Deep learning identifies hyperreflective foci as predictors of geographic atrophy progression

Abstract

Purpose : To characterize the incidence and distribution of hyperreflective foci (HRF) in eyes with geographic atrophy (GA) and to correlate HRF occurrence with GA growth.

Methods : We analyzed 1.106 SD-OCT scans (Spectralis) and corresponding fundus autofluorescence (FAF) images of 171 eyes with GA obtained in 94 patients. Follow-up examinations were performed prospectively every 6 months for a mean of 36 (range, 12-87) months. To establish anatomical correspondence between multimodal images and over time, all examinations were registered by validated deep learning tools using the retinal vasculature as landmarks. The GA lesions were manually annotated on all FAF images by certified, masked readers following a standardized protocol. HRF were automatically segmented in all OCT scans using a deep learning-based algorithm, and overlaid with FAF images. The number and density (volume / mm²) of HRF was analyzed with respect to the distance to the GA border, and compared in areas of GA growth versus non-growth areas.

Results : The algorithm reliably identified and quantified HRF in all GA lesions. The vast majority of HRF were found in the junctional zone immediately adjacent to the GA border, their prevalence and density decreased rapidly with distance to the GA margin towards intact retina. Significantly greater densities of HRF were consistently detected within areas of imminent GA growth compared to non-growth areas resulting in a topographic map of disease activity. The amount of HRF in the retina was not associated with the speed of GA enlargement.

Conclusions : HRF represents a pathognomonic sign of disease activity and imminent focal progression in GA. A locally increased density of HRF is predictive of future GA enlargement. Artificial intelligence allows reliable and fully automated assessment and quantification of HRF with the potential to provide an ideal tool for GA monitoring in clinical practice.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.