SD-OCT B-scan based Automated Hypertransmission Detection and Quantification in Dry Age-Related Macular Degeneration

Abstract

Purpose : Early biomarker quantification and characterization in dry age-related macular degeneration (AMD) is an important opportunity for risk stratification and personalized treatment decision-making. Hypertransmission defects are one such pre-atrophic biomarker with high geographic atrophy (GA) conversion. Historically, hypertransmission has been quantified using a special OCT acquisition protocol and targeted slab-based assessment for enhanced visualization. However, hypertransmission signals are present on standard OCT B-scans, and automated segmentation at this level could provide greater accessibility to this quantitative biomarker of disease burden. The purpose of this analysis is to assess the feasibility and performance of a machine learning-based, OCT B-scan-level quantitative segmentation system for hypertransmission.

Methods : In this IRB-approved retrospective study, eyes with dry AMD that had been previously segmented for GA at the OCT B-scan level were included. Segmented GA regions were utilized for hypertransmission mask generation, such that these masks extended 60 pixels below the Bruchs membrane for Spectralis and 120 pixels below for Cirrus scans. Two separate models were developed, one for Spectralis (Heidelberg) and one for Cirrus (Zeiss) scans. The training data comprised 90,000 images for Spectralis and 100,000 images for Cirrus scans. A modified-Unet architecture with 33 million parameters was trained using a training-testing split of 80:20. Area under the curve (AUC) and intersection over union (IOU) were calculated to assess model performance.

Results : The Spectralis model achieved an AUC of 0.973. The AUC of the Cirrus model was found to be 0.950. The IOU for automated hypertransmission segmentation was calculated at 0.90 for the Spectralis model. This measurement for the Cirrus model was also found to be 0.90.

Conclusions : A high-performance hypertransmission model was achieved using standard OCT B-scans, independent of enface OCT quality. Performance was comparable for Cirrus and Spectralis platforms. This hypertransmission measurement holds significant promise as an accessible biomarker for predicting GA development and progression. Validation studies are underway on larger longitudinal datasets.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

A) Raw OCT B-scan; B) prediction grayscale output; C) prediction mask; D) prediction overlay (red); E) groundtruth (yellow) and prediction (red) overlay