Deep Learning based Segmentation of Geographic Atrophy in Fundus AutoFluorescence: External Validation in Imaging Acquired in a Clinical Trial

Abstract

Purpose : Hypo-autofluorescence in fundus autofluorescence (FAF) images results from the loss of retinal pigment epithelium (RPE) and is an FDA-approved outcome as a sign of geographic atrophy (GA) associated with age-related macular degeneration (AMD). We evaluated the performance of an automatic segmentation algorithm (De Silva et al., ARVO 2022) that performed reliably on the hold-out internal test set (IOU = 0.86), on an external data set acquired in an independent GA clinical trial.

Methods : 704 FAF images (195 eyes, 542 visits) were acquired as part of AREDS2 and were used to develop an automatic deep learning-based segmentation algorithm utilizing DeepLabV3 model. Ground truth binary masks were delineated by trained graders by performing pixel-level annotations of regions with hypo-autofluorescence. All images were fovea-centered and contained multifocal, central and non-central GA lesions. The algorithm's performance was tested on an external data set acquired as part of a clinical trial (Phase II study of Pegcetacoplan (APL-2) therapy in patients with GA [FILLY], NCT02503332) conducted by Apellis Pharmaceuticals (1557 visits). Qualitative and quantitative analyses were conducted by comparing algorithm-predicted and Apellis's reading center-annotated GA areas from FAF images.

Results : A comparison between reading center GA areas and the algorithm-predicted GA areas, measured in mm², revealed a mean absolute error of 0.35±0.7 mm² and a mean relative absolute error of 0.13±0.3 for eyes with average GA lesions of 9.57±5.25 mm². The Bland-Altman plot exhibited robust performance, displaying a mean of -0.1 mm² and limits of agreement ranging from -1.6 to 1.4 mm². The scatter plot demonstrated that most points aligned along the line passing through the origin with a slope of 1 and had an R-value of 0.97, indicating the algorithm's robustness across an external dataset spanning various GA-based disease severities from small unifocal lesions to enlarged multifocal lesions.

Conclusions : In advanced AMD, GA lesions manifested range of morphological changes, featuring multiple localized areas of decreased hypoautofluroscence. The segmentation of retinal FAF images may be useful for monitoring structural changes, potentially identifying eyes at a high risk of atrophy progression and functional decline and identify candidates for future clinical trials.

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

 

Fig 1

 

Fig 2