| Title |
|---|
| EyeFlow: A Self-Evolving Multi-Agent Framework for Knowledge-Grounded Ophthalmic Diagnosis |
| Authors |
|---|
| Bingjie Yan, Xiaolan Chen, Mingguang He, Danli Shi |
| Presenting |
|---|
| Bingjie Yan |
| PURPOSE: |
|---|
| Clinical diagnosis is a multi-step reasoning process that requires structured history-taking, multimodal evidence integration, and iterative hypothesis refinement. We present EyeFlow, a multi-agent framework for ophthalmic diagnosis that orchestrates specialized agents through a clinically-aligned diagnostic workflow. |
| METHODS: |
|---|
| EyeFlow employs four specialized agents: an orchestrator for workflow control, an image analysis agent using Dino-V3 for multimodal interpretation, debate agents for specialist consultation, and a report agent for synthesis. The system incorporates three core components: (1) a separate image analysis pipeline, (2) a multi-agent debate mechanism, and (3) a memory system for knowledge retrieval. We evaluated the system on 169 semi-synthetic ophthalmic cases and conducted systematic ablation to quantify component contributions. |
| RESULTS: |
|---|
| EyeFlow achieved Top-1 accuracy of 80.6% and Top-5 accuracy of 94.8%. Ablation analysis revealed that the image analysis pipeline contributed +11.3% to Top-1 accuracy, the debate mechanism contributed +11.6%, and the memory system contributed +8.1%. The complete system significantly outperformed all single-modality baselines. |
| CONCLUSIONS: |
|---|
| EyeFlow demonstrates that multi-agent collaboration with specialized roles enables structured clinical reasoning in ophthalmology. The framework provides an interpretable architecture for integrating vision-language models, specialist consultation, and knowledge retrieval into a unified diagnostic system. |