An autonomous multimodal AI agent for evidence-grounded ophthalmic diagnosis
TL;DR
AgentEYE is an auditable multimodal agent that routes ocular images to specialized fundus and B-scan tools, retrieves guideline/web evidence, and synthesizes evidence-grounded reports that supports evidence grounding and citation auditability.
Abstract
Summary Multimodal ophthalmic diagnosis requires integrating fundus photography, B-scan ultrasonography, and medical evidence, yet most artificial intelligence (AI) systems remain single-task or weakly grounded. AgentEYE is an auditable multimodal agent that routes ocular images to specialized fundus and B-scan tools, retrieves guideline/web evidence, and synthesizes evidence-grounded reports. In a 302-case internal benchmark, AgentEYE shows higher diagnostic correctness and completeness than large language model (LLM)-only baselines and an ablation without specialized imaging tools; performance remains similar to the no-retrieval ablation, indicating that retrieval mainly supports evidence grounding and citation auditability. Blinded evaluation of 200 cases by three ophthalmologists confirms improved diagnostic correctness, completeness, safety, and citation grounding versus an LLM-only self-citation baseline. External analyses show distribution-dependent performance. These findings support AgentEYE as a traceable decision-support prototype requiring prospective multicenter validation.