Automated Identification of Incomplete and Complete Retinal Epithelial Pigment and Outer Retinal Atrophy Using Machine Learning

Abstract

Purpose : Purpose: To assess and validate a deep learning algorithm to automatically detect incomplete retinal epithelial pigment and outer retina atrophy (iRORA) and complete retinal epithelial pigment and outer retina atrophy (cRORA) in eyes with age-related macular degeneration (AMD).

Methods : Methods: A Resnet18 model was trained to jointly classify the presence of iRORA and cRORA within a given B-scan. The training dataset consisted of OCT B-scans from patients with nonneovascular AMD captured at the Doheny-UCLA Eye Centers were annotated by an experienced grader for the presence of iRORA and/or cRORA: 101 OCT volumes (6,138 OCT B-scans) from subjects with diagnosis of intermediate/late AMD without evidence of macular neovascularization (MNV) and 87 OCT volumes (4,128 OCT B-scans) from 34 subjects with early and intermediate AMD without evidence of iRORA. The algorithm was validated on two separate and independent datasets.

Results : Results: Model performance was quantified in terms of area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC). On an independently collected test set consisting of 1117 volumes sampled from the general population, the model predicted iRORA and cRORA presence within the entire volume with nearly perfect AUROC performance, and AUPRC scores: iRORA: 0.61, 95% CI (0.45, 0.82), cRORA: 0.83, 95% CI (0.68, 0.95). On another independently collected set of 60 OCT B-scans, which was enriched for iRORA and cRORA lesions, the model performed with AUROC (iRORA: 0.68, 95% CI (0.54, 0.81); cRORA: 0.84, 95% CI (0.75, 0.94)) and AUPRC (iRORA: 0.70, 95% CI (0.55, 0.86); cRORA: 0.82, 95% CI (0.70, 0.93)).

Conclusions : Conclusions: A deep learning model can accurately and precisely identify both iRORA and cRORA lesions within a volume scan, despite the performance not being as good for any single OCT B-scan within volume. Notably, the model can achieve similar or better outcomes compared to human graders, thus potentially obviating a laborious and time-consuming annotation process.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Examples of iRORA and cRORA are illustrated in panel A and B, respectively.

 

Model performance using the training data. Receiver Operating Characteristic (ROC- top row) and Precision Recall (PR- bottom row) curves summarizing performance identifying iRORA (left column) and cRORA (right column) within the training set.