Reliability of cone density measurements on adaptive optics images in Stargardt disease

Abstract

Purpose: Quantification of the photoreceptor mosaic in normal subjects has been done using manual and automated methods, however in patients with retinal degeneration, cone cell identification becomes increasingly challenging. It is critically important to determine how reliable cone counting measurements are and what factors affect reliability. This knowledge can then be applied to natural history studies and clinical trials that intend to use it to monitor progression and efficacy respectively.

Methods: Twelve patients (8-37 years) with molecularly confirmed Stargardt Disease (STGD) were imaged using a custom built adaptive optics scanning light ophthalmoscope (AOSLO) designed to simultaneously acquire images in confocal and split detection imaging modes. Cone photoreceptors were identified by two graders (G1 and G2) in the confocal and split detector images acquired at 100 parafoveal regions of interest (ROI). All 200 images were graded twice by each observer. Reliability of cone cell identification was compared between trials, between observers and between imaging modes. Statistical significance for all tests was assessed at the 5% level.

Results: The between trial average variance was 11% for G1 and 19% for G2 (p 0.04, paired t-test). The average variance in cones identified between observers was 15%, which did not result in a statistically significant difference in number of cones identified (p=0.06, t-test). There was a significant difference in mean cones identified (p <0.0001, paired t-test) between imaging modes. For both graders, the between trial average variance was 18% for confocal images compared to 3% for split detection images.

Conclusions: Split detection greatly facilitated more reliable cone photoreceptor identification and will be the mode of choice for natural history studies and clinical trials using AOSLO. There was significant variance between the graders, which may in part relate to the fact that this group is the first one to use non-medically trained graders in a reading-center setting to analyse images. Understanding the factors behind these differences, establishing detailed training and standardised protocols, will lead to increased measurement reliability in order to take the first step towards a reading center-based format of AOSLO analysis for large multi-centre trials.