Longitudinal Comparison of Geographic Atrophy Measurements Using Manual and Deep Learning Approaches

Abstract

Purpose : Change in geographic atrophy (GA) area serves as an important endpoint in clinical trials. It has traditionally been measured through manual or semi-automated annotation of fundus autofluorescence (FAF) images. Recently, artificial intelligence (AI) has been validated as a reliable alternative to these methods. The goal of this study is to compare the measurement of GA and its progression rates between manual and AI methodologies.

Methods : FAF images of 159 eyes (97 subjects) from a phase 2 study sponsored by GlaxoSmithKline (NCT01342926), a GA treatment trial which failed to meet its primary outcome, were annotated at baseline and one year follow-up by certified graders using manual planimetry and by an AI model. The AI model was initially trained (n= 601) and validated (n= 156) on an independent set of manual planimetry annotations. Mean difference in GA area and Dice coefficients were used to assess agreement between the two measurements. Growth rates were also compared over one year follow up.

Results : Mean baseline GA area was 8.26 mm² (SD 4.90) using manual planimetry and 8.16 mm² (SD 4.96) using AI. Mean area of GA at one year follow-up was 9.91 mm² (SD 5.43) for manual planimetry and 9.71 mm² (SD 5.29) for AI. Change in area at one year was 1.65 mm² (SD 1.27) using manual planimetry and 1.55 mm ² (SD 1.26) with AI. Mean difference in growth rate between the two methods was 0.093 mm ²/ year. Dice coefficient including all visits was 0.93.

Conclusions : Manual and AI methods were comparably accurate in quantifying GA in FAF images. Differences in longitudinal measurements may be due to a lack of historical viewpoint when utilizing AI compared to manual planimetry. Whereas human graders are able to reference past annotation and measurement for comparison, the deep learning model employed lacks this ability.

This abstract was presented at the 2024 ARVO Imaging in the Eye Conference, held in Seattle, WA, May 4, 2024.

 

Example of a FAF image of GA (A) as annotated by manual planimetry (B) and AI (C).