AI-derived layer-specific OCT biomarkers for classification of geographic atrophy

TL;DR

This paper presents an AI-based model that segments specific retinal layers and features in OCT images to classify geographic atrophy, aiding in AMD diagnosis and progression monitoring.

Contribution

The study introduces a novel AI approach for segmenting retinal lesions and distinguishing cRORA from hyperTDs in OCT images, enhancing automated analysis capabilities.

Findings

Achieved voxel-level F1-scores of 0.76/0.13 for INL-OPL subsidence

Achieved voxel-level F1-scores of 0.64/0.15 for EZ-RPE disruption

Achieved pixel-level F1-score of 0.80/0.12 for cRORA vs hyperTDs classification

Abstract

Geographic atrophy (GA) is a key biomarker of dry age-related macular degeneration (AMD) traditionally identified through color fundus photography. Hyper-transmission defects (hyperTDs), a feature highly correlated with GA, have recently gained prominence in optical coherence tomography (OCT) research. OCT offers cross-sectional imaging of the retina, leading to the development of the terms complete retinal pigment epithelium and outer retinal atrophy (cRORA) to describe specific patterns of structural degeneration. Within the definitions of cRORA three critical lesions are implicated: inner nuclear layer and outer plexiform layer (INL-OPL) subsidence, ellipsoid zone and retinal pigment epithelium (EZ-RPE) disruption, and hyperTDs. To enable the automated quantification of retinal atrophy progression, we propose an AI-based model that segments INL-OPL subsidence, EZ-RPE disruption, and hyperTDs. Additionally, we developed an algorithm that leverages these segmentation results to distinguish cRORA from hyperTDs in the absence of GA. We evaluated our approach on a comprehensive dataset of eyes with AMD and healthy eyes, achieving mean voxel-level F1-scores of 0.76/0.13 (mean/standard deviation) for INL-OPL subsidence, 0.64/0.15 for EZ-RPE disruption, and 0.69/0.04 for hyperTDs. For distinguishing cRORA from hyperTDs, we achieved an average pixel-level F1-score of 0.80/0.12 for segment cRORA from hyperTDs. This method demonstrates significant advances in the quantitative analysis of retinal atrophy, offering a promising tool for improved AMD diagnosis and disease progression monitoring.