Abstract
Purpose :
Given the rarity and diversity of inherited retinal diseases (IRD), patients experience diagnostic delays and unnecessary referrals. AI-assisted image processing can enhance IRD diagnostics and offer insights into the disease's natural history using extensive data. This project utilizes optical coherence tomography (OCT), a noninvasive imaging technique widely used in ophthalmology. Our objective is to employ an automated OCT segmentation algorithm to characterize and classify images from various IRD diagnostic classes. Additionally, we aim to establish correlations between genotypes/phenotypes and retinal features.
Methods :
Our study encompasses 3184 images from 200 IRD patients and 146 controls, including healthy subjects and age-related macular degeneration patients. Automatic segmentation of 6 retinal layers and detection of 9 AMD-related biomarkers including intra- and subretinal fluid, was performed on retinal OCT images using the AI-based available RetinAI Discovery tool. A random forest classifier was used to diagnose macular versus retinal IRD versus controls.
Results :
We document retinal layer thicknesses and OCT biomarkers in a large cohort of clinically and genetically characterized IRD patients, with longitudinal data spanning up to 16 years. Distinct OCT patterns characterize retinal IRD, including reduced perifoveal photoreceptor and outer nuclear layer thicknesses, alongside increased retinal nerve fiber layer thicknesses. Macular IRD shows significant parafoveal photoreceptor atrophy and foveal outer nuclear layer atrophy, among other observed changes. Based on these and other derived features, our OCT-based diagnostic model reaches 86% mean validation accuracy via four-fold cross-validation, underscoring its effectiveness in IRD classification based solely on OCT data. Correlations between OCT features and genotypes reveal significant differences in gene-specific incidences of pathologic markers such as macular edema, among 48 different IRD genes.
Conclusions :
The specificity of information provided by automated segmentation of macular OCT scans in IRD is noteworthy, enabling an acurate classification of images into classes such as macular and retinal IRD. To validate our findings, expanding this analysis to more IRD and control images is crucial. Such findings offer hope for less diagnostic uncertainty in IRD, and the quantitative assessments may contribute to evaluating treatment efficacy in the future.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.