NEAR-INFRARED REFLECTANCE IMAGING FOR THE ASSESSMENT OF GEOGRAPHIC ATROPHY USING DEEP LEARNING

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Geographic atrophy (GA) is a severe form of dry age-related macular degeneration.
• Near-infrared reflectance (NIR) imaging is an accessible tool for retinal assessment.
• Current GA assessment methods can be time-consuming and subjective.

Purpose of the Study:

• To develop an automated deep learning system for GA detection using NIR imaging.
• To evaluate the performance of deep learning models in classifying and localizing GA.
• To establish NIR imaging as a reliable modality for GA assessment.

Main Methods:

• Retrospective analysis of NIR images from GA patients (≥50 years) and healthy controls.
• Training and validation of deep learning models, including Vision Transformer B16 and YOLOv8-Large.
• Performance evaluation using metrics like accuracy, precision, sensitivity, F1-Score, IoU, and DICE coefficient.

Main Results:

• Classification models achieved >95% accuracy; Vision Transformer B16 showed 98.5% precision and sensitivity.
• YOLOv8-Large demonstrated 91% sensitivity and 91% precision for GA localization.
• High accuracy and reliability in detecting GA using deep learning on NIR images were confirmed.

Conclusions:

• Deep learning models reliably identify geographic atrophy on NIR images.
• Automated GA assessment via NIR imaging can streamline patient selection for new treatments.
• This approach enhances the utility of widely available NIR imaging technology.