Near InfraRed Reflectance Imaging for the Assessment of Geographic Atrophy Using Deep Learning.

Abstract

Purpose: Near-infrared reflectance (NIR) imaging is a widely available but underutilized modality for assessing geographic atrophy (GA), a late-stage manifestation of dry age-related macular degeneration. This study aims to develop and evaluate a fully automated deep-learning-based approach for detecting GA on NIR imaging.

Methods: NIR images of patients aged ≥ 50 years with GA, confirmed by two retinal specialists, were analyzed at Rabin Medical Center. The control group included NIR images of patients with healthy-appearing retinas. Models were trained and evaluated based on accuracy, precision, sensitivity, F1-Score, and DICE coefficient.

Results: A total of 113 GA patients and 119 controls were included. The classification dataset contained 330 images, and the localization dataset included 659 images. Classification models performed well, with accuracy above 95%, while Vision Transformer B16 achieved the best results (precision=98.5%, sensitivity=98.4% and accuracy=98.5%). For GA localization, YOLOv8-Large achieved 91% sensitivity, 91% precision, an IoU of 84%, and a DICE coefficient of 88%.

Conclusion: GA can be reliably identified using NIR images. Deep learning models can assist in evaluating GA on this routinely available imaging modality, aiding in the selection of patients who may benefit from emerging therapies.