Automated Identification and Quantification of Ellipsoid Zone At-Risk Using Deep Learning on SD-OCT in Dry AMD

Abstract

Purpose : In outer retinal diseases, such as age-related macular degeneration (AMD), attenuation of outer retinal layers and the retinal pigment epithelium can result in subsequent vision loss. Identifying eyes at greatest risk for progression could be crucial for selecting patients who would benefit most from future treatment and for enriching clinical trial populations. The purpose of this study was to develop a deep learning (DL) model for detection and quantification of abnormal ellipsoid zone (EZ) bands (i.e., EZ At-Risk) that might be at-risk for future disease progression in dry AMD.

Methods : In this IRB-approved retrospective analysis, 100,266 spectral domain optical coherence tomography (SD-OCT) B-scans from 341 patients with dry AMD were utilized for model building and testing. Ground truth annotations for retinal layers were automatically segmented via a previously-validated machine learning model and ground truth masks for EZ At-Risk were generated following accuracy confirmation by trained expert readers. EZ At-Risk was defined as regions of EZ-RPE thickness of less than 10 microns, excluding areas that had already progressed to GA. The DL model was trained using a U-Net architecture with approximately 20 million parameters and 41 layers. Eighty percent of the initial dataset was used for testing, ten percent was used for periodic validation during training, and ten percent was used for hold-out testing of the final model.

Results : Automatic EZ At-Risk detection in a single OCT B-scan had an accuracy of 87%, sensitivity of 96%, and specificity of 73%. Automated EZ At-Risk quantification measurement by pixel across the macular cube achieved an accuracy of 90%, sensitivity of 90%, and specificity of 84%. The percentage area of regions with EZ At-Risk automatically detected using the DL-based model showed an ICC of 0.83 (p<0.001), compared to the thrice-validated ground truth model.

Conclusions : The DL model successfully detected areas of EZ abnormality with high performance characteristics. These results are promising for further study for automated detection of areas of photoreceptor loss and potential progression risk in dry AMD. Future research will focus on further validation and assessment of the potential of EZ At-RIsk as a predictive biomarker.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.