Exploratory study on a novel automatic quantification software by artificial intelligence for geographic atrophy associated to age-related macular degeneration.

OBJECTIVES: To determine the accuracy and repeatability of novel artificial intelligence (AI)–based software for automated quantification of geographic atrophy (GA) associated with age-related macular degeneration (AMD) using swept-source optical coherence tomography (SS-OCT) images. METHODS: This cross-sectional observational study included 33 eyes from 33 patients with GA secondary to AMD and one healthy control eye, all without a history of choroidal neovascularization. Each eye underwent three consecutive 7 × 7 mm2 SS-OCT scans (Topcon Corporation, Tokyo, Japan). Two independent observers manually delineated GA areas on en face OCT images in a masked fashion at two time points separated by 15 days. Individual and mean manual measurements were compared with automated AI-derived GA measurements, yielding a total of 165 measurements. Lesions were categorized according to morphology (regular vs. irregular) and number (single vs. multiple). Agreement between manual and AI measurements was assessed using Lin’s concordance correlation coefficient (CCC). Software repeatability was evaluated using the intraclass correlation coefficient (ICC) based on a two-way mixed-effects model. RESULTS: The mean patient age was 73.4 ± 11.7 years. Twelve eyes exhibited regular lesion morphology, and 11 eyes had single GA lesions. Intra-observer repeatability was excellent, with ICCs of 0.999 (95% CI, 0.997–0.999) for Observer X and 0.999 (95% CI, 0.999–0.999) for Observer Y. Agreement between the mean manual measurements of both observers was also high (ICC = 0.997; 95% CI, 0.989–0.999). Agreement between AI-based measurements and the mean of the four manual annotations was strong for the overall cohort (CCC = 0.992; 95% CI, 0.981–0.998). Stratified analyses showed CCCs of 0.989 (95% CI, 0.816–0.997) for regular lesions and 0.988 (95% CI, 0.944–0.998) for irregular lesions, and 0.999 (95% CI, 0.998–1.000) for single lesions and 0.984 (95% CI, 0.963–0.995) for multiple lesions. Agreement by lesion size was lower for small lesions (< 3.7 mm2; n = 12; CCC = 0.903; 95% CI, 0.742–0.974) and higher for medium (3.7–9.4 mm2; n = 13; CCC = 0.978; 95% CI, 0.940–0.994) and large lesions (> 9.4 mm2; n = 8; CCC = 0.956; 95% CI, 0.676–0.998). CONCLUSIONS: This AI-based software demonstrates high accuracy and excellent repeatability for automated GA area quantification across different lesion morphologies, numbers, and sizes, with minimal processing time. These findings could support its potential utility in routine clinical practice and research applications. CLINICAL TRIAL NUMBER: Not applicable.