Modeling rod and cone photoreceptor cell survival in vivo using optical coherence tomography – Scientific Reports

In this paper, we present a new computational method for utilizing in vivo retinal layer thickness data captured by optical coherence tomography (OCT) to estimate the degree of rod and cone photoreceptor loss. Our model is registered to photoreceptor density data from landmark literature on human photoreceptor topography¹ and takes advantage of freely-available software^10,11 that performs retinal sublayer segmentation to isolate the ONL thickness for quantification of photoreceptor survival. Rather than attempting to quantify the absolute density of photoreceptor cells, which varies widely from person to person, the model uses the relative proportion of rods and cones because these relative proportions are much more consistent across individuals. In doing so, the relative loss of rod and cones can be estimated by comparing ONL thickness to healthy controls. To demonstrate potential clinical utility of this model, we used it to estimate the degree of photoreceptor loss using a dataset of patients with molecularly confirmed STGD1. Leveraging retrospective OCT data acquired during routine clinical care, the model estimates survival of each photoreceptor cell type in STGD1 patients and reveals subtle loss that may not be readily apparent on qualitative analysis even by expert clinicians.

Rods and cones exhibit differential survival patterns across various forms of retinal degeneration, including age-related macular degeneration and rarer conditions such as inherited retinal diseases. Distinguishing between cone death and rod death can help the clinician narrow the range of considered diagnoses²⁰, unravel mechanisms of progression², and categorize patients by disease stage. Despite the importance of this problem, quantifying the degree of rod and cone loss in vivo remains a substantial challenge using available retinal imaging. Advances in adaptive-optics (AO) coupled with existing modalities, such as OCT, scanning laser ophthalmoscopy (SLO) or flood-illuminated ophthalmoscopy, have enabled quantification of photoreceptor cells. However, AO systems are expensive, not widely available, technically challenging to use, easily subject to artifacts such as patient movement, and typically restricted to visualizing a few degrees of the retina at a time. Moreover, commercially available AO systems can only reliably count cones but not rods^21,22,23. Furthermore, due to optical limitations, these systems have difficulty resolving cones in the fovea, where cone density is highest and the most important for visual prognosis.

Compared to AO-imaging, OCT offers a wider field of view, faster acquisition, and greater familiarity among physicians and photographers due to its widespread use. The model proposed in this paper enables the estimation of rod and cone loss using readily available OCT data alone. An OCT-based computational model of photoreceptor loss has numerous advantages compared to AO, including its ability to analyze large, retrospectively collected OCT datasets acquired during routine clinical practice, the availability of longitudinal OCT data (in many cases over a decade or more) to compare changes over time, and clinician familiarity with this technology to facilitate interpretation of model data with structural information conveyed in the OCT B-scans (e.g., Fig. 4).

Our model proposed here relies on several assumptions. We chose to build our model on the publicly available mean cell density data from Curcio and coworkers’ landmark 1990 study of retinal anatomy¹. The accuracy of this dataset was recently confirmed using a larger cohort⁷ and imaged with a custom AO-SLO system²⁴. This follow up study concluded that, while the average cell densities originally reported in Ref.¹ are slightly higher than those measured by AO-SLO, especially within 300 µm of the fovea, the overall pattern of cell densities were highly concordant between the two studies⁷. It is possible that registration to other density datasets might yield different estimations of rod and cone loss. Our model could readily register to alternative datasets or incorporate variability in expected cellular composition as well as averages for a study.

As a first approximation of rod and cone loss for this model, we focused on two regions based on the ETDRS grid, the central subfield and the outer ring. Clinicians are familiar with using the ETDRS grid for evaluating macular pathologies since the grid was first introduced in 1991¹⁵. Cones predominate in the center subfield and rods predominate in the outer subfield, allowing these two regions to serve as biomarkers for cone and rod health, respectively. By excluding the inner ring covering the perifoveal region, we bypassed the region where the proportion of rods and cones is expected to vary the most between individuals. Future versions of our model could refine the areas beyond just the center subfield and outer ring. For example, instead of analyzing changes within ETDRS regions, which can be subject to floor effects²⁵, our model could be extended to estimate expected rod and cone survival for all A-scans, making the model more robust for modeling regional changes or patchy areas of loss. Our current model presupposes that the proportional loss of rods and cones is globally consistent. For some diseases, such as AMD, however, the pattern of loss is localized, either through secondary cell death (i.e., loss of support from CC and RPE, as in patchy loss observed in choroideremia), or in patterned, local loss of rods in early AMD^26,27. Thus, our current model is better suited to cases where cell death is relatively symmetric and continuous across the retina.

Our model also makes simple assumptions regarding the direct correlation between retinal thickness and proportion of rod and cone loss per region. However, the process of retinal degeneration is complex and nonlinear, involving multiple processes. Scarring and macular edema, for example, can increase retinal thickness, suggesting that obvious gliosis or edema are exclusion criteria for our model. Even apart from retinal disease, the gradual loss of cones in normal aging may result in paradoxical thickening of the retina. A recent study paired OCT with AO-SLO to compare the combined thickness of the HFL and ONL to photoreceptor density in a group of younger eyes (8 subjects, mean age 27.2 years) to a group of older eyes (8 subjects, mean age 56.2 years)²⁸. That study indicated that, contrary to expectation, the combined HFL and ONL thickens with age, despite the loss of cones. (However, a different group, using flood illuminated AO, failed to observe a statistically significant loss of cone density in a comparison with similar ages and larger sample size²⁹.) Neither of these studies, however, presented a null model of the expectation of ONL thickness lost under cell loss. For similar study designs, our model can be used to construct a null model of ONL thinning based exclusively on expected cell loss. Future versions of the model can also utilize larger datasets including from healthy control patients to make the estimates of photoreceptor loss more statistically robust.

A model similar to ours was implemented by Cideciyan and coworkers (see Supplementary Information in³⁰). Unlike our model, which is based on histology, their prediction model incorporates OCT data from 3 CEP290-LCA patients, where only cones survive. They computed the difference in ONL thickness between these patients and the thickness of controls. This difference represents the proportion of the ONL occupied by rods (see Supplemental Fig. 3 in³⁰). They were then able to compare measurements of rod function in RP patients (as assessed by a custom system for two-color dark-adapted perimetry) against the predicted number of surviving rods.

In summary, we developed a computational model to estimate the proportion of surviving rods and cones using in vivo OCT data alone, without the need for advanced retinal imaging such as AO. Our model has many potential broad applications for research or clinical use. For example, future automated estimation of rod and cone loss could facilitate diagnosis of rod- or cone-specific diseases by OCT alone. Characterizing the rate and pattern of photoreceptor loss may help identify patient outliers within age classes of inherited retinal diseases such as STGD1 to provide new insights into pathophysiology or focus investigations of genetic disease modifiers.