Repeatability and reliability of cone density measures in patients with retinitis pigmentosa (RP) or Usher syndrome using adaptive optics scanning light ophthalmoscopy (AOSLO)

Abstract

Purpose : To assess the intraobserver repeatability and interobserver reliability of AOSLO-derived cone density measures in patients with RP or Usher syndrome.

Methods : Twelve subjects (ages: 9-59 years, 9F) with genetically-confirmed RP or Usher syndrome were imaged with AOSLO. Regions of interest (ROI) within 1° of the peak cone density locus (55x55µm) and in the parafovea (85x85µm) were extracted for grading. Four observers, masked to ROI location and pathology, completed semi-automated cone identification four times each for 10 foveal confocal ROIs (avg. eccentricity: 0.64°) and 10 pairs of parafoveal confocal and split-detection ROIs (avg. eccentricity: 3.32°) (30 total ROIs). Five days separated each grading session, and each session included nine distractor ROIs minimizing intersession learning or image recognition (39 ROIs/session). For each ROI, observers rated image quality from 1 (worst) to 5 (best). A signal-to-noise (SNR) based measure of image quality was also derived.

Results : Average cone density across observers was 80,769 cones/mm² for foveal confocal ROIs, 44,755 cones/mm² for parafoveal confocal ROIs, and 29,693 cones/mm² for parafoveal split-detection ROIs. All observers had excellent repeatability of cone density measures for foveal confocal and parafoveal split-detection ROIs and worse repeatability for parafoveal confocal ROIs (Table 1). A similar pattern was seen for interobserver reliability. Observer quality scores were highest for foveal confocal ROIs and lowest for parafoveal confocal ROIs, though SNR did not show the same trend (Table 2). There was a significant negative correlation between interobserver variance and image quality score (r = -0.87, p < 0.0001), but not image SNR (r = -0.31, p = 0.0917).

Conclusions : Greater repeatability and reliability in parafoveal cone density measures derived from split-detection ROIs relative to confocal ROIs at the same locations suggest that split-detection is more suitable for estimating parafoveal cone density in patients with RP or Usher syndrome. Discrepancy between subjective grading of image quality and SNR score suggests an AI-based approach may be more viable as a predictive tool in AOSLO workflows. Regardless, subjective grading of image quality may be useful in analyses of longitudinal measures of cone density in trials utilizing AOSLO.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.