Simultaneous Segmentation of Geographic Atrophy in Longitudinally Acquired Fundus Autofluorescence Images.

Purpose: Accurately measuring geographic atrophy (GA) progression in clinical trials is challenging, owing to its slow and variable nature. This study develops deep learning algorithms that can measure GA growth rates with improved accuracy and robustness, by performing automated annotation of pre-existing versus newly expanding GA regions, based on simultaneous grading of fundus autofluorescence (FAF) image pairs. Design: Retrospective analysis of data acquired in 2 independent prospective clinical trials (i.e., 830 images in AREDS2 [Age-Related Eye Disease Study 2] and 273 images in METformin for the MINimization of Geographic Atrophy Progression [METforMIN]). Participants: One hundred seventy-four AREDS2 and 44 METforMIN participants. Methods: A standard DeepLabV3+ (DL) segmentation model that automatically segments GA from a single FAF image (DeepLabV3+ Main Outcome Measures: Image pixel overlap (using precision, recall, and Dice), GA area, and GA growth rates were measured with manual contours as the gold standard. The study also quantified the effect of longitudinal alignment of images to GA growth measurements. Results: When compared to manual expert annotations, the DL Conclusions: The simultaneous segmentation of longitudinal FAF images yielded improvements in accuracy and robustness by maintaining the longitudinal consistency of measurements. Such improvements are critical to the translation of automatic algorithms and minimizing measurement variability to effectively assess clinical trial outcomes in GA. Financial Disclosures: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.