SRS61 - Data-driven diagnosis: clinical utility of ai image analysis for retinal disease identification and surgical decision-making

Background and Aims Artificial Intelligence (AI) is influencing the field of ophthalmology by increasing diagnostic precision for prevalent and rare eye diseases. The study assessed the diagnostic performance of different AI models using a comprehensive and diverse fundal image dataset. Methods In total, 47 853 fundal images were collected and 35 000 were analyzed. Subsets included 5000 images from common conditions and 2000 from rare conditions from database sources including, the IDRiD, EyePACS, Messidor-2, JSIE, DIARETDB0/1, and APTOS 2019. Seven AI models were assessed using an identical prompt structure for consistency: ChatGPT-5, Sonar, DeepSeek, Qwen, Microsoft AI, Grok, and Google DeepMind. One-way ANOVA was used to test model performance, followed by Tukey’s HSD tests. Results ANOVA supported the null hypothesis that model performance was variable across the models (F = 12.37, P < 0.001). Post-hoc analysis showed Google DeepMind was significantly better than all (P < 0.05). Microsoft AI and ChatGPT-5 both had similar performance relative to one another (P = 0.11) and both were significantly better than the Sonar, DeepSeek, Qwen, and Bing AI models (P < 0.05). Conclusions The Google DeepMind model had the best diagnostic accuracy followed by Microsoft AI and then ChatGPT-5. In addition, for all models performance differed more for common conditions than rare conditions.