Geographic Atrophy Enlargement Using Manual, Semi-Automated, and Deep Learning Approaches

Abstract

Purpose : Enlargement of Geographic atrophy (GA) serves as an important endpoint in clinical trials. It is traditionally quantified using manual annotation or semi-automated software utilizing fundus autofluoresence (FAF) images. Recently, artificial intelligence (AI) has been validated as a reliable alternative to the status quo. The goal of this study was to evaluate methodological agreement in the quantification of GA and its progression rates.

Methods : FAF images from 159 eyes (97 subjects) from the GSK (NCT01342926) GA clinical trial were annotated at baseline and 1 year follow-up by certified graders using Heidelberg Eye Explorer semiautomated software (Region Finder), manual planimetry, and an AI model, as depicted in Figure 1. The AI model was initially trained and validated on an independent set of manual planimetry annotations. Mean difference in GA area and Dice Coefficients were used to assess agreement between the 3 measurements. Growth rates were also compared over 1 year follow-up.

Results : Mean baseline area of GA was 7.88 mm² (SD 4.71) using Region Finder, 8.26 mm² (SD 4.90) using manual method, and 8.16 mm² (SD 4.96) using AI. Change in area at one year was 1.55 mm² (SD 1.26) using Region Finder, 1.65 mm² (SD 1.27) using manual method, and 1.55 mm² (SD 1.26) with AI. Mean difference in growth rate between manual and AI was 0.093 mm²/year, Region Finder and AI was -0.007 mm²/year, and Region Finder and manual was -0.118 mm²/year. Dice Coefficient including all visits was 0.93, 0.93, and 0.94 respectively.

Conclusions : AI, manual, and semi-automated (Region Finder) methods are comparably accurate in quantifying GA in FAF images, with a slightly lower measurement observed with Region Finder. Enlargement rates were comparable between the three methods. Semi-automated and manual approaches are time consuming and require significant time and effort to annotate while deep learning models can accurately detect GA and assess area in a fraction of the time without compromising quality.

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

 

Figure 1: Raw FAF image (a) annotated with AI (b), manual planimetry (c), and Region Finder (d).