Abstract
BACKGROUND: Geographic atrophy (GA), the endpoint of dry age-related macular degeneration (AMD), is irreversible. The recent approval by the Food and Drug Administration of a complement component 3 inhibitor marks a significant breakthrough, highlighting the critical importance of early detection and management of GA. Consequently, there is an urgent and unmet need for efficient, accurate, and accessible methods to identify and monitor GA. Artificial intelligence (AI), particularly deep learning (DL), applied to noninvasive retinal imaging, offers a promising solution for automating and enhancing GA management. OBJECTIVE: This systematic review aimed to assess the performance of AI using noninvasive imaging modalities and compare it with clinical expert assessment as the ground truth. METHODS: Two consecutive searches were conducted on PubMed, Embase, Web of Science, Scopus, Cochrane Library, and CINAHL. The last search was performed on October 5, 2025. Studies using AI for GA secondary to dry AMD via noninvasive retinal imaging were included. Two authors worked in pairs to extract the study characteristics independently. A third author adjudicated disagreements. Quality Assessment of Diagnostic Accuracy Studies-AI and Prediction Model Risk of Bias Assessment Tool (PROBAST) were applied to evaluate the risk of bias and application. RESULTS: Of the 803 records initially identified, 176 were found through an updated search. Subsequently, 200 papers were assessed in full text, of which 41 were included in the final analysis, 10 for GA detection, 20 for GA assessment and progression, and 11 for GA lesion prediction. The reviewed studies collectively involved at least 24,592 participants (detection: n=7132, assessment and progression: n=14,064, and prediction: n=6706), with a wide age range of 50 to 94 years. The studies spanned a diverse array of countries, including the United States, the United Kingdom, China, Austria, Australia, France, Israel, Italy, Switzerland, and Germany, as well as a multicenter study encompassing 7 European nations. The studies used a variety of imaging modalities to assess GA, including color fundus photography, fundus autofluorescence, near-infrared reflectance, spectral domain-optical coherence tomography (OCT), swept-source (SS)-OCT, and 3D-OCT. DL algorithms (eg, U-Net, ResNet50, EfficientNetB4, Xception, Inception v3, and PSC-UNet) consistently showed remarkable performance in GA detection and management tasks, with several studies achieving performance comparable to clinical experts. CONCLUSIONS: AI, particularly DL-based algorithms, holds considerable promise for the detection and management of GA secondary to dry AMD with performance comparable to ophthalmologists. This review innovatively consolidates evidence across GA management-from initial detection to progression prediction-using diverse noninvasive imaging. It has strong potential to augment clinical decision-making. However, to realize this potential in real-world settings, future research is needed to robustly enhance reporting specifications, ensure data diversity across populations and devices, and implement rigorous external validation in prospective, multicenter studies.