Adaptive optics retinal imaging with deep learning cone segmentation enables area-based analysis in RHO-associated retinitis pigmentosa

Abstract

Purpose : Adaptive optics (AO) retinal imaging with single-cell resolution has enabled the extraction of quantitative information about the human photoreceptor mosaic, but characterizing small changes remains challenging. In diseases such as RHO-associated retinitis pigmentosa (RP), where rod loss precedes cone loss, the structural integrity of the cone mosaic could be altered due to rod-dependent cone survival. In this study, we investigate area-based metrics enabled by cone segmentation and their utility in characterizing subtle disruptions in the cone mosaic of patients with RP.

Methods : Non-confocal split detection retinal images were acquired using a custom-built AO scanning light ophthalmoscope. Images from two patients with molecularly-confirmed RHO-associated RP were compared to images from two healthy volunteers at matched eccentricities (2-3 mm temporal from fovea). Cone boundaries were localized using deep learning segmentation (PMID: 31701095), and cone centers were derived from the extracted cone boundaries. Cone boundaries were used to generate area-based metrics such as cone size (area within each cone) and percentage of retinal area occupied by cones, and cone centers were used to generate point-based metrics such as cone density.

Results : Subtle disruptions to the cone mosaic were observed in patients with RHO-associated RP, including variable cone size and local inter-cone area (area enclosed by nearest neighbors). There was a 24.5% decrease in cone density in patients with RP (healthy: average +/- SD, 10747 +/- 947 cells/mm²; RP: 8631 +/- 656 cells/mm²; n=4,216 cones; p < 0.01, two-tailed t-test). The percentage of retinal area occupied by cones was similar across healthy (45.6 +/- 3.8%) and RP (46.0 +/- 3.2%), which could be explained by a 24.1% increase in the average size of cones (healthy: 42.4 +/- 8.1 µm²; RP: 52.7 +/- 10.4 µm²; n=4,216 cones; p < 0.01, two-tailed t-test). These metrics show how area-based cone metrics enrich the interpretation of changes in cone density and help characterize subtle structural changes to the cone mosaic.

Conclusions : Diseases involving progressive photoreceptor loss, such as RP, may benefit from the introduction of area-based metrics alongside existing point-based metrics for characterizing early changes to the cone photoreceptors secondary to rod degeneration.

This is a 2021 ARVO Annual Meeting abstract.