Retina-RAG: AI Framework for Retinal Diagnosis and Report Generation
A new framework called Retina-RAG has been introduced by researchers, integrating a retinal classifier with the Qwen2.5-VL-7B-Instruct vision-language model and enhanced via Low-Rank Adaptation (LoRA). This tool aims to evaluate diabetic retinopathy severity, detect macular edema, and produce clinical documentation. Its retrieval-augmented generation module enhances diagnostic performance by incorporating ophthalmic data. Retina-RAG boasts impressive results, scoring 0.731 for diabetic retinopathy grading and 0.948 for macular edema detection, significantly surpassing previous zero-shot Qwen performance. This versatile and economical system is designed for seamless integration of components. The research is published on arXiv under the identifier 2605.06173.
Key facts
- Retina-RAG jointly performs DR severity grading, ME detection, and report generation.
- Architecture decouples a retinal classifier and Qwen2.5-VL-7B-Instruct adapted via LoRA.
- RAG module injects ophthalmic knowledge and structured classifier outputs at inference.
- F1-score of 0.731 for DR grading and 0.948 for ME detection.
- Outperforms zero-shot Qwen (0.096 for DR, 0.732 for ME).
- Low-cost modular framework enabling flexible component integration.
- Aims to improve diagnostic consistency and reduce hallucinations.
- Published on arXiv with ID 2605.06173.
Entities
Institutions
- arXiv