism:

1. The OS is cylindrical and has a constant volume of 136.6μm³ across participants.^3,64,65

2. The ratio of OS width to inner segment width varies linearly as a function of OS length.⁶⁶

3. The cone mosaic is arranged in a perfect crystalline lattice.⁶⁷

These parameters, though optimized on a cohort of individuals with albinism,⁴³ should apply to all foveal cones with typical dimensions in a contiguous packing array. It is important to note that this model assumes the location of maximum OS length measured on OCT coincides with the location of peak cone density measured on AOSLO – an assumption not always physiologically valid. No corrections were made in this study to systematically account for differences in retinal location between maximum OS length measured on OCT and peak cone density measured on AOSLO. However, there are key differences in methodology between the analysis used to generate the model described above and this study. We utilized raw maximum OS length measures extracted from LRPs over a 200 μm region – we did not use a Gaussian fit of the distances marked between the EZ and IZ bands over a 500 μm region.⁴³ A smaller region of the retina was analysed in this study to better capture the change in OS length that occurs in healthy foveae. OS length was not fit to a Gaussian curve in this analysis due to a poor fit over the retinal region assessed where differences in raw and Gaussian-fit maximum OS length varied upwards of 10 μm.

Results

Direct measures of peak cone density from AOSLO images

Results for maximum OS length and peak cone density values are shown in Table 1. Peak cone density values measured on AOSLO reported here have been published previously.^{38, 48} Peak cone density ranged from 121,319 to 239,410 cones/mm² (mean ± SD: 172,318 ± 23,898 cones/mm²). For further discussion on AOSLO results, see Cava et al., (2020)³⁸ and Wynne et al., (2022).⁴⁸

Table 1:

Demographic data and results for all participants.

ID	Sex	Age	AL	OS Length: Peak-to-Peak	OS Length: Slope	PCD: Peak-to-Peak	PCD: Slope	PCD: Direct
JC_0007	M	42	27.49	47.33	49.13	169,978	187,229	141,521
JC_0077	F	14	24.15	45.54	47.93	153,813	175,606	160,493
JC_0200*	M	30	24.66	43.14	50.33	133,891	199,351	162,200
JC_0616	M	30	24.24	49.13	51.53	187,229	211,983	180,678
JC_0878	F	12	24.02	45.54	50.33	153,813	199,351	161,664
JC_10145	F	53	24.66	46.13	47.93	159,083	175,606	157,194
JC_10220	F	28	22.92	47.33	50.33	169,978	199,351	159,479
JC_10312#	M	18	27.06	48.53	48.53	181,356	181,356	129,349
JC_10339	F	30	23.55	40.74	43.14	115,756	133,891	157,355
JC_10549	M	26	23.99	44.94	47.93	148,661	175,606	167,277
JC_10567	F	27	22.32	45.54	50.33	153,813	199,351	154,198
JC_10591	M	28	23.58	49.73	49.73	193,227	193,227	218,381
JC_11068	F	32	22.36	44.34	49.13	143,624	187,229	166,767
JC_11103#	F	55	24.46	50.33	56.32	199,351	267,810	181,892
JC_11159	F	29	23.63	47.93	50.33	175,606	199,351	149,018
JC_11295*	F	30	22.95	47.93	52.73	175,606	225,134	179,777
JC_11314	F	23	24.11	47.93	49.13	175,606	187,229	121,319
JC_11321	F	30	23.75	46.13	49.13	159,083	187,229	135,693
JC_11354	F	24	24.53	44.34	47.93	143,624	175,606	158,782
JC_11409*	F	23	23.48	47.33	50.33	169,978	199,351	195,403
JC_11441	F	25	23.2	47.93	50.33	175,606	199,351	152,094
JC_11442*#	M	25	23.64	50.93	55.12	205,603	253,037	239,410
JC_11444	F	27	24.03	50.33	50.33	199,351	199,351	171,732
JC_11467	F	61	23.63	47.93	53.92	175,606	238,815	186,417
JC_11469*	M	27	23.3	49.13	51.53	187,229	211,983	226,029
JC_11575#	F	28	22.93	51.53	52.73	211,983	225,134	182,087
JC_11584*	M	27	24.44	50.33	52.73	199,351	225,134	192,458
JC_11591	M	22	24.44	50.33	52.73	199,351	225,134	175,079
JC_11610*#	M	24	23.96	44.34	44.34	143,624	143,624	191,458
JC_11613	F	28	24.94	50.33	50.33	199,351	199,351	183,543
JC_11617	F	48	24.06	45.54	49.13	153,813	187,229	164,928
JC_11631	M	25	25.64	45.54	45.54	153,813	153,813	160,183
JC_11655#	M	18	25.31	45.54	51.53	153,813	211,983	178,476
JC_11658	F	24	22.18	46.13	47.93	159,083	175,606	196,114
JC_11660*#	F	24	22.74	55.12	55.12	253,037	253,037	182,505
JC_11666	M	25	24.76	48.53	50.33	181,356	199,351	199,040
JC_11685	F	25	25.21	50.33	50.33	199,351	199,351	182,785
JC_11686	F	23	23.51	46.73	46.73	164,470	164,470	182,613
JC_11830	F	49	25.48	50.33	52.73	199,351	225,134	160,535
JC_11857	F	23	23.29	45.54	47.93	153,813	175,606	147,215
JC_11867	M	25	24.11	41.94	44.34	124,605	143,624	179,800
JC_11923#	F	24	23.48	49.13	51.53	187,229	211,983	164,409

AL: Axial length; OS: Outer segment; M: Male; F: Female; PCD: Peak cone density. AOSLO peak cone density measures are labeled Direct. Age is measured in years, AL is measured in mm, OS length is measured in μm, and peak cone density is measured in cones/mm².

The asterisks (*) symbol indicates the left eye was used for analysis and AOSLO peak cone density measures were produced from one rater.³⁸ For all other participants, the right eye was used for analysis and was compared to an average estimate of PCD from five raters.⁴⁸

The pound (#) symbol indicates eyes where a single frame from an OCT volume scan was utilized for analysis.

Interobserver reproducibility of OS length measures

There was moderate reproducibility between observers for maximum OS length values using the peak-to-peak method (ICC = 0.53, 95% CI: 0.31–0.75), while the slope method showed worse reproducibility (ICC=0.44, 95% CI: 0.19–0.68). While each method differed in their overall reproducibility, there was a high degree variance, both with large, overlapping ICC confidence interval ranges. Bland-Altman comparison of the maximum OS length (Fig. 2) across observers revealed mean differences of 1.97μm (95% CI: 1.20μm to 2.74μm) for the peak-to-peak method and 1.60μm (95% CI: 0.60μm to 2.60μm) for the slope method. While the mean bias for either method did not overlap with a mean bias of 0, these data equate to a difference in axial distance on the OCT scan of less than one pixel (2.4μm/pixel scale). Less overall variance in maximum OS length measures were noted across observers when using the peak-to-peak method compared to the slope method. This result aligns with the more subjective nature of the slope method, as the range of options possible varies between zero and multiple second derivative points available to select for each LRP. As such, for subsequent analyses, we averaged the maximum OS lengths from both observers within each method.

Figure 2.

Predicted OS length compared between observers for peak-to-peak and slope methods. A) Bland-Altman comparison of the maximum OS length derived from peak-to-peak reflectivity across observers showed a mean bias of 1.97μm (95% CI: 1.20μm to 2.74μm; upper limit of agreement = 6.80μm, 95% CI: 5.36μm to 8.24μm; lower limit of agreement = −2.87μm, 95% CI: −1.43μm to −4.31μm) B) Bland-Altman comparison of the slope method across observers demonstrated a mean bias of 1.60μm (95% CI: 0.60μm to 2.60μm; upper limit of agreement = 7.90μm, 95% CI: 6.02μm to 9.78μm; lower limit of agreement = −4.71μm, 95% CI: −2.83μm to −6.59μm). In each panel, a solid black line denotes the mean bias in OS length across observers, while black dashed lines represent the upper and lower limits of agreement, and gray shading indicates the 95% confidence intervals for each estimate. The black dotted line indicates no bias (difference of zero).

Predicting peak cone density using the peak-to-peak method

Maximum OS length determined with the averaged peak-to-peak method ranged from 40.74 to 55.12μm (mean ± SD: 47.46 ± 2.81μm), and there was a significant positive relationship between AOSLO-derived peak cone density and maximum OS length (Spearman r = 0.38, p = 0.01). Peak cone density predicted from the peak-to-peak method on OCT ranged from 115,756 to 253,037 cones/mm² (mean ± SD: 172,473 ± 26,419 cones/mm²; Fig. 3A). Peak cone density measures derived from OCT and AOSLO were not significantly different on average (paired t-test, t=0.04, df=41, p=0.97). Bland-Altman analysis (Fig. 3C) for cone density predicted from OS length and direct AOSLO measures demonstrated a mean bias of −155 cones/mm² (95% CI: −9,198 to 8,887 cones/mm²). Eighteen of 42 participants (43%) had predicted (OCT) peak cone density values within 10% of their direct measures from AOSLO (Fig. 3E).

Figure 3.

Evaluation of predicted peak cone density derived from OS length for peak-to-peak and slope methods. A) A significant positive relationship was observed between direct measures of peak cone density and raw, highest OS length from the peak-to-peak method (Spearman r=0.38, p=0.01). B) A significant positive relationship was found between direct measures of peak cone density and maximum OS length derived from the slope method (Spearman r=0.36, p=0.02). The model relating OS length and cone density in panels A and B is denoted as a black dashed line.⁴³ Individual data points are represented as open black circles. C) Bland-Altman analysis comparing peak cone density values predicted with the peak-to-peak method and that directly measured with AOSLO showed a mean bias of −155 cones/mm² (95% CI: −9,198 to 8,888 cones/mm²; upper limit of agreement: 56,721 cones/mm², 95% CI: 39,750 to 73,690 cones/mm²; lower limit of agreement: −57,031 cones/mm², 95% CI: −74,001 to −40,061 cones/mm²). D) Bland-Altman analysis of the predicted and direct measures of cone density for the slope method showed a mean bias of −24,895 cones/mm² (95% CI: −34,378 to −15,410 cones/mm²; upper limit of agreement: 34,761 cones/mm², 95% CI: 16,961 to 52,560 cones/mm²; lower limit of agreement: −84,551 cones/mm², 95% CI: −102,351 to −66,752 cones/mm²). In panels C and D, a solid black line denotes the mean bias between direct and predicted measures of cone density, while black dashed lines represent the upper and lower limits of agreement, and gray shading indicates the 95% confidence intervals for each estimate. E) Eighteen of 42 participants (43%) demonstrated a 10% or less difference in direct (AOSLO) and predicted (OCT) peak cone density measures utilizing the peak-to-peak method. F) When the slope method was used, 10 of 42 participants (24%) demonstrated a difference of 10% or less between measured (AOSLO) and predicted (OCT) peak cone density. In panels E and F, gray shading indicates the area of the graph that shows participants with a 10% or less difference in direct and predicted measures of peak cone density.

Predicting peak cone density using the slope method

As expected, when segmented utilizing the slope method, maximum OS length was on average 2.53μm longer than that from the peak-to-peak method and ranged from 43.14 to 56.32μm (mean ± SD: 49.99 ± 2.83μm). Utilizing the maximum OS length derived from this method, a significant positive relationship was found between AOSLO-derived peak cone density and maximum OS length (Spearman r = 0.36, p=0.02; Fig. 3B). Peak cone density predicted from the slope method on OCT ranged from 133,891 to 267,810 cones/mm² (mean ± SD: 197,213 ± 28,815 cones/mm²) and significantly differed from AOSLO measures on average (paired t-test, t=5.3, df=41, p<0.001). Bland-Altman analysis (Fig. 3D) of the predicted and direct measures of cone density showed the slope method overestimates predicted cone density on average, with a mean bias of −24,895 cones/mm² (95% CI: −34,380 to −15,411 cones/mm²). Only 10 of 42 participants (24%) had predicted (OCT) peak cone density values within 10% of their direct measures from AOSLO (Fig. 3F).

Discussion

Here we sampled the raw maximum OS length from OCT using two methods (peak-to-peak and slope) to predict peak cone density and compare these values to direct measures of the foveal cone mosaic in healthy eyes. Compared to the slope method, the peak-to-peak derived peak cone density values showed better inter-grader reproducibility and better agreement with direct density measures from AOSLO (with a mean average bias of only −155 cones/mm²). However, only 43% of participants demonstrated a 10% or less difference in predicted versus direct cone density measures, and significant differences between the OS-derived and direct measures of cone density occurred for some individuals. As such, OS length measures (at least using our methods) appear to be imperfect predictors of cone density for individual retinae. This work builds upon a previous study,⁴³ where no relationship was found between OS length and cone density in controls. Adjusting the approach to extracting OS length on OCT and cone density on AOSLO from the original study has revealed a positive correlation between these measures. Modifications to the model used for predicting cone density, including the cone metric assessed (cone density versus spacing), as well as to the estimation of maximum OS length could improve the agreement between OCT and AOSLO measures.

The observed differences between OS length methods here highlights the importance of defining in studies how the distance between retinal bands of interest is measured, including the location within each band from which to measure. When utilizing LRPs, there is the consideration of the signal’s maxima, minima, and derivative points as well as the capability to manually determine which point best aligns with the band of interest,^{43, 57} especially when multiple peaks exist for one band or peaks from multiple bands blur together.⁶⁸ At least one other study measured the distance between hyperreflective bands utilizing the edges of each band as discussed here,⁶⁹ and others have noted how differences in measured-distance can occur when measuring from peak-to-peak versus the slope of a band’s signal.^{55, 70} Another study defined OS length with a different slope approach as the distance between the derivative point within the declining slope of the EZ peak and that within the rising slope of the IZ peak.⁴⁶ This approach resulted in a strong correlation observed between maximum OS length and peak cone density.⁴⁶ Our slope method of measuring OS length appears to systematically overestimate peak cone density. Utilization of a method like that from Domdei et al.,⁴⁶ discussed above, could provide more accurate measures of OS length, which in turn may yield better predictions of peak cone density from clinical OCT images.

There are additional factors that may contribute to the weak correlation between predicted and direct peak cone density observed here. First, AOSLO-based density measures are sensitive to undersampling,⁷¹ and while the resolution of most AOSLO images are sufficient to resolve most foveal cones, it is possible that not all cones are resolved (resulting in an underestimation of density in some individuals). Based on previous study,⁴⁸ peak cone density measures from AOSLO are accurate within ~20,000 cones/mm², or a count difference of up to ~20 cones between the coordinate file used for the analysis and the underlying anatomy. Exploring the geometrical relationship between cone spacing measures (which are less affected by undersampling) and OS length may be worthwhile in future studies. Second, IZ band detection is inconsistent across the retina,⁷² and produces variable IZ band placement between observers greater than one pixel (~3μm) in 33% of the OCTs analyzed here. These differences may be reduced with additional training of the OCT observers, or by extracting band position with an automated approach that accounts for the attenuated appearance of the IZ.⁴⁶ Third, the time of day and adaptation state of the eye during OCT acquisition was not controlled. OS length is known to vary (up to 2.7μm)⁷³ throughout the day as the OS are renewed and in response to changing lighting conditions.⁷⁴ It has been previously reported that changes in OS length due to light/dark adaptation do not affect OS length estimates derived from the peak-to-peak method,⁷⁰ though we were not able to evaluate this in our current dataset. Finally, it is possible that foveal cone OS are not perpendicular (but can be coiled or curved) to the retinal pigment epithelium,⁷⁵ which would impact the accuracy of OS length measures from OCT images.

Other factors related to the acquisition and analysis of OCT data may also impact the accuracy of OS length measures. For example the axial scale of OCT images is calculated based on device factors and assumed refractive index values;⁷⁶ differences in refractive index of ocular tissues between individuals would result in small variations in axial scale (and thus OS length measures) that we currently do not consider. Additionally, all OCT analysis was performed on log-transformed images, which compresses higher intensity values in to a smaller range and stretches lower intensities into a larger range, altering the appearance and width of individual retinal bands.⁷⁷ Analysis of raw reflectance data may provide different estimates of foveal OS length.

So, what is the impact of OS length measurement variability when applying this model in a clinical setting? Predicted peak cone density from OS length may have utility for pre-screening images on a binary scale categorizing estimates as “within normal” or “outside normal”, even with the level of error reported here. However, OS length measurement error could be detrimental in studies aiming to examine correlations between predicted cone density and another measure, such as visual acuity. Investigation into ways to improve the repeatability, reproducibility, and accuracy of OS length measures on OCT devices may be required for any clinical application of predicted cone density derived from this model.

Beyond measurement errors, there are issues concerning the integration of anatomical data from two different imaging devices. We independently compared the maximum OS length and peak cone density, which is different than comparing OS length and density values at the exact same retinal location. It has been previously established that the preferred retinal locus of fixation is offset from the location of peak cone density,^{78, 79} and neither location matches that of the area of peak visual sensitivity.⁴⁶ Furthermore, a recent study that did find a significant relationship between OS length, cone density, and retinal sensitivity utilized an offset approach to account for positional differences in the peak position of any of these metrics.⁴⁶ Given that various features of the fovea (maximum OS length, fixation, peak cone density, center of the foveal pit, center of foveal avascular zone) are not always anatomically aligned with one another, utilizing both horizontal and vertical line scans or determining OS length from a volume scan may be more effective for this comparison. Additionally, this could enable comparison of OS length and cone density on an aerial basis, as opposed to just a single point.

Future studies may benefit from the use of other OCT modalities, like visible light OCT or AO-OCT. Visible light OCT has finer axial resolution than standard OCT imaging and produces varied reflectivity profiles of the outer retinal bands related to anatomical features of the photoreceptors and allows for disambiguation of rod versus cones.⁸⁰ When compared with standard OCT measurements, AO-OCT’s improved resolution showed that standard OCTs may overestimate both the thickness of the IZ band and affect measured OS length.⁸¹ With ultra-high-resolution AO-OCT, one can clearly identify the inner and outer segments, which allows for visualization of axial displacement of neighboring photoreceptors and for more accurate measurement of the distance between inner and outer segment bands.^{82, 83} Such technology may also afford the ability to measure cone density and OS length in the same images, avoiding some of the aforementioned registration issues when using multiple devices.

In conclusion, OS length measured on OCT (with our methods) only modestly predicts peak cone density in healthy retinae, but a significant relationship was observed here, unlike the original study.⁴³ This suggests further refinement of the method may elucidate a clinically relevant relationship between maximum OS length and foveal cone structure.

Data availability statement:

The data that support the findings of this study are openly available in FigShare at https://figshare.com/s/0d8cdf2bc4d4e99fdf1c, reference number 53.