Machine Learning-Based OCT Detection of Geographic Atrophy Lesions in Non-neovascular Age-related Macular Degeneration

Abstract

Purpose : Advanced non-neovascular age-related macular degeneration (dry AMD) is a leading cause of vision loss. Recent clinical trials demonstrate the potential of emerging therapeutics for potential first-in-class treatment for geographic atrophy (GA). Identification and characterization of GA will be critical for optimal management with new treatment options. The purpose of this analysis was to assess the feasibility of an OCT-based automated machine-learning platform for detection of GA lesions.

Methods : This was an IRB-approved retrospective image analysis study evaluating eyes with dry AMD. A machine learning enabled multi-layer segmentation system was utilized to segment Bruch’s membrane (BM), the ellipsoid zone (EZ), and RPE with manual correction. GA was defined by those areas with the EZ, RPE, and BM all intersection (i.e., total loss of RPE and EZ with only BM remaining. Automated training masks were created based on those areas of intersection of the segmentation lines. A U-Net architecture convolutional model was executed and evaluated on training datasets with varying ratios of annotated OCT images containing (positive) and not containing (negative) GA lesions in dry AMD. Overall, 10,271 B-scans from eyes with dry AMD were utilized including 1,456 B-scans that included GA lesions were utilized for training and tested on a self-generated validation set which including 1027 B-scans.

Results : Following model development, the feasibility of automated GA detection on OCT was demonstrated using a machine learning model (Figure 1). The F-score associated with this model which is tested on validation set was 0.68. Interestingly, the qualitative impression of model performance appeared to be quite good. Scan features that appeared to impact model performance, included low image quality, unusual posterior curvature, and slanted scans.

Conclusions : Automated detection of GA lesions on OCT using a machine learning model is feasible. Future research is needed to optimize performance across a highly diverse scan quality spectrum and evaluate the model for quantification of GA for longitudinal measurements.

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